This paper proposes an arousal and emotional valence-level estimations method using non-contact environmental sensing, which aims to solve some of the problems of conventional emotion estimation methods such as stress and discomfort caused by wearing measurement devices and invasion of privacy caused by using facial images. Construct models using machine and deep learning to estimate arousal and emotional values from environmental data. Although the arousal level estimation accuracy is somewhat low at about 0.68, emotional valence achieves about 0.85 with machine learning with decision trees.

Few disaster rescue support information systems would integrate disaster response robot collected information and vicinity medication institution information . This research proposes such a system that can display the disaster area overview, geographic information, medical institution's emergency current processing capability, 3D building model and the robot captured data. A user can observe disaster situations, mark interested sub-area, pin(associate) some video captured by the robot with the 3D model.

This paper introduces the P4 programmable testbed built by TWAREN. In this testbed, you can define how the P4 switch handles packets by writing P4 language code, providing more customization than the SDN network. However, hardware P4 switches are expensive, and many users often use BMv2 for research due to financial reasons. Therefore, we have selected several nodes on the TWAREN network to build multiple physical P4 switches and connected them via VPLS to form a large remote P4 programmable testbed. With P4 programmable testbed, users can verify their mechanism on a large-scale physical network.

Currently, the demand for innovative technology is increasing, and integrating technology into life in this intelligent society is a significant issue. Society already has many examples, such as smart door locks, refrigerators, and air conditioners. Intelligent home appliances refer to household and similar electrical appliances that adopt one or more intelligent technologies, have one or more intelligent features and must have specific smart features that consumers can perceive. In this paper, we use the Raspberry Pi as the development environment, which uses Mediapipe to recognize the gesture node, and selects the gesture to switch the light on and off.

This paper applies ant colony optimization (ACO) algorithm for the dual-pin flying probe circuit board inspection optimal path searching problem. First, the proposed approach creates a representation for circuit inspection path, which groups and encodes the measurement endpoints into a sequence. Moreover, the ant colony algorithm with the sequence can be used to solve the complex circuit optimal problem. The experimental results show that the best solutions of ACO are performed better than other algorithms.

The future of 5G will focus on massive machine-type communication (mMTC), and many applications will be based on this application and gradually mature. These applications include smart cities, e-health, and the Internet of Vehicles (IoV). In addition, the Narrowband Internet of things (NB-IoT) is a critical technology in mMTC. NB-IoT can bring many benefits, such as efficient bandwidth use and long-distance transmission. However, although NB-IoT has many advantages, issues such as throughput, energy consumption, and security still need to be solved. In order to effectively solve these problems, the choice of cluster heads is very important. Considering that the energy consumption of the CH will be more severe than that of other devices, which will also shorten the network; therefore, we will focus on the dynamic cluster head selection. To solve this problem, we introduced the concept of the pseudo-rigid body into this paper and proposed a method similar pseudo-rigid body to determine the CHs; at the same time, we also defined the maximum range of each cluster. The experimental results show that the proposed method can dynamically select the CH and increase the network lifetime.

Wearable sensors have a significant increase in both research and commercialization as a kind of consumer electronics. In the field of healthy science, wearable sensors provide affordable solutions for massive screening or long tern monitoring, for instance, wearable dry electroencephalography (EEG) for brain and electrocardiogram (ECG) for heart. Human brain and heart are the most important organs and the targets for healthy monitoring. Given that brain and heart sometimes have comorbidities of each other as they are reciprocally connected, it is important to monitor their relations. However, only a few studies addressed the relationships between them under healthy states. This study aims to examine the brain-heart interactions(BHI) under different states, including two different resting states and one attentional state using wearable sensors. Twenty subject were recruited and performed a memory task in a virtual supermarket. Five-channel dry EEG and two-channel ECG were acquired before, during, and after the task. Pearson correlation was employed to analyze the relations between EEG features and heart rate variability (HRV). Two sample t test and machine learning method were employed to analysis the difference between states. We found that higher frequency oscillations in the brain were used to communicate with the heart during the pre-rest state and task state respectively, then switched to lower frequency for resetting the BHI after task. Moreover, the combination of EEG and ECG features can best distinguish the pre and post resting states. Our findings suggest that BHI are dynamic and state-dependent and studying the BHI can provide better understanding of the states in the body to aid the diagnosis and/or treatment for diseases affected both brain and heart.

With machine learning models continuously growing in size and short release cycles of GPUs, hardware becomes outdated very soon. To cope with the ever-growing model sizes, we seek out ways to better utilize the computing power we already possess. This paper implements a makespan-aware distributed training framework called Train ‘N' Play (TNP) to make training on large models and large datasets possible for systems that originally could not accomplish.

Sign Language is one of the important communication methods for the hearing-impaired person. As it is hard to learn, it has been researched several times throughout various motion capture technologies. In this paper, MediaPipe Hand, Google's AI framework, and LSTM, a type of RNN, were used for training our dynamic Sign Language recognition system. Unlike in previous researches, our trained model translates the dynamic Sign Language in real-time with displaying the translated sentences. Sign language differs from existing English grammar in terms of language structure. To solve this problem, there is a way to use Neural Machine Translation, but due to the limitation of the number of words used in this research, we used a simple rule-based structure converting the recognized word sequences into sentences with English grammar.

Mukbang ASMR has become a popular form of audio-visual interaction in recent years, and many YouTube channels related to Mukbang ASMR topics have millions of subscribers and high viewership. The "Rainbow Candy Mukbang ASMR" in this study was used as an example to analyze visual behavior and detect hearing and emotion through eye-tracking, heart rate, and galvanic skin response (GSR). The results showed that rainbow jelly had the best TFD and TFF performance in eye movement among five types of candies, and the chewing sound of Mukbanger significantly decreased the GSR of viewers, indicating that the chewing sound affected viewers' emotions.

Emotions experienced by learners play a significant role in both their cognitive and affective learning gains, and thus, the comprehension and measurement of emotions in technology-rich learning environments (TREs) are essential aspects of 21st-century education. This study aims to design an application interface that can measure the topic emotions and epistemic emotions experienced by children after reading books. The application flow and interface design are presented, featuring emotion wheels for both topic and epistemic emotions. Data collected with this application can further be used to explore the correlation and causation between the folksonomy topic structure and emotion categories by analyzing the data. The data could potentially help in training machine learning models for sentiment analysis in future studies.

We have developed a transmissive LED touch display for engineered marble that can be used for such as cooking support on a kitchen counter. This device has an LED matrix for display and a reflective infrared sensor matrix for touch detection, and both are placed on the back of the engineered marble. The dots displayed by the LED matrix are transmitted through the engineered marble, and the blurred image is displayed on its surface. The reflective infrared sensor matrix can detect the positions of objects existing on the surface of the engineered marble. By incorporating these two functions into one device, we were able to develop a unique device suitable for smart home applications.

E-commerce has become a major shopping channel for people in recent years, and women's bags are one of the most popular products in online stores. This study uses eye-tracking data to analyze the effects of the brightness of product images and the legibility of text within products on subjects' attention and visual tendency. The results showed that more visual attention was obtained when the brightness of the product photo was increased and placed on the right side. Harder-to-read fonts were placed on the left side to get more visual attention. Regardless of the visual presentation, when placed on the right side, a change on the original Z-shaped reading line was observed.

Interactive picture books on tablets have become increasingly popular among parents and children. This study investigated parental decision factors for purchasing interactive picture books by conducting semi-structured interviews with 10 parents of preschool children in the Jiangsu, Zhejiang, and Shanghai regions of China. The results showed that educational value and interactivity were important factors that influenced parents' purchasing decisions. These findings can help manufacturers better understand consumers' needs and adjust product strategies to produce products that meet the needs of parents and children. Future research could expand the scope of the investigation to identify factors that influence parents' willingness to pay for interactive picture books, providing potential suggestions for their development in mainland China.

This work aims to develop a camera-less photography system, which utilizes the geographic information to retrieve images with similar geographical characteristics from cloud data. The proposed approach contains two parts, insert subsystem and search subsystem. The insert subsystem is used to store the image data in the cloud database through a module which also records the surrounding information. As for the search subsystem, images are retrieved from the cloud database according to the geographic information. Finally, the search subsystem provides some photos possess similar geographic and related information as the in-situ image acquisition scenario. In addition to software development, the system is also implemented as a hardware device, which can be used in the outdoor applications.

One of the most important issues in smartphones is power consumption. The CPU is one of the most power-consuming devices, and controlling the CPU clock rate affects the device's power consumption and performance. However, there is a trade-off between improving device performance and reducing power consumption. Increasing the CPU clock rate improves performance but increases power consumption. Conversely, decreasing the CPU clock rate reduces performance but helps reduce power consumption. Therefore, it is ideal to decrease the CPU clock rate when and only when its usage is low. However, the Linux kernel installed on Android smartphones controls the CPU clock rate by tracking past CPU usage. Therefore, when the CPU usage changes rapidly, there is a time lag between the CPU usage and the CPU clock rate. To address this issue, in our previous work, we proposed a CPU clock rate control concept based on observation of the behavior of the foreground application. In this paper, we demonstrate the effectiveness of the previously proposed method by implementing a simple CPU clock rate control system and comparing it with the existing scaling controller.

We propose an indoor location estimation method using sensors of mobile devices. The sensors used for the estimation are a triaxial acceleration sensor, a triaxial angular rate sensor, and a triaxial geomagnetic sensor. First, we perform attitude estimation using each sensor. This estimation is to estimate the attitude of the mobile device concerning the earth. By knowing the attitude, the acceleration in the gravity direction and the horizontal direction can always be measured, no matter the mobile device's attitude. Based on the acceleration and other information obtained from the attitude estimation, step detection, step length, and direction of the step are estimated. Finally, the location is calculated using all the estimation results. The location estimation is based on the PDR (Pedestrian Dead Reckoning) algorithm. To eliminate the need to hold the mobile device in place during the estimation process, we configured the method to be able to estimate while walking while looking at the screen and while walking while holding the device in one's hand. Although the accuracy could be higher, experimental results show that the proposed method effectively estimates the location and walking trajectory.

The road navigation system which can be often seen everywhere plays an important role in the modern life. The appearance of this system replaces the traditional paper map, thus it not only can efficiently reduce the time to find the way for the drivers and substantially absolve the trouble of getting lost and risk, but also can increase the driving safe. However, the existing system generally considers the information of planned path length, road speed limit, or road type in the map when performing the path planning, rather than considers other road conditions that will also affect the applicability of path planning, such as the effects of traffic lights and traffic jams, and so on. This paper mainly proposes an improved path planning algorithm for realistic roads by adding the function of further considering the waiting time of traffic lights to the existing one, to meet the requirements of applications in the future intelligent transportation system.

In this study, we focus on the issue of calculation reliability and the overhead in vehicular microclouds. We propose a scheduling solution, model the proposed system, and propose an algorithm to obtain the best scheduling configuration. We use simulation to cross-validate our mathematical model, and our evaluation shows that the proposed algorithm performs best.

Mobile edge computing has been introduced in various fields. The discriminator of the edge device need to work for local optimized problems. However, since the edge devices are worked in limited areas, these cannot obtain novel learning data by themselves. Therefore these discriminator needs to update by using the knowledge of other devices. Thus, in this paper, a method of updating models using weights of the Neural Networks by transferring and fitting is proposed. It is expected models can be updated for solving local optimized problems for sharing the knowledge of another model that are effective for learning.

The COVID-19 pandemic has resulted in an escalation in the demand for online learning, leading to the need for experts, such as experienced math teachers, to classify exercises. However, achieving accurate and effective classification may present challenges due to differing expert opinions and complex exercise concepts. To address this challenge, we propose the use of the Concept Identification Visualizer (CIV). The CIV tool assists experts who lack \emph{engineering programming} knowledge in comprehending the exercises and evaluating student responses. The tool leverages Knowledge Tracing to extract relevant information from students' answers and presents this data in a visual format. By providing a more comprehensive understanding of the exercises, the experts enable more informed exercise classification based on student feedback and improve the overall effectiveness of online learning.

This paper presents a positioned quality evaluation method for posterior breasts in Craniocaudal-view mammograms. The proposed method employs a convolutional neural architecture (CNN) to recognize multiple candidates extracted from the posterior-inner and posterior-outer breasts and determines whether visualize fat posterior to glandular tissues. To find the proper classifier, we experiment with the two-category classifications with various CNN models, including, Squeezenet, Densenet-121, Alexnet, Vgg-19, and Resnet-101. Our method determines the evaluation result of posterior breasts by consensus decision-making from the classification results of multiple candidates. The experiment results demonstrate that Densenet-121 is the best choice for this work.

According to the criteria revealed by the PGMI evaluation system, fat visualized posterior to glandular tissue is a significant criterion for mammography positioned quality evaluation. This paper introduces the quality evaluation method in craniocaudal-view mammograms. The proposed method implements a classifier with a convolutional neural network (CNN) and then analyzes the inner and outer regions of the posterior breast. To find the proper CNN model, we experiment with the classification with five CNNs, including, Squeezenet, Densenet-121, Alexnet, Vgg-19, and Resnet-101. Furthermore, analyzing the gradient-weighted class activation mapping extracted from CNN models interprets what features the model observes. The experiment results demonstrate that Densenet-121 is the best choice for positioned quality evaluation of posterior breasts.

The early detection of pancreatic cystic lesions plays a significant role in the survival chance of a patient with pancreatic cancer. Yet it is still a huge challenge. Unfortunately, most pancreatic cancers were diagnosed when the tumor was metastatic. In this study, the Hybrid Transformer, which is the combination of the VGG19 network and Vision Transformer, is utilized as a learning model to predict the pancreatic cystic symptom types in needle-based confocal laser endomicroscopy. A total of 16,944 images containing five types of pancreatic cystic lesions are collected as the training and validation data. Our method can automatically classify the subtypes of pancreatic cystic lesions in the test videos and record the prediction results frame by frame. In our experiment, the proposed method successfully identifies subtypes of 13 from 18 test videos and achieves an accuracy as high as 72.2%.

In this paper, we propose a Graph Neural Network-based model to detect fraudulent users on e-commerce platforms without relying on rating scores. Utilizing user-product bipartite graphs and timestamp data, we capture temporal patterns and neighborhood information, creating a graph with multi-dimensional edge vectors. Our model demonstrates competitive performance compared to state-of-the-art methods, effectively identifying fraudulent users under data-insufficient conditions and enhancing the overall reliability of online platforms.

In this paper, we apply wireless technologies to implement automatic abnormal event detection and E-clinical cart information systems for long-term care institutions. The E-clinical cart can follow medical personnel, easing the physical workload of manually pushing the clinical cart. For autonomous patrol work, some abnormal actions, such as falls and squats, can be detected using the proposed human action recognition with few-shot data. Moreover, the proposed monitoring platform with a graphical user interface (GUI) can improve the E-quality and efficiency of nursing work. The experimental results show that the intelligent data monitoring system can effectively carry out the task assignment of the E-clinical cart and event monitoring and warning in long-term care institutions

The main objective of this paper is to utilize only RGB images to let a robotic arm grasp a target object without the related 3D position information. The advantages of the proposed method include image-to-action translations that apply to a class of general robotic arms, and the mathematical analysis of the inverse kinematics is not necessary. We employ the RGB images and the proximal policy optimization (PPO) algorithm to train the reinforcement learning network in the Gazebo simulated environment. Finally, an illustrative example shows how effective the proposed strategy is.

Intelligent manufacturing practice fields often have poor information security; common security flaws in industrial control networks include messages are sending in plaintext, lack of source authentication, and lack of message integrity verification. In this study, methods for secure NC program updating, two-factor user authentication, and secure messages transmission module were proposed to improve the security of industrial control systems. Attacks were performed on a CNC machine to demonstrate that attackers could control the unprotected machines. We also verify the performance of the proposed M2M authentication scheme, which ensure message freshness, in preventing man-in-the-middle, impersonation attack, and replay attack.

Energy transmission using radio frequency (RF) technology became prominent. Utilizing RF technology, wireless charging vehicles (WCVs) can be employed to recharge the sensors, leading to the development of wireless rechargeable sensor networks. This study divides the WCVs into Recharging WCVs (R-WCVs) and Server WCVs (S-WCVs). The R-WCVs are responsible for recharging the sensors, while the S-WCVs are responsible for recharging the R-WCVs. This paper proposed a hierarchal charging vehicle scheduling (HCVS) algorithm for R-WCVs and S-WCVs, aiming to improve the energy usage efficiency of R-WCVs in each recharging cycle. In addition, this study investigated the relationship between two consecutive recharging cycles of R-WCVs. Experimental studies reveal that the HCVS algorithm yields the best result compared to the literature.

Based on the suspend and resume mechanism, this paper proposes fine-grained scheduling with a fragmented resource-reusing approach to improve the utilization of a GPU sharing system. The HRRN and the best-fit schemes are also adopted to alleviate the starvation problem for large-scale machine learning jobs submitted to the cloud. Experiments are simulated based on the trace logs and benchmarking data collected from a real cluster and the TF-Slim tool. Various image classification models are applied to the experiments to simulate heterogeneous job submissions to the GPU sharing system. Experimental results show that our approach improves GPU resource utilization by around 4.3 times and reduces overall makespan by at least 3.9 times compared to the sequential execution without GPU sharing.

Recommender systems are widely used to assist users in finding potentially interesting content. Most methods analyze personal preferences and provide users with personalized recommendation results. The personal preferences are also easily influenced by their relatives and friends. In this paper, we propose a group recommendation system based on graph attention network. The first stage is to group users according to the mean genre score of the movies, and the similarity between users. The second stage is based on a graph attention network (GAT), we predict users' rating for the movies by the preferences of the users and group relationship. The simulation results show that the accuracy of our recommendation mechanism is better than the existing recommendations

In recent years, recommendation systems have been widely used in the commercial field, and many e-commerce platforms and merchants have adopted the model of recommendation systems as a way to sell goods. However, when the product's features are insufficient, or the user's evaluation of the product is not much, it is difficult for the recommender system to recommend the product that meets the customer's expectations. For example, when recommending books in the library, you will find out that the books cannot be like clothes and can use rich and diverse ontology features such as brand, material, color, and length as the basis for the recommendation system. In addition, we found that only a few users will Evaluate or comment on books during the user's borrowing process. To sum up, this study will explore how to recommend books through natural language processing in the case of insufficient book features and insufficient user ratings and reviews.

With the development of the economy and the improvement of people's living standards, more and more people start to pay attention to health and exercise, and the smart bracelet is born. The designed smart bracelet in this paper can collect data such as heart rate, exercise, temperature, humidity, altitude, atmospheric pressure, and light intensity in the current environment. The collected data are processed by the STM32F103C8T6 controller. Then the processed data is displayed in the OLED display module. In addition, the corresponding data can be uploaded to the mobile phone for real-time display through the Bluetooth communication module. The test results show that the system is stable and reliable.

This paper designs a crib monitoring system based on the STM32 microcontroller. The system uses temperature and humidity sensors to collect the ambient temperature and humidity. The microcontroller processes and analyzes the collected data, and displays it through the display module. When the temperature is higher or lower than the preset threshold, the fan module or heating module is started to adjust the temperature. The temperature threshold is set by the key module. The system will give an alarm when the baby touches the fence to prevent the baby from falling. The system can adjust the motor speed through the key module, aiming to control the crib swing speed. Lullabies are played through the voice broadcasting module to help the baby relax and fall asleep; In addition, the temperature and humidity information is uploaded to the mobile terminal for real-time display through the Bluetooth communication module, and the mobile terminal can also control the motor speed, playing, and music switching. The test result shows that the crib monitoring system has the characteristics of fast response speed, high identification accuracy, and good stability. It can provide a theoretical basis for the actual crib production, and has a certain reference value.

This paper designs an automatic tracking car that uses machine vision technology to realize QR code recognition. The car uses the MSP430F5529 chip as the main controller. It uses the OpenMV camera as the image acquisition sensor and the OpenCV library to process the acquired images. It can locate the target of the QR code by monoculture visual positioning technology. The car can obtain the distance between it and the leading vehicle, and compare the actual distance with the expected distance. The main controller can control the motor operation according to the distance deviation obtained, aiming to realize the automatic following function of the car.

This study aims to use patent mining to explore the impacts of patent diversity on enterprise finance performance. The selected companies are from the list of Derwent Top 100 Global Innovators 2020 by Clarivate Analytics, patents are from the United States Patent and Trademark Office (USPTO) and Google Patents, and the Financial Information is from the macro trends website.

In recent years, Big Data analytics has gained widespread recognition and been applied in various fields, including marketing, finance, and medical domain. Text analytics, a subfield of Big Data analytics, has been a trending topic due to the emergence of fast improving Natural Language Processing (NLP) techniques and the increasing amount of text data on the web, such as blogs, forums, social media, and reviews. Leveraging text data presents an excellent opportunity for enterprises to acquire valuable insights and make data-driven decisions. However, there is no one-size-fits-all solution for companies adopting Text analytics. Strategies for implementing algorithms and tuning parameters may vary depending on different domains and scenarios. This paper introduces a web-based workflow system called Text Analytics and Retrieval Flow (TARFlow) for non-technical people to make use of NLP and machine learning techniques without programming, enabling them to concentrate on generating insights.

Facial recognition technology has been increasingly applied to daily life; however, differentiating identical twins remains a challenging task. This paper investigates the performance of facial recognition models on identical twins and introduces fine-grained image classification as a potential solution. We created a dataset of 54 pairs of twin images and tested various models on three datasets (LFW, SLLFW, and our homemade twins dataset) with different degrees of similarity. The Facenet model was chosen as the backbone network for our fine-tuned model due to its outstanding performance. The fine-tuned model showed improved performance in distinguishing negative pairs compared to the pretrained model and had slightly better accuracy than human recognition.

Drawing upon the social affiliation model, we build theory examining the consequences of daily interaction with AI. Specifically, we theorize that the more employees interact with AI in the pursuit of work goals, the more they experience a need for social affiliation during the work day, which then motivates subsequent behavior to cope with the daily surge in such need (more helping behaviors and less interactional avoidance). We test our hypotheses with 128 financial analysts who work with AI on a daily basis in a Fortune 500 European financial institution.

A digital twin technology is one emerging technique to mirror physical devices into the cyber world and then keep the real and digital worlds in synchronization, especially for industry and manufacturing. Regarding the industrial Internet of Things (IIoT) requirements of low latency and high reliability, IEEE Time Sensitive Networks (TSN) becomes one necessary and critical wired networking technology. This paper utilizes one famous digital-twin open-source project called Eclipse Ditto to monitor the TSN operations and provides the prioritized trans-mission capability. Based on our initial experimental results, adequate TSN policies can keep critical traffic flows as best effort as possible even in the competitive situation.

This study proposes an intelligent roadside system based on a V2X vehicle network and intelligent traffic signals to address the important issues of traffic accidents and congestion in traffic management. With the V2X vehicle network, vehicles can obtain real-time information from traffic signals and use better methods to control their speed and direction, thus reducing the occurrence of accidents and congestion. In addition, the system uses CNN technology to improve detection performance under required conditions, and YOLOv7 technology can accurately identify vehicles. With these technologies, the intelligent roadside system can more sensitively detect vehicle behavior on the road, better coordinate traffic flow, improve road traffic efficiency, reduce traffic congestion, and also reduce the occurrence of traffic accidents. Overall, this intelligent roadside system based on a V2X vehicle network and intelligent traffic signals is expected to play an important role in urban traffic management, effectively reducing the occurrence of traffic accidents and congestion, improving road traffic efficiency, and providing a safer and more convenient driving experience for the public.

The modern idea of gunplay has its origins in games and TV shows, and survival games have emerged as a new form of competitive sport. To assist users in survival games, a system has been developed to help them gather enemy intelligence and gain an advantage for their team during gameplay. The system is designed to detect the enemy's position and transmit the information to the user, enabling them to respond quickly to the intelligence and significantly reduce the risk of being ambushed from behind. YOLOv4 has emerged as a powerful tool for addressing simple but labor-intensive tasks across a range of industries. For example, in manufacturing, YOLOv4 is used to detect product defects and prevent defective items from reaching the market. Similarly, in facial detection, YOLOv4 is employed to compare facial features for purposes such as face clocking and criminal tracking. During the pandemic, YOLOv4 has been used to detect masks and social distancing in many manufacturing facilities. This modified YOLOv4 system uses real-time vehicle detection at intersections to regulate traffic flow and enforce speed limits.

In this paper, we design a high accuracy neural network based on SPVCNN, which combines voxel-based processing and point-based processing. The neural network can not only precisely segment the scene on the road but also capture small objects in the scene. The experiments show that the performance is excellent compared to other literature.

In this study, we prepared a potentiometric dopamine biosensor based on copper oxide (CuO) thin films modified with tungsten trioxide nanoparticles (WO3 NPs). We used the excellent anti-interference ability of WO3 NPs to modify CuO thin films. And the differential pair system is used to measure, and the measurement content includes linearity, sensitivity, response time, and interference influence. The sensor of this research project has the advantages of simple structure, easy manufacture, low cost and small volume.

A magnesium oxide (MgO) thin film glucose sensor was proposed, which used enzymes to modify the MgO layer and deposit MgO on a flexible printed circuit board (FPCB) by a radio frequency (R.F.) sputtering system. Due to the combination of MgO and enzyme, the sensor has an excellent glucose-sensing ability. Measured by the voltage-time (V-T) measurement system, the average sensitivity of the glucose sensor is 11.60 mV/mM, the linearity can reach 0.98, and the repeatability and selectivity experiments have verified that our glucose sensor has superior stability and anti-interference ability.

In this paper, an enzymatic potentiometric biosensor based on Nafion-modified indium gallium zinc oxide (IGZO) thin film has been developed for the measurement of uric acid (UA). Sensing characteristics, including sensitivity, linearity, response time, etc., are measured by a voltage-time (V-T) measurement system. The IGZO UA biosensor proposed in this study has a high sensitivity of 21.9 mV/(mg/dL) and maintains a sensing characteristic of linearity of 0.998.

This thesis proposes a system integration and optimized hardware acceleration design for the lightweight YOLOV3 model in the object detection network architecture, including the Convolution Layer, the Maxpooling Layer, and the Detection Layer, the hardware integration of the Shortcut layer, and the optimized implementation of the three output layers. In addition, this paper is verified and implemented in hardware on the Xilinx Zynq UltraScale+MPSoc ZCU102FPGA platform. The operating frequency is 180 MHz. The usage of bandwidth for the Convolution and Maxpooling Layer Fusion and Shortcut and Convolution Layer Fusion can be reduced by 85.33% and 45.27%, respectively. While optimizing Maxpooling Layer and Shortcut Layer, the running time is faster than ARM CortaxA53 15 and 26 times, respectively. Furthermore, the realization and the results of the system integration are exhibited through the HDMI monitor.

Winograd convolution has almost become a standard for reducing computation complexity for convolutional neural network. However, its efficiency is limited due to the surge in the number of additions. In this paper, we analyze the upper bound of addition count of Winograd convolution with fixed-point filters and an estimated arithmetic weight density, and then proposed a ternary-coded binarized (TCB) convolution that can eliminate all multiplications and reduce more than 99% of addition operations.

Most recent advances in artificial intelligence mainly focus on further acceleration, reliability and security in safety-critical applications including 6G StarLink, and automotive. We find that Winograd Convolutions possess both potentialities on acceleration and fault tolerance theoretically. Therefore in this paper, we firstly take the single-fanout property of Winograd products to develop an effective and efficient single-interpolation error model. Secondly, we improve the state-of-the-art interpolation-point selection algorithm by taking into the arithmetic weights of integer multipliers. And finally we propose some single-error detection and correction methods using redundant-interpolation techniques. From both derivations and evaluations, the proposed redundant-interpolation Winograd convolutions not only possess a beautiful and concrete theoretical basis, but also an error correction capability with only quite few area overhead.

This paper proposes a three-dimensional tree architecture. This architecture consists of 32 tree architectures. Each tree architecture is responsible for all operations of a kernel, so that each kernel can be processed in parallel. The inner product operation in each kernel can also use the characteristics of the tree architecture to achieve parallel operations. Operations in two different dimensions achieve the goal of parallel processing.

This study proposes a cognition-adaptive approach to enhance the safety of production line workers through IoT data by artificial intelligence. The study addresses the challenges associated with estimating mental states and focuses on the sequential faults made by production line workers, using a Hidden Markov model on top of a human cognitive model trained on operation logs from installed data recording hardware. The study asserts that merely retaining workers' attention is inadequate for efficient production tasks, and tracking the state of perception is crucial. Fatigue states are verified through photoplethysmography imaging (PPGi) and heart rate variation (HRV) methods. The PPGi technology has the added benefit of rejecting noise interference. The study introduces a robust fatigue index, verified through a simple facial video. Robustness is considered a significant contribution, as the containing algorithms are improved to resist random motion and to mislabel, which previously impaired performance. To avoid gradient step noise interference for the learning algorithm used in previous studies, we utilize a gradient-free Hamiltonian Monte Carlo step in error minimization. The proposed algorithm was tested on an assembly line with five workers collaborating. The experimental results demonstrate superior reliability in assessing a person's fatigue.

In stock market, the common way to find trading signals is using trading strategy that can be generated by technical indicators which is often used to predict trends of a target stock. To find useful technical indicators and parameters, the previous work presented an optimization algorithm based on the genetic algorithms and simulated annealing. It used return and risk as the fitness function to evaluate the goodness of a candidate solution. Observably, it is not enough. Hence, in this paper, we propose an improved optimization algorithm to find a suitable set of technical indicators and their parameters for identifying low-risk trading signals based on the previous approach. The main different between them is that we add the buying and selling ratio of the institutional investors to design the fitness function. Experiments were then conducted to show the effectiveness of the proposed approach.

Extending the lifetime of sensors is a crucial research topic in wireless sensor networks. Several charging methods, including mobile vehicle charging and environmental charging, have been proposed to address this issue. To ensure the sustainable operation of sensors, an optimal deployment strategy should consider factors such as target coverage, environmental charging conditions, and the length of the charging path required for mobile vehicles. Hence, a novel sensor deployment method that uses a genetic algorithm to maximize the environmental charging efficiency of sensors while minimizing the frequency and distance of charging vehicle visits is proposed in this paper. In the target field, a sensor is deployed for each target, and the sensor can obtain energy from the environment. This paper describes how to use a genetic algorithm to determine the deployment location of sensors and designs relevant objective functions and constraints. Through experimental simulations, it is demonstrated that the proposed deployment method can effectively reduce the frequency of charging required through charging vehicles.

In this paper, we propose a novel Prior-Guided Parallel Residual Bi-Fusion Feature Pyramid Network (PPRB-FPN) for accurate obstacle detection in unmanned surface vehicle (USV) sailing. Our method tackles the challenge of detecting small objects, which are prone to information vanishing. To the end, we leverage the PRB-FPN for small object detection and YOLOv7 as a single-stage object detector to effectively identify obstacles. Our experimental results on the Obstacle Detection Challenge dataset at the 1st Workshop on Maritime Computer Vision (MaCVi) demonstrate that our method outperforms both Mask R-CNN and YOLOv7, achieving an F_avg score of 0.514.

In recent decades, there has been biomimetic research focus on imitating the human olfactory system by the electronic nose. Rapid advances in the Internet of Things and artificial intelligence have made low-cost electronic nose feasible. But, arranging multiple gas sensors into rows and columns based on how they respond to the presence of gas could generate a great deal of high-dimensional data and easily cause the curse of dimensionality, making the late-stage AI model training more difficult. For this reason, this paper proposes to use the Max-Relevance and Min-Redundancy (mRMR) algorithm for feature selection of e-nose data and optimal feature subset selection for training AI models

This paper discusses the Design of Operation and Monitoring Architecture of the Content Delivery Network (CDN) service deployed in National Center for High-Performance Computing from 2021 to 2022. In 2011, we deployed an experimental node in Tainan. In 2022, we deployed two production odes in Hsinchu and Tainan. The CDN service is deployed for the research and education units in Taiwan. Furthermore, the CDN service can be applied in another research and development fields, such as Internet of Things (IoT), Big Data analysis. Taking advantages of CDN service, the web users can have better user experiences when browsing the websites of government and academic units.

In the early period of the COVID-19 pandemic, Taiwanese officials and people are doing their best to suppress the epidemic, leading to notable changes in the congestion levels of emergency departments. This phenomenon offered an opportunity to study the influence factors of emergency department (ED) crowding. In this work, we conduct data analytics on collected data from the ED of major hospitals in Taiwan. The data we analyzed comprise nonprivate information, ensuring the confidentiality of patients. Moreover, the data comprise several hospitals in Taiwan for a comprehensive view of the situation during the COVID-19 pandemic. We aim to provide valuable information on the appropriate timing of nonemergency patients' visits to the ED and to help nonemergency patients make informed decisions about when to visit the ED, ultimately improving their experience and the overall quality of medical care. The findings of this study have potential applications in developing intelligent systems or mobile applications that could offer valuable insights into optimizing nonemergency patient visits, thereby relieving the ED crowding problem.

Optimization to Refactoring reduces the queuing delay between the 5G core Network Functions and considers the allocation cost of 5G core Network Functions. This paper studies the 5G core Network Functions Refactoring Optimization problem. The problem is formulated as the Mixed Integer Quadratic Constraint Programming (MIQCP) and two-objective problem. We implement the model by using the optimization tool, GUROBI. Finally, we compare and analyze the performance of the other architectures that are designed by different 5GC refactoring methods. The performance result shows that 5GC refactoring by GUROBI can improve the performance, with less queuing delay and allocation cost.

HCI (Human-Computer Interaction) is a trend in technological art, it can build an immersive experience in an interactive field. This paper focuses on intelligently and dynamically capturing dancers' motion, quantifying the emotional representation of dancers. We define the gesture's emotional type (positive and negative), and the gesture's definitions given are expansion and contraction of the limb. Finally, the interaction between the 3D coordinates and the movement of the robotic arm can be captured by the pose of the dancer.

This paper provides a system that balances speed and accuracy to recognize students' behaviors. This system consisted of a skeleton extraction method and two behavior classification models. This paper used two stage classification models (CNN+SVM) which is new and different from other papers. We also make some modifications to the input of skeleton information so that they can represent the differences between similar behaviors more effectively.

This research develops an intelligent automatic production loading and unloading solution for a multi-process part. According to the posture required for each processing operation of each part, a novel gripper and carrier that can implement multi-position clamping are designed. The Beckhoff controller is adopted as the host controller in the control system to integrate the robotic arm (including the vision system), the automated guided vehicle (AGV), and the computer numerical control (CNC) machine. The main topic of this research is to meet the requirements of automatic loading and unloading in the processing of the entire production line. It will significantly reduce the workforce consumption of the processing production line and avoid human error, thereby improving processing reproducibility and product production stability.

This paper proposes a wire break prediction and avoidance method for wire electrical discharge machining equipment. This method enables the equipment to automatically adjust the machining parameters based on the predicted wire break results, thereby saving the time and effort required for automatic or manual threading after wire break. Ultimately, the goal of achieving equipment intelligence and improving production efficiency can be accomplished.

Following the rapid development of smart agricultural technologies and applications, many smart mobile vehicles have started to be used in crop fields in order to reduce the labour pressure and burden on farmers. The pineapple crop is particularly important in the central and southern regions of Taiwan. In order to meet the need for pineapple production throughout the year, artificial flower induction is used. In this paper, we propose a prototype system for smart mobile agricultural vehicles for the flower induction process in pineapples. The prototype system combines a deep learning image recognition system with the YOLOv5 framework for the identification of pineapple flowers, to help farmers take advantage of smart mobile vehicles for the placing of calcium carbide water.

Human motion recognition technology has started to be applied in many important research fields, for example to assist inpatient rehabilitation and for the perception of dangerous activity. Hence, improving the accuracy of human motion recognition is a very important research topic. At present, human action recognition systems have only been developed for actions by a single person. In this paper, we propose to use LSTM technology for multi-person human action recognition model training and identification. We compare the performance of our approach with that of the ST-GCN human action recognition framework, and find that it improves the recognition accuracy by 33.33%.

Although the operation of an honest store can reduce the pressure on farmers' selling costs, the quality of fruit will gradually decline as the storage time increases. The sales of most fruits can be promoted by reducing prices to stimulate consumers to buy and enjoy them. Since the current selling prices of fruit products in an honest store cannot be dynamically adjusted based on their appearance and quality, this paper proposes the design of a smart fruit quality recognition system, and describes the implementation of a prototype system. The quality recognition system uses a NVIDIA Jetson Nano embedded device as a system integration platform, which is placed in the honest store. Consumers place the fruit they want to buy in the evaluation area of the quality recognition system to obtain a suggested price. The fruit quality recognition system uses a webcam to capture real-time images which are passed to a deep learning image recognition model to identify the quality of the appearance and to complete the proposed price adjustment. This approach can dynamically adjust the price of fruit based on its appearance and quality identification, which can reduce the number of times farmers need to visit the honest store to modify the price.

With the prevalence of deep learning technology in various application domains, a recent trend is to deploy convolutional neural networks onto the sensors in the field enabling new smart applications, such as poaching detection, discovering endanger species, and estimating wildlife populations in environmental monitoring. Among the ultra-low-power hardware platforms with microcontrollers for the computations, the sub-mA platforms (i.e., lower-end ultra-low-power platforms) would be desired for monitoring in the field since they can achieve a longer lifetime given a fixed amount of power budget provided by the batteries. In this work, we propose a new approach to develop convolutional neural networks to run on such sub-mA platforms, usually with less than 10 KB memory for keeping the runtime data. The new approach combines the concepts of the One-vs-All (OVA) strategy and the transfer learning technology to build networks with small memory footprints. Our experimental results show our approach achieves the accuracy of 77.0% of a four-class classification problem for a poaching detection application as fast as 1.1 seconds.

In the context of computer numerical control (CNC) machinery, fault diagnosis traditionally involves complex formula conversions to extract characteristics and categorize faults. However, such a method is unsuitable for hardware implementation due to high resource usage. This paper proposes a convolution neural network (CNN) approach for fault classification and hardware acceleration using ternary quantization and batch normalization techniques to reduce data access for weights and improve accuracy. The proposed CNN hardware accelerator is implemented on FPGA (VC707) and reduces memory usage by 83.8% compared to floating-point operations. Furthermore, the proposed method achieves 97.6% accuracy in CNC machinery fault classification.

In recent years, the development of IoD (Internet of Drones) has become more and more important, since multiple drones can accomplish much more tasks than single drone. However, most drone control software requires one operator for one drone, which is not scalable. In order to control multiple drones efficiently and reliably, we uses AMQP broker for connecting multiple drones and ArduPilot for autonomous flight of a single drone. Furthermore, we designed a drone control GUI that allows the operator to easily switch controls between drones while watching all streaming videos at the same time.

Drones have been widely used in many applications in recent years. We have designed a drone specifically for the police and security guards. Our drone has the following features: First, all drones are connected to a server using IoT technology, and users can easily control drones with a keyboard. Second, a siren is installed at the bottom of the drone to warn criminals. Third, the drone can be equipped with four cameras to provide panoramic video streaming. All the components are off-the-shelf or 3D printable. The drones are called Guardian Drones

A farm field monitoring system was developed to solve two problems of a tea manufacturer. The system allows the manufacturer to check the environmental information of the field in real time. We also analyze the acquired data and discuss the relationship between the data and the problems in the tea fields.

This paper presents Ureka, a decentralized and secure open system for managing personal IoT devices. Ureka comprises three key components: a personal private system, a blockchain smart contract accounting system, and a ticket module for IoT device access. The personal private system provides a secure means for individuals to manage their personal data and identities, while the smart contract system audits access records. The ticket module ensures secure access management. By adopting a decentralized approach to personal and IoT device management, Ureka offers enhanced security and privacy compared to existing proprietary systems.

Recently, many service providers or telecom operators provide the services of hybrid cloud with edge computing to consumers, such as gaming, streaming, or manufacturing. In the global ESG policy, the energy management of the data center is emphasized for enterprises. Therefore, we propose the energy management architecture for hybrid cloud with edge computing including physical and virtual resources. and propose the energy saving policy to improve resource utilization and cooling efficiency.

Due to the limited physical space and training facilities, we propose one efficient immersive training method to integrate a virtual reality (VR) simulation system with a body area network (BAN). With the Customized deep neural network algorithm, the body-worn inertial sensors are capable to recognize the activities of participants and avoid mismatched actions. Moreover, the neural networks have been utilized to provide greater access to physical actions of the VR real-time training environment. In this paper, a quaternion based deep neural network algorithm is developed and implemented for human activity recognition (HAR). We share the experience on the VR application that has the potential to fulfil multi-user immersive VR system on HAR.

Since short-distance transmission technologies are susceptible to interference from obstacles, which may not be suitable for indoor communication, the robustness, reliability, and pervasive coverage capabilities of the Long-Range (LoRa) transmission technology may be an alternative to tackle this problem. By combining the strengths of both TDMA and FDMA scheduling schemes, this work establishes a hierarchical network to collect data in indoor environments via a tree-based network topology with End Nodes, Routers, and a Gateway automatically. Based on the hierarchical network architecture, a distributed cooperative scheme is further developed to provide indoor positioning service via received signal strength indicator (RSSI).

This paper proposes a dual-sensor fusion system that combines a Frequency Modulated Continuous Wave (FMCW) radar and an Intel RealSense RGB+Depth (RGBD) sensor as non-contact physiological signal sensing devices. Through the fusion of point clouds obtained from the dual-sensor system, the spatial information and vital sign signals can be perceived.

The escalating popularity of fitness, driven by advancements in fitness-related technology, has led to a rising incidence of over-exercise, resulting in an array of health concerns including injuries and burnout. Current fitness tracking apps and wearables focus mainly on documenting past workouts, neglecting the forecasting of future activities, thereby inadvertently promoting over-exercise. To counter this issue, we propose a novel application named Proactive Over-Exercise Monitoring (POEM). POEM goes beyond merely recording past workouts and proactively considers future workouts, alerting users about the likelihood of over-exercise. It derives daily workout duration from multimodal data, incorporating the user's exercise history, location details, and daily schedules. Through its proactive alerts on potential over-exercise, POEM seeks to instill safer and more sustainable exercise practices.

Mental health disorders is one of critical illnesses that an individual tries to avoid talking about it. Thus, many people would seek information about mental health disorders by searching the information on Webs through available search engines. Despite the fact that the search engine is able to retrieve, these resources are normally scattered. The aim of this paper is to propose a chatbot so-called mentBot that is able to function as a repository for mental health disorders. Human users can chat and retrieve specific information from the mentBot without being judgmental. This paper presents an architecture of mentBot and a step by step process on how to design and implement mentBot.

Virtual reality (VR) is a simulated experience in a three-dimensional (3D) computer-simulated world. The experience could be similar to real life or completely different. Despite the fact that VR has been cited as a fantastic choice for teaching and training, the existing publications do not specifically demonstrate how to construct training application with VR technology. The purpose of this paper is to present our personal experience in designing and implementing First Aid application on Oculus Quest.

aching at the right moment can be effective in communicating with children to correct wrong behavior. In some cases, such teachings could even be more effective when they occur later in the day when emotions are calm. This study presents an innovative web-based system that allows parents to generate a print bedtime storybook that addresses timely issues through children's favorite characters, helping parents to use the right material to teach at the right moment.

Optical coherence tomography (OCT) is a non-invasive medical imaging technology that captures high-resolution images of biological tissue in real-time. However, OCT images are often plagued by noise, artifacts, and poor contrast, which can hinder their clinical utility. The study explores various image enhancement techniques using Python and OpenCV to improve the quality of OCT images. The study begins by discussing the basic principles of OCT and the challenges associated with OCT image processing. Next, the demonstration on how to preprocess OCT images by removing noise and using various filters and denoising algorithms. Finally, the study was demonstrated to real-world OCT images and evaluated their effectiveness with the aid of a developed interface.

This paper proposes a medical visual question generation model for generating higher-quality questions from medical images. The visual question generation model can guide the diagnostic process and improve the utilization of medical resources by reducing the dependence on physician involvement. Our model uses cross-attention and the large language model to preserve inherent information and addresses the issue of poor results in the medical domain due to limited data. We also control the category of generated questions by setting guidance sentences that include interrogative words. The experimental results demonstrate that our method generates higher quality questions than previous approaches.

We propose a method to estimate amyloid-β positivity from axial slices of Quantitative Susceptibility Mapping (QSM) considering patient age. Although QSM has received much attention for the early diagnosis of Alzheimer's disease, the slices reflect not only iron deposition derived from Alzheimer's disease but also physiological iron deposition due to age. Our method uses both QSM images and patient age to train the model, and thus alleviates the negative effects by referring to iron deposition in patients of similar age. Experimental results show the effectiveness of using both QSM slices and patient age.

With the development of the Internet of Things, applications of the wireless sensor network are more and more widespread. In wireless sensor networks, the sensor is basically stationary and performs sensing tasks at the deployed location. The mobile charger moves to the sensors for recharging them. However, the energy of the mobile charger is limited. It usually consumes lots of energy for moving and recharging. As a result, some sensors die due to the lack of timely recharged, which affects the monitoring quality and recharging efficiency. In this paper, the sensors are mobile. They can move to the recharging station in order to be recharged by the mobile charger, which reduces the cost of the mobile charger's traversal. At the same time, the proposed algorithm considers the cooperation of multiple mobile chargers, aiming to shorten the recharging cycle. In addition, the sensors also locally adjust their sensing frequencies for improving the sensing data quality according to the starting recharging time. Experimental results show that the proposed recharging algorithm improves the sensing quality compared to the existing algorithms.

Traditional facial recognition has been widely adopted in the technology industry, and its recognition accuracy has reached its peak. However, it is still difficult to implement a cost-effective attendance system using low-cost webcams in a small classroom measuring approximately 8.5m*6m, which can accommodate 20-25 people. The main reason is that webcams are designed for personal use; their facial recognition distance is only about 2 meters, which cannot recognize the face of every person in such a classroom. In this study, we used YOLOX for small face detection, enhanced the features of distant small faces through Super-Resolution(SR), extracted facial features, and finally used Dlib for facial recognition. We also used OpenCV's optical flow to track the movement of the person's body frame. Therefore, even if the person looks down or turns around, our system can still determine whether the person is still in the classroom. Experiments have shown that this system can extend the webcam range to about 4-5 meters. In addition, under clear facial conditions, this system can also recognize whether the person has closed their eyes, as a reference for classroom participation.

Low-resolution infrared (IR) images potentially provide information for estimating social distance between persons without capturing personally identifiable information. In this study, we have implemented a social distance estimation system using Boosting-CNN (AdaBoost-based convolutional neural network). Furthermore, the system is embedded on a low-power microcontroller. We have compared the inference time between Boosting-CNNs and single CNNs processed on the microcontroller, and have found that specific Boosting-CNNs are faster than single CNNs. We have also constructed a Boosting-CNN using CNN models with different parameters, and it has been observed that the accuracy is improved by the order of training.

In the current era, there has been rapid advancement in the field of information technology. When scholars compose their research papers, it can be challenging for them to conduct an exhaustive examination of the existing literature. Therefore, it is essential to create a system that can suggest suitable citations to researchers during the writing process. This study aims to recommend relevant citations to researchers by utilizing weighted bag of words and BERT models. The proposed mechanism provides the following benefits: (1) Clear and transparent representation of document vectors; (2) A range of diverse natural language processing techniques, and (3) Avoidance of redundant citation suggestions.

The problem of floating debris in rivers and oceans is growing. To clean floating objects on the water more effectively, IoT-based unmanned boats were chosen for autonomous cleaning. However, the strong light reflections of riverside objects on the water surface pose challenges for vision-based object detection systems to detect small targets. By modifying the pooling module in Spatial Pyramid Pooling and using the TS-YOLO structure to retain the original spatial pyramid advantage, we improve the accuracy of floating litter for detecting objects on rivers. In the experimental results, our proposed method was tested on Pascal VOC, FLOW, and WIDER FACE, which showed good detection capability on mAP with 2.86%, 1%, and 2.28% improvement over the original YOLOv4.

Due to the complex track environment, the use of only one stage for track component identification is prone to inaccurate component positioning, resulting in misjudgment of the target coordinate system and other problems. This paper proposes a two-stage recognition method based on YOLOX-nano and ResNet34, hoping to solve the problem of inaccurate component positioning in the existing classification system and also improve the recognition accuracy. In the first stage, the entire image is preliminarily screened through YOLOX-nano, so that the system can understand the image structure, obtain the possible range of components, and then obtain the leftmost and rightmost positions of the track through Hough Transform. Next, calculate the intersection with the sleeper range obtained in the first stage, and calculate the possible relative position of the component base on the intersection, thereby locking the range where the component is located, and handing this range to ResNet34 in the second stage for component defect detection.

Performing machine learning tasks on low end devices enables the development of various smart applications. Especially, these low end devices are often equipped with ultralow-power microcontroller units (MCUs) that have weak computation power and few memory resources. It is a more challenging work to put these machine learning tasks on those end devices powered by harvested ambient energy, which are often referred to as energy-harvesting (EH) devices, since the unstable ambient energy can lead to the execution failure of the machine learning tasks. This paper proposes an adaptive energy-aware design to coordinate multiple EH devices to accomplish multi-class classification computation. It also leverages the concept of the One-vs-All (OVA) strategy turning a multi-class classification into multiple binary classifications. The experimental results show our work performs better than the widely used round-robin policy and self-greedy policy in consideration of time and energy consumption.

This study proposes an innovative stacked battery management system (BMS) architecture for monitoring 20s lithium titanate oxide (LTO) or lithium batteries, which can be used for applications including 48V energy storage systems, multi-cell electric vehicles, and ultra-high voltage energy storage systems. The proposed stacked BMS uses the BQ76942 series Battery Monitor and Protector IC produced by TI and implements a low-side MOS control architecture, which offers safety and cost advantages over the common high-side architecture. This research demonstrates the feasibility of using BQ76942 to realize a generic stacking architecture BMS for multiple cell applications, which can also be used for different rated voltages of lithium batteries.

The design of a distributed power flow system (DPFS) is vital to focus on how to optimally assign the intermittent renewable energy (RE) resources to the unpredictable power loads with the integration of energy storage system (ESS). This is because the renewable energy resources such as solar photovoltaic, fuel cell, and so on are strongly depending on the environmental changes and the power loads are tightly operating according to the unplanned human activities. Thus, it is a highly significant challenge to design the energy storage system efficiently in terms of minimum energy loss and optimum storage capacity before the deployment phase. In this paper, the DPFS that incorporates an optimization approach has been investigated with real experimental dataset of smart home environment by considering two power storages and two home appliances based on human daily activities in both summer and winter seasons. Numerical simulation results reveal that the energy loss and the optimum storage capacity varies greatly depending on the algorithms of the power flow assignment and load shifting in the presence of power fluctuations from both power generations and power loads.

The integration of RESs such as wind generation system and photovoltaic generation system increases the risks of power fluctuations from the supply side. The uncertainty due to seasonal data, temperature variations, and consumer activity increases the risks of power fluctuations on the demand side. To accommodate the uncertainty of both generator side and consumer side, the power grid has to increase its ability to accommodate fluctuating power. Based on this, the robust energy balancing between power generators, consumers, and storage devices in a periodic operation is introduced. This paper proposes the minimum requirements on controllable generators, loads and storage devices for supply-dominated energy balancing of a system which includes fluctuating generator and load. The discussions on the periodic operation of the system enable us to apply the result to a long-term system operation.

The proliferation of distributed energy resources (DER) and the widespread adoption of the Internet of Things (IoT) in Prosumer Nano-Grids (P-NGs) pose severe challenges in existing energy management schemes, thus necessitating the need for the development of novel solutions which can guarantee system efficiency and security. To address such challenges, we propose a novel scheme leveraging the benefits of distributed power flow coloring, blockchain technology, and Transactive Energy Management (TEM). The proposed system model offers P-NGs a secure, transparent, trustworthy, and efficient energy management and transactions platform that can: maintain the overall voltage stability and specified power ratios against power fluctuations within the P-NG, offer an increase in energy security, provide real-time monitoring, and reduce energy waste.

Green energy sources, including wind turbines and solar generating systems, play a significant role in power systems due to their low environmental impact. However, their generated power is highly variable, posing an unpredictable danger of power fluctuations. As a result, power fluctuations can affect the quality and stability of the power grid, hence making the power imbalance a formidable problem. Traditional control and management techniques have many limitations in measuring, transmitting, and controlling power data between power devices. To achieve power balancing in each time instance, artificial intelligence (AI) techniques are needed to realize truly real-time power control. This paper promotes the application of distributed AI framework for a power system with renewable power sources, loads, and storage devices to achieve power balancing.

One of the characteristics of electric vehicles (EVs) is that they are capable of providing Vehicle-to-Grid (V2G) service. However, most V2G research considers only passenger cars, which typically already have their own usage purposes and different travel schedules. In contrast, electric school buses (ESBs) are used at specific times and are often parked in idle mode. Because of that, ESBs are more practical for providing V2G services when the grid sends incentive demand response signals. In this paper, we study V2G model considering electric school buses by formulating the problem as Stackelberg game. The model considers the utility company (UC) as the leader who determines the optimal incentive price for the demand response event, while the ESBs in various schools within the area act as the followers trying to find the optimal amount of discharged energy from their battery. Furthermore, the existence and uniqueness of the Stackelberg equilibrium are also proven and guaranteed. The results show that as the number of ESBs increases, the optimal incentive price offered by UC decreases.

The study aims to improve brainstorming by using Generative Artificial Intelligence (Gen AI) to generate more ideas and increase novelty. A comprehensive literature review was conducted to examine the rules and techniques of brainstorming. An experiment was conducted in which Gen AI was added to improve the brainstorming process. The results showed a significant positive correlation between participants' confidence in the information provided by Gen AI and an increase in the number of non-redundant ideas generated by co-ideation. In addition, participants' expectations of co-ideation with Gen AI were significantly positively correlated with an increase in both non-redundant and unique ideas. Gen AI can be an extension of human thinking in brainstorming, leading to improved idea generation. However, the effectiveness of Gen AI depends on the user's attitude towards AI.

A Tablet PC is now used not only for reading books and watching videos, but also for various work tasks such as creating business documents and analyzing data. However, while graphics performance and other specifications of the tablet PC have improved over the years, the basic application usage is similar on all devices, that is, a user maximizes any application window and uses it in full-screen mode. There is the question of whether the maximized single window is truly appropriate for performing various tasks on the tablet PC. In fact, multi-window environments facilitate multitasking on desktop or laptop PCs. We analyzed the usability of the multi-window environment on a tablet PC through experiments in which subjects performed simple tasks while switching between two or more windows. The results revealed that while the multi-window environment has great potential to perform multitasking on a tablet PC, the ability to automatically adjust the positions/sizes of multiple application windows is needed.

Handwritten messages can often impart personal touch to the text. However, this personal touch cannot be achieved in computer-mediated communication. Although features such as emojis, emoticons, or stickers can help enhance the expressiveness in the computer-mediated messages, they cannot lend the messages a unique individuality. In this study, we propose a chatroom website utilizing a person's handwritten font that can automatically select different glyphs for the same character, thereby creating the perception of human handwriting and providing a personal touch to digital communication.

Interactive picture books have become popular reading materials for preschool children in recent years, as they employ digital technology to present content in a multi-sensory way and encourage children to interact with the materials. In this study, we observed the behavior of 5-year-old children while they interacted with interactive picture books. Our findings suggest that children prefer dragging than tapping in the interactive picture books used in our study, and these features may have a potentially positive impact on story memory which we suggest further studies. Based on the analysis of children's behavior in interactive picture books and relevant research, this paper proposes issues that should be noted in the design of interactive features in interactive picture books.

Maintaining real-time communication quality in metaverse has always been a challenge, especially when the number of participants increase. We introduce a proprietary WebRTC SFU service to an open-source web-based VR platform, to realize a more stable and reliable platform suitable for educational communication of audio, video, and avatar transform. We developed the web-based VR platform and conducted a preliminary validation on the implementation for proof of concept, and high performance in both server and client sides are confirmed, which may indicates better user experience in communication and imply a solution to realize educational metaverse.

Nowadays, elementary school students are so good at using technology to learn that even the subject of dance can be taught with the help of technology. In the past, many scholars have shown that using Kinect as a teaching material can help students learn dance, especially in the inspiration of movement. However, different students have different needs for using Kinect as a teaching material. Thus, some researchers have developed DANZOO teaching materials based on Kinect technology to investigate the needs of elementary school students. It is expected to understand how different groups of elementary school students have different needs and demands on DANZOO. The conclusions are proposed for future reference in the design of educational materials based on Kinect technology.

We study the use of PPO (Proximal Policy Optimization) algorithm for trading ETFs belonging to different asset classes. The studied asset classes include common stocks, bonds, REIT (Real Estate Investment Trust), gold, and future contracts on agricultural commodities. When properly training the PPO agent with the proposed ratio allocation strategy, the agent outperforms the static, periodical rebalancing approach.

The prevalence of mobile phones and messaging apps has made smishing attacks a growing concern. Attackers use smishing to deceive individuals into revealing sensitive information, downloading malicious software, or performing other harmful actions. These disregarded responses cause significant financial loss to individuals or organizations. We proposed using a machine learning model to analyze message contents and detect phishing URLs. Our experiments collected and manually annotated text messages from mobile phones, and both natural language processing techniques and machine-learning algorithms were applied to the dataset simultaneously. The results show that our proposed method achieved an F1 score of 94%, showing high effectiveness and efficiency in accurately detecting Smishing attacks.

In this paper, we will discuss the challenges of Financial-grade API, focusing on prevention mechanisms for "Cuckoo's Token Attack" and delve into the current security authentication mechanisms of OAuth 2.0, FAPI 2.0 and OpenID Connect. Finally, we found that the OpenID Connect method can mitigate the attack by using a one-time token and issuer identifier, but a one-time authentication method without the OpenID Connect requirement is also needed.

Acceleration in technology transformation for financial sectors has been impacting the way how relevant institutes should store and protect their data and resources from untrusted entities. Zero Trust as an enhanced security model in which every access request must be verified on a per-session basis is believed to be an effective way of providing fine-tuned and granulated secure access control compared to the traditional network perimeter-based security model, where approval of access requests is based on the network location. In this paper, we proposed a machine learning-based trust management model that is integrated into a Zero Trust framework, aiming to develop a practical example specifically for financial sectors by offering a guide on how trust management could be handled in the zero trust architecture.

Cryptocurrency is a blockchain application that has attracted the attention by many users. The proof-of-work consensus mechanism is adopted by several cryptocurrencies, such as Bitcoin, Litecoin, and Dogecoin. This work discusses how miners choose mining pools (MPs) and how MP operators adjust fees to maximize their revenue. This work proposes a mining difficulty adjustment method and a randomly distributed round-batch computing power competition algorithm for mining simulation. The simulation results show that by increasing the MP fee, the large MP can increase the revenue by about 60%, and the ability to withstand the risk of changes in computing power. Small MPs can sacrifice about 35% of the revenue in exchange for the willingness of miners to stay by lowering the operation fee.

Traditional machine learning typically requires the training dataset to be stored locally on a machine or data center for training. However, this approach may have issues related to data privacy and security. Therefore, Google proposed the concept of federated learning in the same year. However, federated learning may have problems like Non-Independent and Identically Distributed (Non-IID) data and private security. Non-IID can lead to lower training accuracy than expected, and there may be a risk of privacy leakage in the data uploaded by clients. Therefore, this paper proposes a Clustered-based Hierarchical Federated Learning Framework with Differential Privacy and Secure Aggregation (CHFDS). Before training begins, we cluster all clients so that the data distribution between clients in each group is similar. This means only a random subset of clients from each cluster is selected in each training round instead of all clients participating in the training. We can use this method to adjust the data balance of participating training. Finally, we add differential privacy and secure aggregation to the clustering and training process to improve the privacy and security of the proposed clustered federated learning framework.

The paper realizes a 12-bit 1Gs/s segmented digital-to-analog converter (DAC), with a six-bit thermometer code and six-bit Switching-Minimized Monotonic (SSM) code. The six most significant bits (MSB) are converted into thermometer codes. In order to reduce chip area, a minimized Switching Monotonic code is proposed for the six least significant bits (LSB). A more flexible coding method improves the issue of excessive switching of binary codes. It can achieve better linearity. By considering the performance of low power, static performance, and small die area, the 6+6 segmented architecture is applied with the proposed coding. The proposed design is implemented by TSMC 90nm process, and the sampling rate and resolution are 1Gs/s and 12 bits, respectively. The proposed design consumes a power of 23mW. The simulated results demonstrate that the proposed coding technique improves the overall linearity and dynamic performance.

Device layout is very important for one Integrated Circuit (IC) performance. From different metal routing layouts, ICs can own different achievements. In this study, simple device layouts with only two poly fingers are discovered for obtaining the optimizing device characteristics. In order to create the uniform current distributions, more metal routing segmentations are preferred, but many metal segmented amounts will induce less effective metal widths. Through the simulation tool, optimizing metal routing can be achieved. Finally, one suggested metal routing method for normal applications can be obtained from many simulation results.

On the one hand, Silicon Controlled Rectifier (SCR) is often applied for Electrostatic Discharge (ESD) protections. Since SCR owns the latch-up phenomena, SCR holding voltage is required to be larger than the operation voltage. On the other hand, direct or parasitic bipolar transistors are more frequently than SCR to be used as ESD protections. Similar requests for larger holding voltages are often proposed for bipolar transistors. SCR owns the positive feedback circuit characteristics so it owns latch-up properties. However, bipolar transistors don't have the positive feedback circuit so the latch-up event should not occur. The latch-up characteristics of SCR or bipolar transistors are discovered in this study through the silicon verifications.

In this paper, a 0.18μm high voltage BCD process is used to design octagonal Lateral-diffused MOSFET(LDMOS), with a modified width of the DPW layer in the drift region. The width of the DPW layer of the reference device is designed to be 36µm, and the DPW layer is designed to be reduced by 12µm each time in the direction of the drain terminal. It is found that with the reduction of DPW layer, the trigger voltage (Vt1) increases from 58.805 (V) to 99.8 (V) when the DPW layer is 24µm (DPW24), an increase of 69.7 %; and the secondary breakdown current (It2) increases from 2.067 (A) to 4.486 (A) when the DPW layer is completely removed (DPW00), an increase of 117%. Finally, the best FOM value from the TLP test data was found to be the DPW00 device.

In this paper, high-voltage LDMOSs via a TSMC 0.18-μm BCD process are used to enhance ESD discharge capability of nLDMOS devices. Schottky diodes and SCRs are parasitic at the drain side, which take advantage of the Schottky barrier property to increase the series conduction resistance at the drain side, and the low conduction resistance characteristic and strong discharge current capability of the SCR. It can be found that as compared with the nLDMOS with fully parasitic SCR, the holding voltage of adding both Schottky diodes and SCR increases from 16.66 V to 21.81 V by 30.9% increasing and retains the same good secondary breakdown current It2 of 9 A.

Transmission Line Pulse Generator (TLPG) is often applied for Electrostatic Discharge (ESD) discoveries. Since this machine is designed from the transmission line, unlimited to the common voltage source, it is often applied for Devices' electrical Safe Operation Area (eSOA) measurements with high currents. Different transmission line pulse widths can result in different device data. HBM event is often with 100 ns pulse width, but is 100 ns suitable for eSOA? Through silicon data analyses, the pulse width equal to 200 ns is proposed for Large Array Device (LAD) characteristics measurements in this study.

Dance has played a major role in expressing the characteristics of each country and culture. It is necessary to have a teacher who can teach dance accurately. This paper proposes a system that can analyze skeleton data from recorded videos and calculate similarities between teacher's and student's motions. The system extracts the positions of the joints of each dancer in the video using OpenPose and identifies how different are the dance movements using dynamic time warping (DTW), Pearson correlation, and time lagged cross correlation (TLCC). The results obtained in this study show that the information of skeletons identified from recorded videos can be accurately extracted and the similarity between the teacher's and student's motions can be determined for scoring the dance quality.

To assist self-study of C programming, we have proposed three types of introductory exercise problems: grammar-concept understanding problem (GUP), value trace problem (VTP) and phase fill-in-blank problem (PFP). GUP and VTP aim at code reading study whereas PFP aims at code writing study. Students can solve them using the browser-based answer interface that has the automatic marking function. In this paper, we present their application results to first-year undergraduate students taking a C programming course in Nihon University, Japan. The same instances were assigned twice this time, where they were assigned once in the previous year. The results show that the correct answer rate was improved for any type as the course elapsed, where that for PFP was no sufficient. Thus, the effectiveness of the proposal is confirmed for basic C programming study.

JavaScript is the lightweight and interpreter-based programming language. Since it is essential in web application developments, the popularity is increasing day by day. To assist self-study by novice students, we have developed the JavaScript programming learning assistant system (JSPLAS). JSPLAS implemented the grammar-concept understanding problem (GUP) for the first-step study of basic concepts and client-side developments. However, it does not cover the server-side developments, although they are also essential in web application developments. In this paper, we extend Grammar-concept Understanding Problem (GUP) for web server-side JavaScript programming. A GUP instance consists of a source code and a set of questions describing the definitions of the key grammar concepts in the code, where the answer is the keyword itself. The string matching technique is used for the correctness of the answer from a student. For evaluations, we generated 10 GUP instances by selecting 55 keywords and the corresponding questions for them, and assigned them to 41 students in Japan and Indonesia. Their solution results confirm the effectiveness of our proposal.

This paper discusses a method for detecting gas leaks in collapsed buildings in disaster areas using multiple micro-drones. In this paper, we develop our own drone with a mass of less than 100 g equipped with a gas sensor and con-duct gas detection experiments in an experimental environment. Computational Fluid Dynamics (CFD) is also used to visualize the air flow around multiple micro-drones. The results of the experiments and numerical simulations will show that leaking gases can be detected using gas sensors mounted on the drones by actively rolling up the gases with the propellers of the drones.

This paper proposes a content-sharing method to operate pure P2P at the lower network load by porting the functions in the CCN router that are operated in the network layer to the application layer where the conventional P2P works. Furthermore, the consistency control function using the timestamp of the content updates is added. The effectiveness of the proposed method is evaluated using computer simulations.

This study developed a framework based on deep learning algorithms for processing 360° videos for detecting pedestrian spatial behavior in urban spaces. Information divergence is determined through the sampling and conversion of spatiotemporal behavior data for pedestrian flow analysis. Traditional videos, such as those captured by one-way security cameras, cannot be used to fully analyze the flow of pedestrians in cities. Therefore, a 360° camera is used to capture panoramic videos of city spaces over time. Subsequently, deep learning algorithms are used to process the videos and obtain pedestrian trajectory data for analyzing their spatial behavior and interactions. The results of real-world implementation indicate that the proposed method and analytical framework can be used to detect pedestrians and collect data related to pedestrians' spatial behavior. However, the sampling rate and application of pedestrians' trajectory data must be explored in future studies.

Night markets are usually crowded and highly complicated in terms of spatial layout, making it challenging for visitors to navigate and find what they are looking for. This study is to develop a map-based application for addressing the difficulties faced by visitors when navigating night markets. This innovative application uses a map-based interface, integrated with the data of mobile positioning data (MPD) and space syntax (SS), to provide three different types of wayfinding information: 1. the locations of shops, stalls, and other amenities; 2. real-time crowd density; and 3. spots that can potentially cause disorientation. The application helps visitors plan their route, avoid crowded areas and enhance wayfinding experience, thus increasing the overall enjoyment of their visit to the night market. With an easy-to-use interface and the support of empirical evidence, this application not only can provide a novel solution to wayfinding problems in night markets, but also can enhance the visitors' satisfaction and further lead to better business outcomes.

In this work, we propose a method for road step detection via a 2D LiDAR and a binocular camera on the mobility scooter. Some visual experimental results can verify the correctness of our results.

As Virtual Reality (VR) technology becomes more advanced, VR can create scenes that closely resemble the real world. Therefore, automotive industry has started to use VR environments as test bed for autonomous driving algorithms. However, to the best of our knowledge, there is no systematic evaluation for the performance of real-world object detectors in VR worlds. To solve this issue, we used Unreal Engine to construct a virtual environment simulating real street view and AirSim plugin to generate bounding box ground truths. We employed the state-of-the-art object detectors YOLOv7 and YOLOv8 to detect common traffic objects. Our experiments can reveal the strengths and weakness of each detector.

Environmental issues have received much attention, especially the issue of sewage treatment. Many people do not understand how to deal with the sewage generated from the house cleaning. The lack of knowledge about water resources will bring negative impacts on the environment. In order to let the people to understand the methods and processes of sewage treatment, most of the water resources recycling center (WRRC) are open to the public. However, most of the visitors are conducted in the form of one-way explanations, slideshows, and filling in study sheets. There is a lack of interesting interactions and impressive guides, which cannot arouse the interest of the public. In order to make the visiting process more interesting and let visitors easily learn the professional knowledge of sewage treatment, this study developed an interactive sewage treatment learning game system. The game uses somatosensory detection such that the visitors can learn about sewage through the game to achieve the effect of entertaining. The game can give the visitors the correct knowledge of water resources and sewage treatment.

This study presents the design of a portable PPG and ECG synchronous acquisition system for non-invasive blood pressure measurement. The synchronous acquisition of pulse wave signal and ECG signal is achieved by designing PPG circuit and ECG circuit, and the integration of data and Bluetooth sending of data are realized by nRF52832. Finally, the collected data are input into the trained classification model for BP prediction, and the device achieves wearable, synchronous and real-time acquisition. The experimental results show that the signals acquired by the device can be used for non-invasive continuous BP prediction.

This study proposes a design of portable EEG acquisition equipment for collecting SSVEP signals. The EEG acquisition hardware is mainly designed, and the hardware interface can be connected to the EEG electrode cap. It can record the EEG signals of eight channels of SSVEP completely and stably. The hardware circuit design realizes signal acquisition, conversion and data transmission. The sampling rate that can be set in the software program is 250Hz to 1KHz, which can be used for this device. The device meets the practical requirements of small, handheld and real-time acquisition. The experimental results show that the device can provide high-quality EEG signals.

An electrocardiogram (ECG) is a low-amplitude biological signal, and it is easily contaminated by external biological noise. Removing noise from ECG signals is a fundamental step for ECG signal analysis. Therefore, this study presents the results of digital filter bandwidth selection of the ECG signal analysis. The digital filter combination in this study includes one notch-filter, two low-pass filters and one high-pass filter. The results show that the series connection of the 55-65Hz notch filter, 100Hz low-pass filter, 40Hz low-pass filter and 10Hz high-pass filter presents a discernible ECG signal.

This paper presents an area efficient analog-front-end (AFE) amplifier for electroencephalography (EEG) signals. The primary circuit system is divided into an AC-coupled IA and a programmable gain amplifier (PGA). In the proposed AFE, DC servo loop (DSL) is utilized to suppress the offset voltage and MOS capacitors are utilized to reduce the large chip area occupation in order to operate at the low frequency of 1kHz. The proposed area efficient AFE circuitry is implemented in TSMC 90nm CMOS process with 1V supply voltage. The chip area occupies 0.122 〖mm〗^2.

The recognition of serial code on a container is a key to manage the containers. Traditional approach to identifying the code manually is tedious and error prone. To tackle the problem, scholars have proposed various approaches based on traditional text detection and text recognition methods from the images to automatically recognize the code in the past. However, the approaches usually suffered from low recognition accuracy due to complex environments or inexpensive photographic equipment. In recent years, deep learning techniques have proved effective in real-time object detection and recognition. In this study, we propose to use deep learning coupled with image processing techniques to tackle the problem. Experimental results indicate the proposed approach is feasible and promising.

In the medical field, acquiring a sufficient number of medical samples can be challenging, and the collected datasets may be imbalanced and small. To address these issues, we propose a weighted SMOTE algorithm that targets imbalanced datasets. This technique has been applied to a dataset of breath biomarkers of liver disease as a feature set and a supervised learning model. Our results show that the proposed method significantly improves the prediction probability and classification performance of the chosen model in both the original imbalanced dataset and the balanced dataset. This study demonstrates the potential of the proposed approach to enhance machine learning performance while dealing with small and imbalanced datasets in medical applications.

Based on two methods of multi-view geometrical map modeling and multi-view loop closure optimization in conventional multi-view RGBD visual simultaneous localization and mapping (SLAM) and navigation, this paper proposes two specific methods, forward-backward-view depth-visual point cloud and fusion of upward-view visual odometry and wheel odometry, to simultaneously improve the accuracy, reliability, and immediateness of depth estimation, localization and mapping, scale estimation, and loop closure. Experimental results show, in terms of depth estimation, localization and mapping, scale estimation, and loop closure, the proposed forward-backward-view depth-visual SLAM and navigation based on upward-view visual odometry and wheel odometry has better estimation accuracy, noise immunity, and convergence speed than conventional multi-view RGBD visual SLAM and navigation, especially under the conditions of sharp rotation, surrounding moving objects, and white/textureless background environment.

The study uses machine learning to determine the accuracy of different algorithms for predicting MetS. Early prediction of MetS through routine health screening is an important goal for preventive medicine. The subjects in this study were employees aged 20 to 65 years old who signed a consent form for in-hospital health screening data were collected from 2015 to 2020. Total of 3538 individuals who underwent health screening were included in this study. Baseline characteristics of the 315 participants diagnosed with metabolic syndrome and 3222 participants without metabolic syndrome. This is a machine learning approach to identify metabolic syndrome using biochemical values and lifestyle habits. We found that the F1 score for detection of metabolic syndrome by neural network is the most accurate and higher compared to other models. Our study also identified some new indicators, including heart rhythm, RBC, WBC, Hb, and uric acid, as candidates for better detection of metabolic syndrome.

In this paper, a polar-coordinate-based PointNet++ is proposed to classify objects such as roads, cars, motorcycles, and pedestrians using point clouds of outdoor scenes captured by LiDAR. Since the outdoor point clouds obtained by LiDAR are sparse in the distance, if the point clouds are partitioned by a grid of Cartesian coordinate, the number of points in each partition varies greatly, which causing the bias of the learning network. With SemanticKITTI data set, the experimental results of polar-coordinate-based PointNet++ show that the proposed method can improve the overall accuracy and processing speed. The results show that Mean Intersection over Union (mIoU) is 52% and overall accuracy (oAcc) is 78%, processing speed per frame improve from 300ms to 30ms.

This paper uses an underwater unmanned vehicle to detect the distance and swimming direction of tilapia. As the underwater vehicle is equipped with a single camera system that lacks depth information, the YOLO3 architecture of deep learning is used to determine the relative distance of fish and further analyze the swimming direction of the fish group, which is of great help in analyzing fish group.

This study aims to generate Kinmen-style text using the GPT-2 model with transfer learning. Kinmen-related literature is mainly preserved in paper format, and thus, this research collected texts from six works and documents in various fields using web crawling, OCR, and text-to-speech conversion technologies. After data cleaning, the collected texts were segmented and fed into the 12-layer Transformer-Decoder GPT-2 model for training. Three types of experiments, namely Train From Scratch, Pretrained Model, and Transfer Learning, were conducted to evaluate the effectiveness of the proposed method. Results showed that the output text in Train From Scratch contained errors in usage and phrasing, while Transfer Learning corrected the errors and preserved the unique Kinmen-style language usage.

In modern society, the interaction between human and machine has become more and more important. Compare with the traditional interface such as buttons and joysticks or the touch screen, which requires familiarity with the interface of input devices, operation with the gesture is more intuitive and attractive. With the flourish of computer vision and artificial intelligence, deep learning can share the computations and the parameters by sharing the convolution layers, in order to eliminate the multistage system. Once the difference tasks share the same feature, the network can be merged into a multi-task deep neural network. For hand gesture recognition task, hand detection and gesture classification can share the same image features. In this work, we focus on hand detection and segmentation task, enforce the data on extreme situations scene, and use iterative training to further reduce the errors. Finally, we achieve a robust hand segmentation model with F1 score over than 90%.

Human activity recognition (HAR) using the deep learning method has caught the attention of researchers thanks to its automatic feature extraction and accurate prediction capabilities. However, for applications based on a wearable sensor, such as an inertial measurement unit (IMU), the process of collecting and hand-labeling large amounts of data is complicated and labor-intensive, meaning that there is a limited amount of data available for model training. Therefore, there is a need to propose and develop data augmentation approaches to generate high quality data for the growth of HAR research. We propose a head-mounted virtual gyroscope signal generator to alleviate the problems caused by the lack of data in head movement-related applications. Unlike previous work, our system only generates head-motion related gyroscope data, minimizing system complexity. We trained a deep-learning model in a head motion-based application with different generated sensor data ratios, and show the viability of our proposed data generation method.

As operators increasingly adopt environmental, social, and governance (ESG) policies, the development of green radio access networks (RAN) is becoming a future trend. Local microgrids are often situated on-site and can provide direct power supply, coupled with the use of power-efficient equipment and lighter on-site units, making it more feasible to develop renewable-powered RANs. In this paper, we have summarized the net-zero strategies of operators and proposed a conceptual framework that combines RAN and microgrid by incorporating the RAN intelligent controller (RIC) and energy management system (EMS). We have also elaborated on the benefits of the proposed framework and identified future works of study.

This paper presents a 57.6 V high voltage bidirectional current sensor for the autonomous underwater vehicle (AUV). 4 source followers are used to trace the external high voltage and to overcome the difficulty of processing the high voltage of the 57.6 V battery module. Moreover, the dual sensing loops are employed to detect the bidirectional current, respectively. The proposed design is implemented with a 0.18 µm HV process. The core area is 557 × 489 µm2. The detection range is from -3 A to 3 A. The maximum detection error of the linearity is simulated to be 0.89%.

This study tried to solve the high repetitiveness of routines in fitness game, and improve player engagement. Our research proposes an AI which is able to be empowered through data collected by motion-sensing device during player's fitness activities, and makes fitness game more diversifying and challenging. Also, we use GPT-3 to generate suitable opponent based on the player's needs, providing fitness activity's pre-assessment. The results show a higher level of player participation in fitness game.

With the rapid progress of six-generation (6G) wireless communication technologies, Low Earth Orbit (LEO) satellite global network technology played a very important role and promoted millimeter-wave (mm-wave) systems gained demand. However, security always is a very important issue in the design and use of wireless networks. In wireless security, physical layer security (PLS) addresses security concerns by taking advantage of the fundamental ability of the physics of radio propagation to provide certain types of security through the exploitation of the physical properties of the wireless channel. Despite this, PLS is vulnerable to cybersecurity attacks, the Sybil attack has a significant threat to millimeter-wave base LEO satellite communications. The attacker targets physical layer communication to create multiple identities and gain unauthorized access to a system. To address this, we propose a method based on an artificial intelligence technique, machine learning algorithms, in cybersecurity. This approach handles significant security dataset training, using different techniques such as decision trees and Recurrent Neural networks which allow feature selection and Sybil attack classification. We also refer to guidelines and specifications of orbit satellite standards in ISO 16164-2015 and ISO 27852-16 to validate the satellite transmitter. Simulation results arise an improvement in security and the ability to detect and prevent Sybil attacks. This approach makes a significant contribution to the field of satellite cybersecurity by providing a standard-compliant solution.

The Open Radio Access Network (O-RAN) uses a virtualization architecture to disaggregate components and connect them through open interfaces, and also uses software defined networking (SDN) technology for decoupling. The O-RAN architecture is an innovative telecommunication architecture, which also brings some potential problems for security, including but not limited to the followings: new architecture (disaggregation, Real Time RAN intelligent controllers (Near-RT RIC), virtualization), the open interfaces (O1, A1, and E2, etc.), artificial intelligent and machine learning (AIML) workflow, and standardization issues. Facing the new structure, in addition to the original risks, new risks also need to be considered. "Secure in design" is imperative to mitigate the information security risks. This study will look at the risk assessment of O-RAN from the perspective of testing and do security analysis.

An Intrusion Detection System (IDS) is used in the field of cyber security to prevent and mitigate threats during transmitting data between the devices and server. An IDS using Machine Learning (ML) can be applied to identify and classify security threats. Various ML techniques have been applied to improve the model performance of IDS. In addition, the quality of the dataset is also an important determinant that can greatly improve the detection accuracy. The authors propose an IDS framework based on SVM ensemble with feature selection and balancing dataset. The experiment results show that the proposed model can achieve a high detection performance.

This paper presents a framework for the robustness evaluation of AI-based malware detectors. The framework is based on the idea that a robust detector should be able to detect malware from multiple, different sources and should be able to adapt to new threats. The framework is composed of three elements: a dataset, a set of evaluation metrics, and attack experimental steps. The dataset consists of benign and malicious software from multiple distinct sources. Researchers can upload their own dataset, feature extraction scrips, or trained model in the framework. The evaluation metrics measure the accuracy, F-1, false positive, and false negative rate of the detectors. The evaluation framework is utilized to assess the robustness of detectors using the dataset and evaluation measures. The evaluation results are then used to confirm effectiveness and efficiency. The proposed framework aids in resolving the issue that researchers don't need to possess an entire experimental setting, especially considering that some research sources are difficult to collect and integrate.

Since 2020, the number of cyber-attacks has doubled, seriously affecting the quality of private internet use and causing significant economic losses. Threat intelligence analysts use threat reports to analyze malicious actors and organizations to prevent cyber-attacks. This study proposes a method that leverages the diamond model of intrusion analysis and deep learning models to provide a comprehensive view of malware attacks. We examine the comparative efficacy of Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transfer (GPT) in threat intelligence reporting. The results show that GPT achieved an accuracy rate of 61%, a 17% improvement over BERT. It can effectively assist malware analysts in their decision-making process.

Low Earth orbit (LEO) satellites play an important role in modern communication systems. Recently, Khan et al. designed a robust authentication and key agreement scheme (RAKS) for LEO satellite communication. They also performed security analysis of RAKS to verify the declared safety. Upon careful examination, however, we discovered several security flaws in RAKS, including insufficient information for user identification, vulnerability to elaborate DoS attacks, and inaccurate estimation of computational costs. In this paper, we highlight these flaws and suggest potential modifications to improve the scheme.

This paper proposes a novel Chaos-based Advanced Encryption Standard technology for both image and data protection, it can be applied to the micro control unit for the industrial internet of things(IOT). There are two main characteristics in the proposed Chaos-based Advanced Encryption Standard (Chaos-AES): (1)Logistic map plays an important role in the generation of private key for data as well as image encryption, which provides rapid as well as light-computational key series. (2)The encryption round is flexible, which can be adjusted according to the level of protection to the ciphered data. Experimental results demonstrate that the proposed Chaos- AES structure is effective as well as feasible to the data protection during transmission

Recently, information security has become an important issue. People receive more and more information and often download lots of data and files through the Internet, but there is no way to verify the correctness of the data source. In this paper, we introduce a file verification mechanism based on Verkle trees without the intervention of third-party institutions. Furthermore, it will restore the tampered files through the blockchain network and Inter-Planetary File System (IPFS). In addition to designing the file verification mechanism, we will discuss the overall efficiency and compare the differences between the Verkle tree and Merkle tree architectures.

This paper proposes a design methodology and its hardware design of fast 3D-sorter engine in successive cancellation list (SCL) for 5G applications. Our hardware only requires fewer comparators as compared to the popular sorters, such as the odd-even sorter and Bitonic sorter. Meanwhile, the fast 3D-sorter engine has less computing time, owning the fast-computing feature for 5G demands. Our proposed work only occupies 7243.3 um^2 after synthesized with TSMC 40-nm CMOS technology. In comparison, our circuit outperforms the odd-even sorter and Bitonic sorter by 2.84 and 1.64 times, respectively, in terms of operating speed.

This paper describes a filter circuit based on one second-generation current conveyor (CCII), one differential voltage current conveyor (DVCC), three resistors, and two grounded capacitors. Outputs of three voltage-mode filter responses, which are highpass, bandpass, and lowpass are available in the proposed circuit. The proposed circuit has the feature of high input impedance. It employs only grounded capacitors and is ideal for integration. Simulation results are carried out to confirm the theoretical analysis.

Gate camouflaging is a useful method to protect integrated circuits from reverse engineering attacks. However, the design flow for the adoption of gate camouflaging has not been well studied. In this paper, we use camouflaged cells as spare cells. Then, a metal-only ECO (engineering change order) algorithm is proposed to maximize the hardware security by replacing normal cells with camouflaged cells under the same clock period constraint. Experiment results consistently show that the proposed ECO approach can effectively improve the hardware security without violating any timing constraint.

Computing-in-memories (CIM) is recognized as a promising solution to overcome the Von Neumann bottleneck. To increase the fault-tolerance, redundant memories can be added into the CIM. However, the integration of redundant memories with the MAC operation has not been well studied. In this paper, we propose a fault-tolerant near-memory MAC design by using redundant memories. To ensure a correct near-memory MAC operation, the outputs of redundant memories are also considered as the inputs of adder tree. Then, we use a partial product matrix to implement the corresponding adder tree. Experiment results show that the overhead of the fault-tolerance approach is approximately linear to the overhead of spare rows for the memory.

Due to the development of modern artificial intelligence, CNN has been widely used in many application domains. Because the architecture of CNNs has become more and more complicated in recent years, the amount of computation time increases dramatically. In order to meet the need of real-time computing, some research use the Winograd algorithm to reduce the number of multiplications. However, Winograd algorithm does not get computation benefit for all kernel size and stride. In this paper, we propose an integrated PE (processing element) architecture that can support both Winograd and traditional convolution computations. Winograd architecture is used for the most common 3x3 kernel with stride-1 sliding window calculation, and convolution operation is performed on the systolic array architecture for the other kernel size or non-stride-1 sliding window. Based on the proposed integrated Winograd and convolution PE architecture, we can reduce quite amount of the CNN computation cycles.

This paper proposes a framework for simulating RRAM-based CIM. The framework is designed to explore changes in accuracy under the constraints of ADC precision, resistance on/off ratio, and activated WL number. We also propose an accuracy-aware quantization methodology to reduce the number of quantization bits required. Experiment results show that the quantized IFMs only reduce the accuracy by 0.09%, while the accuracy of quantized weights dropped by 10.55%. In the future, we will add R on/off ratio and current error to further study the relationship between number of activated WL and CIM accuracy.

It is recognized that there is a positive correlation between the disease of obstructive sleep apnea (OSA) and the abnormalities of craniofacial morphology. Thus, an analysis to craniofacial morphology can help doctors to quickly judge the possibility of OSA. To reduce the overheads caused by manual process, an automatic facial analysis system is needed. However, previous systems are designed for adults. Moreover, previous systems do not deal with the calculation of length. Different from previous works, in this paper, we propose an automatic facial analysis system for the detection of pediatric OSA. Based on Hough transform, our system can automatically perform length calculation. Compared with the manual process, experiments on real cases show that the error rate of our system is acceptable.

In this paper, the operation of coordinate rotation digital computer (CORDIC) based circular rotation made (CRM) is studied to enhance the iterative microrotation speed. Moreover, the angle recoding (AR) method of CORDIC CRM implementation still involves more iteration loops. The proposed method in this paper can improve by about 12% iterations reduction to AR method. Besides, the extended elementary angles set recoding method using 2-dimension (2D) angles set searching still possesses more computational complexity to find the closest elementary angle. Based on the 2D angle sets, we propose the successive cancellation of the coarse and fine angles rotation, which uses 1-dimension (1D) angles set searching to find the closest elementary angle. Therefore, the proposed method provides lower computational complexity than 2D searching method. Finally, simulation results show that the proposed method involves the low iterations being close to the 2D searching optimum method. Furthermore, to simulate the rotated accuracy of CORDIC CRM, the proposed method provides more precision than the conventional AR method.

This paper develops a writing and walking support system using haptic and wearable devices. In the system, a user can write characters in calligraphy by using a haptic device. After the user finishes writing characters, he/she can choose to exhibit his/her work in a virtual 3D exhibition room or not. In the exhibition room, his/her avatar in the virtual space is displayed, and the avatar is moved with a foot sensor attached to the user's waist and feet. Multiple avatars can communicate with each other in the room.

To achieve early detection and prevention of dementia, this paper enhances an object identification game using the olfactory and haptic senses which can be enjoyed by elderly people. By analyzing data such as the correct answer rate and response time of each correct answer obtained in the game, we can expect to achieve early detection and prevention of dementia. We will prototype the game system and investigate its effects through experimentation.

This paper evaluates the QoE of video and audio transmission over a full-duplex wireless LAN with interference traffic through a computer simulation and a subjective experiment. We employ a simulation environment with a pair of audiovisual transmission and reception terminals and a pair of interference traffic transmission and reception ones. We investigate the effect of the transmission rate of interference traffic and communication distance in a wireless channel on the output quality of the video and audio stream at the reception terminal. We perform a subjective experiment with the output timing of video and audio obtained by the simulation.

In this paper, two users operation of a remotely-controlled robot having a force sensor by using haptic interface devices is compared with one user operation for the cooperative work of carrying an object under network delays through experiment. The objective is to find strategies that apply smaller force to the carried object. As a result of our experiment, we demonstrate that under large network delays, the two users operation outperforms the one user operation. We also show that when two users have different network delays, the user with the smaller delay can help the other user.

Industry 4.0 is expanding in manufacturing. With digital technology, manufacturing lines are expected to become more efficient. This research aims to collect the activity status of Automatic Guided Vehicles (AGVs) on the production line and compare it with simulation results by FlexSim. We analyzed the activity status of the AGVs by attaching beacons to the AGV and the parts carrier vehicles AGV carry and collecting their positions of them. As a result, we confirmed that the activity status is similar to the simulation results and demonstrated that our proposal benefits AGVs tracking.

To improve productivity in small and medium-sized manufacturing businesses, it is necessary to use digital systems. Therefore, we designed a sensor device that requires no additional work to install and built a ball screw torque monitoring system that can be applied to old-type ball screws. We proposed an algorithm that can divide collected data into cycle waveforms. This algorithm allows analyzing the difference among cycles. The experiment showed that the proposed system could detect the torque difference between bearing with grease and without grease.

Central management has used and maintained many portable medical devices(MDs). However, the central management lends medical staff the MDs only with managing by a ledger; therefore, missing or subletting may be caused. In this paper, we develop and propose a device that can measure MDs' location and power consumption and a visualization system. We implemented 85 developed devices in 4 hospitals and confirmed that the system adequately managed MDs.

This paper presents a time series analysis of queueing systems by using machine learning. Recently, the Digital Transformation (DX) market has been expanding rapidly. Data processing systems that analyze time series events are important to realize DX. Transient performance evaluation of queueing systems is useful for designing data processing systems. Theoretical analysis is a method to evaluate transient performance. However, it is difficult to obtain a solution due to the difficulty of identifying the distribution, arrival rate, and processing rate, as well as the complexity of the theory. In this paper, we propose the time series analysis method by machine learning to solve the problems.

Traditional intrusion detection for home security mostly adopts door sensing, infrared sensing, or image recognition of cameras. In this study, a radio frequency (RF) sensing-based intrusion detection method based on a Wi-Fi Mesh network for home security is proposed. This method combines a Wi-Fi Mesh network and RF sensing technology to achieve home intrusion detection without camera image recognition, infrared sensor, or other equipment. In the experiments, the results show that this method and system accurately detect intrusion behavior with high sensitivity and accuracy. Overall, this method offers a new approach to home security that is cost-effective and reliable.

For improving security, this study proposes a mobile data usage prediction method based on homomorphic encryption (HE)-based similarity calculation without plaintext storage. The cosine similarity, angular similarity, Tanimoto similarity, and Euclidean similarity based on HE according to ElGamal cryptography are proposed and combined in a machine learning method to consider the mobile data usage during the last periods for predicting the mobile data usage during the next period.

With the development of intelligent transportation systems, license plate recognition is now directly performed by the system. However, the system is vulnerable to environmental factors that affect recognition, such as darkness and overexposure. To address the issues of blurry, overexposed, and underexposed license plate images, this study proposes a two-stage approach: the first stage enhances the license plate images through image processing, and the second stage uses a deep learning model to restore the images. By employing this two-stage approach, we obtain license plate images that are recognizable.

The flowering season is an essential part of tourism in Taiwan, but the blooming conditions may vary each year due to weather factors such as rainfall and temperature. Numerous AI models are available in the market for predicting flowering. However, these models only apply to specific regions due to climate variations. Furthermore, Taiwan's climate is known to be changeable. By utilizing Taiwan's climate and flowering date as training parameters, the accuracy of the trained model can be significantly higher than other models that did not use Taiwan's climate information as training data. This paper builds an app called "Booming Blooming" that contains a flower prediction model and real-time information on weather to collect data. Booming Blooming used the weather data from Taiwan's Central Weather Bureau to predict the best flower viewing time and display the forecast information. Most importantly, it can be used as a reference for users to go out to enjoy flowers. In addition to the embedded Google map to help users find nearby stores, traffic conditions, and people around flower viewing locations, Booming Blooming also provides a flower-sharing platform to update the latest information. The proposed flower blooming prediction model and APP can help Taiwanese enjoy flower-viewing activities.

We proposed a framework for defect detection in production lines that leverages deep learning models, out-of-distribution (OOD) detection, and continual learning to address the challenges of unknown defects and catastrophic forgetting. The proposed method divides classifier training into chronicle tasks, each introducing new defect classes and leveraging OOD detection to classify unknown defects. We trained and tested the framework on a highly unbalanced product defect dataset and demonstrated that it outperformed existing approaches, improving the average f-score by 10%. Our method also improved the performance of the PODNet and DER models, but not the WA model due to its poor performance on our dataset. These results suggest that the proposed method has the potential to improve defect detection in production lines, especially for small-quantity-wide-variety production scenarios.

Mura is usually caused by inhomogeneity and material defects in the manufacturing process. According to the JND value, it can be divided into light Mura and serious Mura. In order to optimize the repair process, the factory hopes to distinguish between light Mura and serious Mura before sending them to the repair site. However, the traditional AI model only distinguishes between normal and Mura and is ineffective in distinguishing between light Mura and serious Mura. To address this issue, we propose a Mura Detection System using an image inpainting model with self-supervised technique and an attention module to distinguish light Mura and serious Mura. The experiment results show that the proposed method's Area Under Curve (AUC) can reach 0.85.

Behavioral analysis of the elderly is very important for improving the quality of elderly healthcare services. This paper proposes a Fine-Regularity Feature Extraction Mechanism (FRFEM) based on unsupervised learning algorithm, which can extract fine features including occurrence, duration and frequency regularities. The proposed FRFEM can be applied to detect abnormal behaviors and assess the health status of the elderly. Simulation study exhibits the effectiveness of the proposed FRFEM on feature extraction.

With the growing application of Bluetooth Low Energy (BLE), the sensors embedded with BLE has been applied in elderly home environment to monitor the living states of the elderly. In order to transmit the emergency of the elderly to the gateway, this paper aims to construct a disjoint paths for the emergent sensors deployed in the same room. The proposed mechanism can avoid collision and prevent the emergency transmission from failures. The simulation study exhibits the effectiveness of the proposed mechanism on collision avoidance.

The authors explore the conditions of parameters in a Continuously Decimalized Maximum length (CDM) sequence to realize its uniform distribution bit patterns and evaluate the special characteristics of CDM-sequences with the parameters.

The authors focus on a kind of modified M-sequence to accelerate the development of stream ciphers. This paper proposes a dynamic mapping scheme and considers the correlation property to evaluate a periodic property of the sequence. As a result, it is found that there is a symmetricity and the authors show its proof theoretically for further improvement.

This paper formulates the Smallest Botnet Firewall Building Problem and proposes a heuristic solution based on the Girvan-Newman algorithm. The problem is to build the smallest WH botnet as a firewall to prevent the further spread of a malicious botnet. A brute-force solution available to this problem has been proposed, but it requires an enormous amount of computation time as the size of the network increases. The proposed heuristic solution uses the Girvan-Newman algorithm to find communities covering a specified number or more nodes, then computes the minimum cut between each community and the malicious botnet, and finally obtains the number and placement of needed white-hat bots from the connection of the cuts. The proposed solution yields similar results to the brute-force solution, while significantly reducing the computation time.

Unmanned aerial vehicle (UAV) has received a lot of attention in recent years due to its potential applications in military, surveillance, package delivery, agriculture, and disaster management. Typically, these applications require multiple drones to build a network, transmit and receive data from IoT devices, and offload the task to a cloud server. However, the Internet of Drones (IoD) is vulnerable to a variety of attacks associated with IoT ecosystems, such as eavesdropping and identity spoofing. In this paper, we propose a secure IoD system in which the data transmission between UAVs and IoT devices is protected by physical-layer secrecy and the identity authentication is strengthened by blockchain. To the best of our knowledge, this work is the first to explore the joint design of physical-layer security and blockchain for protecting IoD systems. We demonstrate the feasibility of our designed IoD framework using experiment results from our testbed implementation.

The author has been developing a smart gate for the main entrance of the Library of Konan University. Two sensors are used. One is the FeliCa card reader that can read the information of the ID card issued by the university, and the other is a 2D LiDAR to verify the passage at the end of the gate, that are cooperatively used with the support of a display for the users. The design is intended to adhere the universal design. The system is of very low cost and has various extension capabilities because the author and the university can share any information of the system.

This paper presents a lightweight, low-cost, and low-network-consumption edge-cloud architecture that is suitable for small- and medium-scale fish farmers engaged in precision aquaculture.

The proliferation of Internet of Things (IoT) devices has led to an exponential growth in data generated at the edge of the network. Edge computing, a distributed computing paradigm that enables computation and data storage at the network edge, has emerged as a promising solution for managing this data deluge. With the integration of Artificial Intelligence (AI) technologies, edge computing can provide real-time insights and decision-making capabilities, making it a powerful tool for a variety of IoT applications which poses both opportunities and challenges

A comprehensive visual traits-based recommendation system is designed for proactive retailing in a physical store environment. The proposed system utilizes computer vision algorithms to analyze various visual traits of customers, including facial features, clothing and accessories, to provide targeted product recommendations. The system does not require customers to provide any personal information, making it a less intrusive and more hassle-free approach compared to the conventional membership approach. By providing tailored product suggestions, the system offers retailers the opportunity to enhance their customers' shopping experience and increase sales. This paper provides a detailed description of the system's design and setup

This paper provides an overview of the Internet of Things (IoT) components in the layered architecture. Then we give an overview possible of tests that are applicable in the layers and possible challenges of in conducting tests in the layered architecture. We provide explanation on the integration and interoperability tests with a simple evaluation on how the existing architecture works. We present the need for better tests and an improvement using the combination of path-based and business process components presented in the paper.

This paper presents a comparison of the performance and efficiency of two single-board computers, VisionFive V1 Board and Raspberry Pi 4B model, in the context of Industry 4.0 and IoT applications. Our study focuses on evaluating the compression speed, compute speed, power consumption and performance efficiency of both devices. Our results show that the Raspberry Pi 4B model outperforms the VisionFive V1 Board in all performance metrics. Nevertheless, we note that the VisionFive V1 Board development community is still in a relatively infancy stage compared to the well-developed Raspberry Pi. Our findings contribute to the ongoing discussion on selecting appropriate single-board computers for Industry 4.0 and IoT applications and provide valuable insights for engineers and researchers working in this field.

Google Meet (formerly known as Hangouts Meet) is a video-communication service developed by Google. It is one of two apps that constitute the replacement for Google Hangouts, the other being Google Chat.It replaced the consumer-facing Google Duo in late 2022, with the Duo mobile app being renamed Meet and the original Meet app set to be phased out.

In the early months of the COVID-19 pandemic, Google announced Meet was to be made available to all users, not just Google Workspace users, in which it previously was. The use of Meet grew by a factor of 30 between January and April 2020, with 100 million users a day accessing Meet, compared to 200 million daily users for Zoom as of the last week of April 2020.

Features of Google Meet include:

Two-way and multi-way audio and video calls with a resolution up to 720p An accompanying chat Call encryption between all users Noise-canceling audio filter (depending on the license only) Low-light mode for video Shared whiteboard, reactions, polls, voting, Q&A (depending on the license) Team activities - shared YouTube and Spotify watching, playing mini games like UNO! Mobile, Kahoot! and Heads Up! Google document sharing (documents, spreadsheets, presentations) Integration with the Google ecosystem, e.g. live streaming to YouTube, Google Calendar and Google Contacts for one-click meeting calls Screen-sharing, browser tab sharing Ability to join meetings through a web browser or through Android or iOS apps "Knock Knock" shows a live preview of the caller before the recipient picks up, which Google says is to "make calls feel more like an invitation rather than an interruption". Ability to call into meetings using a dial-in number (US numbers always; international numbers when included in Workspace license) Hosts being able to deny a user's entry to, remove a user from, and control microphone and video access in a call. Video filters, effects, backgrounds and augmented reality masks. Ability to join meetings through a web browser or through Android or iOS apps

Food safety issue always existed in Taiwan, from evil oil event in 2014 to the re-sale of expired product by changing its expiration date in 2017, and finally to the poisonous tea leaf even, they not only affected seriously the food safety of consumers, but also wounded seriously the reputation of food safety of Taiwan, indirectly, it also affected the farmer's income. Therefore, in order to solve such problem, the government has launched food safety badge and traceability for agricultural product to facilitate the consumers and related suppliers to search the traceability system, in this way, consumer's suspicion on food safety can be eliminated, however, it was still insufficient. In this study, the focus was put on the introduction of block chain into traceability certification, and through this, consumer's acceptance on smart agriculture was further investigated. Therefore, the main focus of this study was to investigate the influence of introduction of block chain into traceability certification to see its influence on consumer's acceptance on smart agriculture; in other words, block chain was added into traceability certification to study the influence on consumer's acceptance of smart agriculture. Moreover, SPSS and Smart PLS 3.0 statistical software were used as main tools for data analysis, finally, data was analyzed and theoretical model was verified. From the empirical result, it can be seen that traceability certification will affect positively consumer's acceptance on smart agriculture, and traceability certification will affect positively the block chain, and block chain will affect positively consumer's acceptance on smart agriculture, furthermore, through regression analysis, it was found that block chain had mediating effect, meanwhile, traceability certification, through block chain, showed stronger influence on consumer's acceptance on smart agriculture than the direct influence of traceability certification on consumer's acceptance on smart agriculture, hopefully, the researchers expected to use the research result as reference and practice meaning for future development of smart agriculture.

With the rapid development of internet technology, it has become increasingly straightforward for people to seek and share health information in online health communities (OHCs). However, the current OHCs are mainly centralized systems fraught with data breaches, identity theft, and censorship problem. Some researchers proposed that Web3 technology can provide potential solutions to current problems. In this research, we identify three core components that should empathize in developing OHCs: trust, privacy, and information quality. Accordingly, we surveyed and analyzed Web3 literature and generated the spectrum of Web3 research. Then we applied text analysis to 6 articles we had collected to conduct a co-occurrence network and generate minimum spanning trees. We then discussed the important factors in the technical, application, and service aspects for applying Web3 technology in future OHCs development.

In sports competitions, the media plays a crucial role. Without media coverage, even the most exciting games can only be passed down through word of mouth and not easily known to people. From the earliest news newspapers and radio broadcasts to television broadcasts, the constant innovation of media represents that the audience's demand is always increasing. Nowadays, in sports event broadcasting, in addition to traditional broadcasting methods, various media applications such as sports science analysis, laser projection, AI communication, virtual characters, etc., have gradually been added. Under the trend of the current multi-dimensional universe, content is king, and whoever owns this content controls the audience's attention. In terms of sports event broadcasting, the uniqueness, interactivity, professionalism, and entertainment of the content will be the parts that attract the audience. They not only heat up the atmosphere of the game but also satisfy the viewing experience of sports fans. Therefore, this study proposes a mixed reality (MR) system that integrates sports event broadcasting and the competition venue, hoping to allow viewers to immerse themselves more in the game while watching it.

This paper discusses the high-performance exigencies using field oriented control (FOC) resolver. To overcome permanent magnet synchronous motor (PMSM) under dynamic speed and load disturbance, the drive control features space vector pulse width modulation (SVPWM) is proposed. An adaptive neuro-fuzzy inference system (ANFIS) controller is used in the system to enhance speed performance. A novel efficient PMSM control is verified with digital implementation.

In recent years, the world has faced various challenges due to the impact of COVID-19, and the electronic and electrical industry has received particular attention, developing many products to cope with the pandemic. People have been trying to find ways to have least impact from virus. This study uses a non-contact system to allow users to complete operations without physical contact while also providing health management, intelligence, and versatility. Using the embedded system Raspberry Pi as the core of this project, and extending its functions to include speech detection, AI facial recognition, and gesture recognition. The concept is to introduce an AI system into the Internet of Things, which enables digital devices to be closely connected and have functions such as remote control and integration with other devices. Finally, using a finite state machine to link the entire system together and using a human-machine interface to interact with users through an LCD screen and speaker to display information. It also has a secure connection between two different networks and can respond to user needs by utilizing data from both on-site and cloud sources.

Semi-autonomous vehicles require human drivers to regain manual control when the automated system fails to perform the driving task. The tactile modality has shown the potential in improving the effectiveness of takeover requests (TORs), yet the potential of pattern-based alerts in TORs has been largely unexplored. The present study compares the impact of static and dynamic patterns in directional tactile TORs on takeover performance. Preliminary results show that static TORs create shorter reaction times, lower crash rates, and better aid driver experience.

In an aging society, the rate of dementia would be dramatically increased. It is urgent to provide intelligent facilitative devices and appropriate medical care for those who with dementia symptoms, especially for the elders. Accordingly, the primary research target of this study is to automatically detect and monitor behavior of the elderly suffered from dementia. Although previous research has been conducted to identify various human movement patterns, the experimental data were all received from cameras. Furthermore, these related applications are not specifically designed to identify the behavior of the elderly with dementia. In order to improve previous research methods, our study adopted wearable multi-axis sensors to collect corresponding motion data with protected privacy. In addition to collecting raw data containing time series signals, frequency and spatial domain features were applied to detect human behaviors, and Random Forest algorithm was applied to construct a motion prediction model. The classification system for real-time motion recognition contained two main actions of users including holding/putting hand gestures and walking model. Finally, an additional camera on the chest was designed to be triggered and to take a shot according to previous settings. It could enhance practical applicability of the proposed intelligent system to assist the elderly with dementia to find her/his lost objects. In order to extend the subsequent applications, this study developed an APP to record individual's behavior and to integrate with a chat robot to answer the location and time of the lost objects, or combine it with a medical advice of physical rehabilitation to record all the activities during training processes, which could be used as a reference basis for patient's returning to hospital visits> It provides medical staffs with appropriate recommendations for supporting precise medical care and treatments.

Digitalizing documents has become a crucial aspect of modern information management. However, various factors, such as manual handwriting, stains, and degradation, can significantly impact the quality of digitalized documents. This paper proposes a autoencoder-based network cleaning technology to completely erase manual writing, stains, colors, and background noise from paper documents to enhance their digitalized quality. Our proposed approach utilizes machine learning algorithms to identify and isolate unwanted elements from the document and then replace them with clean and transparent backgrounds. The experimental results demonstrate that our proposed approach significantly improves the digitalized quality of paper documents, making them more suitable for long-term storage and retrieval. Our proposed technology has the potential to revolutionize the field of paper document digitalization and improve the efficiency of information management.

Regarding existing technology, people with disabilities tend to be monitored using video devices for occurrences of falls or dangerous situations. In addition, health monitoring devices such as blood pressure monitors and blood glucose meters can be set to enable automatic data transmission to medical staff and family members. Many older adults hesitate to allow monitoring equipment in areas such as bedrooms, bathrooms, and toilets, where they tend to fall. Moreover, requiring older adults to wear devices involves potential problems such as forgetting or unwillingness to wear them. This study involves to develop and manufacture a care robot featuring an artificial intelligence behavior model that evaluates emotional feedback. Within this paper we will complete the construction of the robot, which involves the use of a multi-sensor data fusion technique for achieving obstacle avoidance and allowing the robot to operate autonomously in dynamic environments and daily life scenarios. Moreover, the robot is equipped with a vision and voice recognition module to collect information on care receivers' expressions, behaviors, and voices, thereby establishing a data storage and management platform.

A real-time Internet of Things (IoT) status update system is considered, where the IoT device needs to send not only the interested sensing information (payload) but also the controlling information (extra) to the destination for improving the quality of service. To simultaneously transmit these two types of information, a promising coding scheme, i.e., free-ride code, is utilized, in which the 5G low-density parity-check coded sensing sequence is superposed on randomly coded controlling sequence and is delivered to the destination without consuming additional bandwidth and power resources. The average age of information (AAoI) is evaluated for both two types of information. Numerical simulations indicate that the AAoI of extra can be reduced by free-ride codes without impairing the one of payload.

Water leakage is a significant cause of drinking water loss, and the detection of leakage in public toilets is often challenging due to the many toilets in these environments and the difficulty in detecting small pipeline or equipment leaks without obvious sound or vibration. This paper presents a deep learning-based environmental sound recognition system for detecting water leakage in public toilets. The system notifies users of the location of current leaks to arrange timely repairs. Experimental results show the feasibility and practicality of the implemented system, achieving accuracy above 95% at night. This research contributes to the timely detection and repair of water leakage in public toilets, improving water conservation.

Driven by China's demand of independent high-end software, IT (information technology) innovation gradually comes into people's lives. In this area, operating system as an important part, its research is bound up with the information security of key sectors in China. Through investigation, upper-layer applications with a Linux basement arise people's concerns on security. To address this problem, we propose the framework of this software, and introduce its implementation in detail in this paper. First of all, we investigate the needs of users to determine the overall framework of the software and functions' design. Secondly, we develop the back-end modules to implement preset logical functions, and design the front-end interface to comprehend users' demands. Finally, we perform the functions of the software on Ubuntu 20.04 to verify the stability and feasibility.

In this paper, we propose a monocular depth ranging method based on a hardware-software integrated system, which composes a deep learning algorithm with a Jetson Nano-based sensor. Firstly, we use the existing model to realize the transformation from single-image to depth map. Secondly, we use the KITTI dataset to fine tune the depth transformation model. Thirdly, we invoke the physical camera and apply the reconstructed model to implement a joint hardware and software robot system. Finally, we conduct a series of experiments to verify the effectiveness of our customized hardware-software system with robot cars.

To improve the current situation of insufficient metro capacity during peak hours in large cities, it is necessary to accurately forecast the subway passenger flow. To solve this problem, first, we carry out the data pre-processing after selecting the appropriate time granularity. Then, based on the PyTorch tool, we implement the LSTM network, employ a training scheme, and search the best parameters preferentially. Besides, we add an L2 regularization to the MSE loss function and utilize the NAG algorithm for optimization. After that, to enhance the model's anti-interference ability to abnormal data, we randomly inject the Gaussian noise into the training dataset. Finally, a set of experiments are carried out to evaluate the performance of our proposed method. The experimental results verify the feasibility and robustness of applying the LSTM model to subway passenger flow forecasts.

Photovoltaic (PV) is a clean and renewable energy source and is considered an effective way to solve energy problems and reduce emissions. However, photovoltaic output is significantly impacted by different weather conditions. To improve the reliability and stability of photovoltaic resource, in this paper, we propose a photovoltaic output prediction approach to accurately deduct the future photovoltaic output capability. First, we construct a weighted undirected spatio-temporal graph to characterize the relationship between distributed photovoltaic sites. Then, we utilize the LSTM network to study time-series features of each node in the spatiotemporal graph. After that, we jointly design the edge and feature decoding architecture to learn the spatio-temporal features of nodes, based on which a multi-layer fully connected network is used to effectively infer the photovoltaic output of on-site nodes. Experimental results show that our proposed model outperforms other state-of-the-art models.

The technologies associated with the Internet of Things (IoT) have great potential for application in the domain of agriculture to increase production and reduce energy consumption. This paper builds a integrated vegetable supply system which integrates planting, storage, logistics and consumption through IoT technology. Using blockchain and web development technology, this system creates traceable smart contract to ensure food safety. Environment data collected by sensors are used to predict the growth requirements of the plants by Long Short-Term Memory (LSTM) neural network. A prototype system is created to prove the feasibility of the system.

Obstacle detection is a key technology for marine intelligent transportation system. In this paper, an obstacle detection technique for marine intelligent transportation system based on Maximum Likelihood (ML) processor is proposed. Different from most of the former source localization algorithms, our proposed algorithm explicitly makes use of both the horizontal L-Shaped array and the underwater acoustic propagation model to achieve a relatively high spatial resolution. Numerical simulations are presented to demonstrate the superior performance of our proposed algorithm.

Obstacle detection can help human to reduce risk on underwater transportation activities. In this paper, a joint DOA and range estimation technique based on compressive sensing is proposed. According to underwater acoustic propagation ray theory, the signal model for horizontal L-Shaped array is first constructed. Then, the joint DOA and range estimation problem in ocean waveguide is reformulated as a sparse representation problem and finally solved by an efficient toolbox CVX. Simulation results verify the feasibility and validity of our proposed method.

Combining data mining technology into housing price evaluation problem has increased great attention in recently years because it improves the prediction accuracy. To facilitate the application, this paper builds an urban second-hand housing price evaluation system based on our Bayesian probabilistic model under location submarket division. Using urban data such as house location, surrounding environment and point of interest (POI) information, a prediction model is constructed based on the second-hand house transaction data crawled from the network. It helps users get the price as well as location, POI information and characteristic attributes of the target house, and query suitable houses meeting some given requirements. The system provides visual display of query results and evolves by using query results.

Music is often described as the "language of emotion," and emotion prediction is important as it can impact future behavior. This paper proposes an audio-based emotion prediction model using a One-Dimensional Convolutional Neural Network (1D-CNN) approach, with Mel-Frequency Cepstral Coefficients (MFCCs) extracted as audio features. Preliminary results show an overall accuracy of 93%, but the imbalanced dataset used may cause bias in the accuracy of each emotion. Further research is needed to investigate the classification of audio features and 1D-CNN layers.

We proposed a novel approach for an automatic attack alerting system that is helpful for citizens. Our proposed model uses the YOLO object detection model, which can detect violent objects and the pose estimation model, which can estimate dangerous actions. Then, the output of these two models is entered into the LSTM model as input at the same time and the LSTM model is used to classify the type of attacks. The whole system is deployed into edge devices. Finally, based on the output from the LSTM model. If the threat is detected, the edge device can automatically alert the victim's family members or friends to minimize harm. The performance proves that the proposed model can detect the objects and accurately classify the attack types.

Retinal layer and fluid segmentation is a critical task in assisting doctors to diagnose retinal diseases. Manual segmentation by experts provides the highest accuracy, but it is time-consuming and inconsistent if segmented by different experts. Deep learning algorithms(e.g. Convolutional Neural Network(CNN)) have provided a faster way to perform segmentation through a computer-aided diagnosis system. Nevertheless, CNN has limitations, such as a limited receptive field and loss of details. In this project, we propose a transformer-based architecture to segment the retinal layer and fluid from retinal images. The architecture is based on Vision Transformer (ViT) and modified to improve performance. The transformer has been trained on a set of training retinal images and evaluated on a separate set of testing retinal images. The transformer-based architecture demonstrated a 0.01 improvement in average dice coefficient compared to the Unet architecture for fluid and layer segmentation. The Transformer-based architecture is better suited for deployment in commercial portable Optical Coherence Tomography (OCT) devices due to significantly faster inference speed. The proposed model is at most 4 times higher than that of the CNN family models. This makes it an ideal choice for resource-constrained environments where computational resources are limited.

Intelligent vehicles rely on millimeter-wave radar and machine vision to perceive their surroundings. However, the considerable differences in the features of radar point clouds and those of image pixels make it difficult for models to perform effective fusion. Moreover, high-frequency noise in images can impede the extraction of meaningful features. This paper proposes a novel 3D object detection method that combines millimeter-wave radar and RGB camera data. Our approach includes a gaussian filter for preprocessing, a hierarchical model architecture for fusing radar and image information, and a training stabilization strategy. We evaluated our method using the challenging NuScenes and Taiwan street databases and found that it outperformed the popular CenterFusion model in terms of detection performance. In addition, our method is applicable to a variety of scenarios in Taiwan.

Video-based point cloud compression(V-PCC)is a point cloud compression standard formulated by the Moving Picture Experts Group(MPEG) organization. The concept of this standard is to project 3D point cloud information onto a 2D plane and generate 2D image, namely geometry map, texture map, and occupancy map. Overlap 2D images to form 3D depth from three images, combine color and position information, and then encode and decode using High Efficiency Video Coding (HEVC) compression. However, the projected image will have obvious holes. These holes are regarded as high-frequency signal in the image, which will have a bad impact on the subsequent compression rate. It is necessary to use image filling to smooth the image, reduce high-frequency signal, and facilitate subsequent compression processing. Therefore, the purpose of this research is to develop a series of anti-noise procedures to fill and smooth images with High-Efficiency Video Coding(HEVC), including mean filter, Smooth Pull Push Algorithm(SPP), etc. This algorithm has been implemented in Moving Picture Experts Group(MPEG) release-v8.0, and the obtained data proves that although PSNR needs to be sacrificed, it can effectively reduce the number of compressed bytes after texture map filling.

The current society is in an era of vigorous innovation and development in digital media and artificial intelligence. We often see the sharing of our media, and this makes the external biometric information be at a high risk of exposure. However, once the biometrics are leaked, it is difficult to update and modify them. Therefore, a biological feature that is difficult to be exposed is necessary in the future. The vein in the human body has this feature, which makes it advantageous for live imaging. With the steady development of deep learning (DL) technology, an identification model can easily have an extremely high accuracy rate, but there are also disadvantages like a high parameter volume, calculation volume, and storage volume. These disadvantages cause the model to be unable to be effectively implemented in the real world. To solve the problems, this paper proposes a model training strategy combined with automatic augmentation, to achieve the advantages of reducing the amount of model parameter and improving the accuracy of the model. As results, the method of this paper can improve the accuracy of the model by 16.9% without changing the parameter quantity.

Fracture is common in clinical medicine, and doctors usually make the initial diagnosis based on X-ray images. Compared to general fractures, diagnosing rib fractures requires more experienced doctors, and the manual diagnosis method requires much time and effort. A computer-aided diagnosis (CAD) system for rib fractures would help reduce the pressure on doctors and the likelihood of missed diagnoses. However, the rib is a circular three-dimensional structure, so there are problems with overlapping ribs, unclear bone features, and edge features in X-ray images. To solve the challenges, this paper proposes a rib X-ray image enhancement method to enrich bone feature information and enhance bone edge information effectively. According to experimental results, the method in this paper can effectively improve the AP by 36.1%.

This research introduces a method for smoothing 3D Point Clouds, which is a two-stage algorithm designed for models that are affected by Gaussian noise. The first stage utilizes a median filter to arrange the positions of neighboring points and determine the most probable relative position. The second stage applies a bilateral filter to achieve surface smoothing. This algorithm adds a low-complexity median filter to the original bilateral filter, which can improve the Peak Signal-to-Noise Ratio (PSNR) by 14%, and reduce the Mean Squared Error (MSE) by 75%. Furthermore, this research preserves both the complexity of the bilateral filter and the number of neighboring point searches. This study represents a valuable contribution to the field of 3D Point Cloud smoothing technology.

With the rise of the internet, people tend to value their safety on the web more and more. This work proposes an EfficientNetV2 model combined using orthogonal projection loss function (OPL) to build a palm-vein recognition system. The experimental results indicate that the proposed model using OPL can improve the accuracy in the multispectral palmprint of the public dataset CASIA. Compared with other palm-vein recognition works, this work has a higher model recognition ability, with a 97.35% rate of accuracy in the CASIA dataset.

Fitness has been a popular and mainstream leisure activity in recent years. Interpersonal relationships play a significant role in fitness participation and are an essential part of fitness activities. However, in existing fitness-related mobile applications, the functionality is limited to recording user information and only plays a supportive role in fitness activities, lacking social interaction and unable to effectively increase fitness participation. This research aims to use machine learning-based AI's dynamic generation to produce content for fitness games and allow users to co-create, filling the social gap in past fitness games and thereby increasing customer engagement in fitness games. 30 subjects were recruited for this research to conduct two experimental measurements, including traditional fitness activities and AI interactive fitness games. By using paired group average tests to examine the result show that the AI assisted fitness game system had a statistically significant difference.

In this paper, a violin bowing gesture training system is proposed. Based on the positions of the skeletal joints of the hands of a user, we propose to use the long short-term memory (LSTM) deep learning model for violin bowing gesture recognition. According to the hands API of MediaPipe, a user can use a built-in camera to learn the violin bowing gesture in practicing. In the developed prototype, the overall violin bowing gesture recognition accuracy is 96%.

Road marking recognition is an important task for advanced driver assistance systems and autonomous vehicle. In the existing research, most techniques focus on detection mainly due to the lack of segmentation datasets. The segmentation of road markings can be used to understand the traffic regulations, as well as for vehicle localization. This paper introduces a road marking segmentation dataset, RLMD. It consists of 700 images annotated with 25 categories. The effectiveness of our dataset is evaluated using state-of-the-art segmentation techniques. The performance comparison is performed with well-known public datasets, Bdd100k and CeyMo. The dataset and code are made available publicly at https://github.com/stu9113611/RLMD.

We are in a fast changing world, and the medical system is constantly advancing with each passing day. Combining "technology" with "medical care" will become "I-healthcare" or "smart hospital". However, most hospitals currently do not use smart hospitals in depth, and at best automate medical operations, but it is really helpful for improving medical quality and patient safety, or as an aid to medical decision-making. This research uses the Analytic Hierarchy Process to construct the framework diagram of this research, and the content is designed by the research framework graph cost research questionnaire. The subjects of this questionnaire are hospital administrators, physicians, pharmacists, nurses, dentists, and related hospital research scholars or experts. 68 samples were collected in this study. The questionnaires are collected and analyzed through Expert Choice software, and the key factors of smart hospitals are explored through expert opinions.

Utilizing big data analytics to predict the election results has been a trend in election strategy development. However, it is questionable whether big data analytics can predict the results precisely. This research utilized a media big data analytics to exame the public opinion volume of Hsinchu City mayor election. It is suggested that the big data analytics platforms can be used to find the hot issues and exam hidden information during election campaign. However, even using big data analytics had reference value, the algorithm needs to be adjusted to predict the results.

Hyperspectral imaging has become an emerging research topic recently due to reduced form factor and low initial cost. However, these sensors and the optics still can't be compared with traditional imaging solutions in terms of resolution, sharpness, and Field of View. These limitations present challenges in the process of image stitching that is required for creating the orthomosaic image. This paper introduces the problem of identifying a low number of features using traditional feature detection methods in low-resolution hyperspectral imaging. Then, a state-of-the-art learning-based detector-free feature identification and matching algorithm is compared with the conventional methods for its ability to identify and match features in low-contrast and high-contrast hyperspectral images. Results indicate that the leaning-based approach outperformed the traditional methods and it is usable for image orthomosaic generation tasks in hyperspectral imaging.

Video understanding is a significant computer vision research subject since online video content is growing exponentially. Feature extraction and representation play a crucial role in video understanding tasks such as classification, segmentation, and recognition. However, the model's learning is ambiguous since adjacent video frames typically have similar RGB features. To address this issue, we present graph-based embedding to enhance video feature distribution. We construct a graph-structured of videos by connecting similar features. Node embedding is generated by utilizing a graph model. Experiments demonstrate that our approach effectively improves feature distribution. The graph attention network (GAT) improves accuracy and editing score by 4% over the visual model.

In this paper, we present a couple of examples of effective linearity improvement of millimeter-wave (mm-Wave) 5G Power Amplifiers (PAs) using Digital Pre-Distortion (DPD). The PAs are successfully linearized using 64-QAM /256-QAM 1 x 100 MHz 5G NR input at PAPR (peak-to-average-power-ratio) of 8 dB, and/or with 256-QAM 1 x 20 MHz 4G LTE input at PAPR of 8.7 dB. These examples suggest wideband DPD is feasible and may be attractive for future mm-Wave 5G and B5G/6G applications.

This paper proposes a method to predict the surface roughness of wire electrical discharge machining (WEDM) workpieces. The method first uses the Gramian angular field (GAF) and the Markov transition field (MTF) to encode the time series data into 2D images, which are then fed into the CBR long short-term memory (CBR-LSTM) neural network to predict the WEDM workpiece surface roughness, where CBR stands for the combination of the convolution, batch normalization, and rectified linear unit (ReLU) layers. The Keras hyperband tuner is used to optimize CBR-LSTM network hyperparameters. The proposed method is shown to have the best prediction when compared with related ones.

To realize high speed processing of parallel applications in high-end IoT devices, multi or many core processors are implemented on a VLSI chip. As the interconnection system of many core processors (nodes), Network-on-Chip (NoC) is attracted attention. In NoC, fault-tolerant routing plays an important role in communication between nodes. Many fault-tolerant routing methods have been studied with the evaluation of communication performance. However, the evaluation of only communication performance is not sufficient; it is also necessary to evaluate execution time of parallel applications. We evaluate the execution time of parallel applications for fault-tolerant routing methods. We use the parallel tasks generated by TGFF, and employed a genetic algorithm to obtain semi-optimal task mapping. The result shows that the fault-tolerant routing method called Passage reduces application execution time by about 11%, compared with the method called Fcube4.

Situs inversus is a rare congenital abnormality where the organs in the chest and abdomen are mirror-reversed. Due to the reversed orientation, medical interventions such as visceral surgery or organ transplant carry an increased risk, as the mirror-image organs can significantly affect the procedure. Failure to detect situs inversus before treatment can result in medical errors. This paper proposes a method that combines the outputs of YOLO and ResNeSt to see situs inversus in X-ray images, providing better diagnostic information and treatment options for physicians. Experimental results show that our method achieves a MAP of 99.85% on YOLO and a weighted accuracy of 99.99% on anomaly detection. This proposed method can effectively detect the possibility of situs inversus.It can serve as a reminder for healthcare professionals to recognize orientation during diagnosis and treatment, thereby reducing the potential for medical errors.

The prediction of pedestrians' intentions is an important topic in the transportation field. For example, in the automated driving application, by understanding the intention of pedestrians and implementing robust pedestrian detection and tracking, the risk of pedestrian collisions can be reduced. In this work, the experimental scene is on a busy street adjacent to the University of California, Berkeley. According to the absolute distance of pixels (ADPs), the outcome is classified into five intention levels. The accuracy result shows that our designed model has a better result than traditional and previously published neural network methods. In order to analyze the attention of deep learning networks, Gradient-weighted Class Activation Mapping (Grad-CAM) was implemented in the designed neural network. Traditional machine learning methods, such as Gradient Boost (GB) and Random Forest (RF), has the ability to determine important feature. Therefore, this work also compared the important feature of machine learning methods and the attention heatmap of Neural Networks.

Image inpainting has been researched for many years. From traditional methods to current CNN models, they all pursue two targets (structural stability and texture consistency). In this paper, we propose the multi-shift CSWin Transformer (MSCSWin Transformer) and the HSV loss to focus on colors to inpaint images for these two targets. At last, we compare our model with state-of-the-art methods on the Places2 dataset to confirm our proposed module is indeed working.

This paper presents a method for shoot event prediction in soccer considering expected goals based on the players' positions. To quantify players' and teams' performance, various ways based on the chance of shoot events have been proposed in recent years for soccer analytics. In soccer, since the players' positions in soccer change little with respect to the soccer court, it can be difficult to directly introduce the tracking data of players, that is, players' positions, into the shoot event prediction model. We tackle this problem with expected goals estimated from the field position as the player's importance. At the end of this paper, we confirm the effectiveness of our method through experiments using actual soccer videos.

The newest video coding standard, H.266/VVC, improves compression performance by about 30~50%, as compared with the previous H.265/HEVC, with the cost of higher time complexity.
We proposed to speed up the VVC intra-coding while maintaining good compression performances. Statistical heuristic and deep-learning methods are used in the coding tree structure to eliminate unnecessary rate-distortion optimization operations. Experiments showed that the proposed method can save 46.73% of execution time with 1.16% BDBR increment, as compared with the VVC test model, VTM 14.0.

In brain research, brain tissue segmentation techniques have offered vast aid and possibilities for quantitative analysis of the brain to detect brain tumor and other brain issues. In this paper, UNet-based multiclass brain tissue segmentation is presented. To achieve the best performance of the model, data augmentation techniques are applied. The training, validation and testing of the model is performed on the BrainWeb dataset. The model has achieved improved accuracy and IoU after data augmentation. Both pre and post data augmentation results are presented which show clear improvement in model performance after data augmentation. The achieved IoU for white matter, gray matter and cerebrospinal fluid are 0.84, 0.92 and 0.95, respectively. Our model accurately performed the multiclass brain tissue segmentation and this study proposes a valuable medical aid model.

Moving object tracking in satellite imagery is a hot topic in remote sensing. The size of the object's motion is tiny in scale compared to the entire frame. Because satellite images are in continuous motion, the momentum of the motion of the moving object is disturbed by the satellite motion. In order to obtain moving objects in satellite images, this paper proposes an improved algorithm for target tracking in satellite images based on the DBSCAN to track targets. The experimental results show objects are successfully tracked in satellite images using the algorithm proposed in this study.

This paper proposes a meta-learning classification weight transfer network to generate masks as a few-shot image segmentation architecture. It generates good prior masks via a pretrained classification weight transfer architecture, and uses a pretrained feature extraction architecture on query images and support images. The network architecture exploits a top-down path in a feature augmentation module to adaptively transfer information from fine to coarse features for extracting features from query images. Finally, the classification module predicts the segmentation of the query image. The experimental results show that using the mean intersection of joints (mIOU) as the evaluation mechanism, the accuracy of the 1-shot experimental results is 1.7% higher than that of the baseline. In the 5-shot experimental results, the accuracy is also improved by 2.6%. Therefore, compared with the base-line, it clearly shows that the mask generated by the me-ta-learning classification weight transfer network can effectively help improve the performance of few-shot image segmentation system.

This study presents to promote the behaviour of a celebrated speech enhancement framework, the Conformer-based Metric Generative Adversarial Network (CMGAN). CMGAN adopts a generative adversarial network (GAN) structure and employs a Conformer-based metric to evaluate the synthetic data from the generator work. In CMGAN, the loss function for the generator network is mainly from three sources: the time-frequency domain loss, the waveform-domain loss and the loss fed by the discriminator network. We argue that this loss function can be further revised by adding the phone-fortified perceptual loss (PFPL). PFPL reflects the mismatch between the synthetic data and the real data in the respective latent phone-level representation created by the wave2vec encoder. The preliminary experiments conducted on the VoiceBank-DEMAND task indicate that when integrating the PFPL in the learning of CMGAN, the objective metrics (PESQ and STOI) scores for the test data can be significantly promoted.

The prominent neural network-based speech enhancement (SE) methods, FullSubNet+, and its adaptive version, adaptive-FSN, employ a full-band and sub-band fusion model to fulfill the SE task. In these methods, the magnitude and complex spectrogram are exploited to learn the network that mainly comprises multi-scale time-sensitive channel attention (MulCA) modules and stacked temporal convolution network (TCN) blocks. To capture the phase information of input time-domain signals more simply, we propose using the discrete wavelet transform (DWT) features as an alternative to the complex spectrogram to be an input source to learn these SE models. The preliminary experiments conducted with the VoiceBank-DEMAND task indicate that exploiting DWT features in adaptive-FSN achieves higher objective speech quality and intelligibility in terms of PESQ and STOI metric scores, respectively, for the test set compared with the original adaptive-FSN arrangement.

In a cloud computing environment, a virtual machine placement (VMP) problem is considered for traffic load balancing, where each VM is placed on a physical host according to traffic loads in the physical hosts. In particular, VM is referred to as a node and a pair of nodes is considered a flow. To achieve load balancing, identifying flows that transfer heavy traffic volumes is necessary. We propose a VMP method with compressed sensing (CS) based traffic volume estimation. In the proposed method, all flow traffic volumes are estimated by CS from the traffic volumes that nodes transmit and receive. VMs are placed according to the estimated flow traffic volumes by solving a combinatorial optimization problem. We validate the proposed method via simulation experiments.

In this paper, we propose a method for determining the abnormal web pages using a Deep support vector data description (SVDD) and HDBSCAN (Hierarchical Density-based Spatial Clustering of Applications with Noise). Deep SVDD replaces the kernel of the SVDD with a neural network and minimizes hyperspheres containing normal data. The proposed method applies HDBSCAN, which is density-based clustering method, to remove abnormal data from the training data before using Deep SVDD. Through experiments on a web page dataset, we showed that HDBSCAN can remove anomalous data and that the performance of Deep SVDD is stabilised by using the removed data as training data.

In-network caching is a technology that allows routers in a network to cache contents, enabling users to download the contents from the routers in addition to content servers. As one of routing methods enhancing the performance of in-network caching environments, a multi-armed bandit-based routing (MAB-routing) has been proposed in the past. To further enhance the performance of MAB routing, in this paper, we propose a content prefetching method. In the proposed method, each router periodically retrieves contents according to their popularity, and then notifies its adjacent routers of caching information to efficiently utilize MAB-routing. Through simulation experiments, we show the effectiveness of the proposed method.

In this paper, we discuss the optimization problem of Unmanned Aerial Vehicle (UAV) placement for UAV-based edge computing systems. In UAV-based edge computing systems, UAVs have computing resources for dealing with tasks of clients. Two UAV placement problems that determines the location of UAVs with/without considering task offloading to cover all devices with a small number of UAVs are discussed in this paper. We provide Integer Linear Programming (ILP) formulations for the placement problems. Through numerical calculations, we evaluate the performance of the UAV placement based on the ILP.

In this paper, we propose a two-stage malware detection using calling API (Application Programming Interface)-based BiLSTM (Bidirectional Long Short-Term Memory). In the first stage, to identify the overall characteristics, we use all the calling APIs of the application as input to determine whether the application is malware. By contrast, in the second stage, we divide the calling APIs into small windows and analyze each window separately. Through simulation experiments, we demonstrated the effectiveness of the proposed method using two-stage detection.

This paper considers a human assistive system of the rehabilitation for lower-limb using the exoskeleton. When the patient feels the fatigue in the rehabilitation, the power assist is needed to continue and to stop safety. First, the proposed power assist equipment using exoskeleton is indicated. The DC motor, the torsion spring, worm wheel, worm gear are used. Next, the control system using PI controller is proposed. The encoder attached to the knee joint is used as the reference value and the encoder attached to the motor is controlled by the motor. Finally, an experiment of repeated flexion and extension of the knee joint is conducted. It is verified that the motor can be controlled to be same angle of the knee angle. Therefore, the power assist by the exoskeleton is useful for the rehabilitation.

This paper investigates the effect of bitrate control methods on QoE of multi-view video and audio streaming with MPEG-DASH. We adopt three bitrate control methods for conventional single-view video streaming to the MVV-A system with MPEG-DASH. We conduct a subjective experiment changing available network bandwidth and investigate the effect of the methods on QoE.

This paper enhances the robot movement control using force sensor which was previously proposed for cooperative work between two remote robot systems with force feedback. The control alleviates large force applied to the robot arm in the direction of sudden change in the robot position owing to obstacles such as holes and stones by taking account of mobile robots. The purpose of the enhancement is to reduce large force in multiple directions when the robot position suddenly changes. By experiment, we demonstrate that the enhanced control can largely alleviate the force.

In this paper, we examined the effect of force adjustment control on cooperative tasks performed by two remote robot systems with force feedback. Two force adjustment controls were introduced: manual tuning and control acquired through reinforcement learning. The experiment involved an operator who carried an object between the robot arms from both sides. The results showed that both of the two force adjustment controls improved control performance.

In this paper, we have proposed a smartphone application for evaluating its upload speed. We implemented the application and evaluated the upload speed results for an actual 5G mobile network. We proved that the application was capable of evaluating the performance of mobile networks.

In this study, we implemented an uncompressed 8K-video frame conversion software to evaluate the availability of Message Passing Interface (MPI) for real-time processing. We measured the processing time using an actual PC cluster and MPI environment for the conversion. We demonstrated that this environment is suitable for real-time processing.

We propose a novel travel route control for autonomous mobile robots (AMRs) ad hoc networks. In this method, relaying AMRs (referred to as "nodes") can provide high throughput for a longer time by relaying other nodes data and taking a route that mitigates interference from external networks. Relay nodes form the ad hoc network as they move. At this time, it takes a longcut route on its own path for the best performance, so that it can take the best position for relaying. In ad hoc networks formed with moving nodes, it is difficult to comprehensively consider effects of influence from external network and effects of communication distance with adjacent communication node. Thus, it is also difficult to set up an optimal route in which relay nodes can take a longcut route and take the best position for relaying. The evaluation results showed that the proposed method can maintain high throughput in 72% of the routes.

In this paper, we proposed a method to improve the accuracy of stock price forecasts using Prophet, an open-source software for time series analysis, by taking into account the date of closing of accounts or the trends of stock price fluctuations of multiple stocks. In stock price forecasting, it is important to capture recent trends because price fluctuations cannot be predicted. However, Prophet automatically selects trends in time-series data, which causes trends to diverge from the current situation. We have identified that the days when price fluctuations are most likely to occur are the fiscal days on which the company's information and status are publicly announced, and that by considering the days on which the company closes its accounts, it is possible to bring the trend closer to the time of the forecast, thereby improving the accuracy. We conducted stock price forecasts for multiple individual stocks with the addition of fiscal year end dates, and found that stock price forecast accuracy improved for approximately 76.2% of all companies, and up to approximately 19.9% more accurate than when no fiscal year end dates were added. In addition, we verified the improvement in stock price forecasting accuracy by focusing on changes in social trends given by the state of fluctuation of stock prices in the same industry and affiliated industries. The addition of multiple stocks improved stock price forecasting accuracy by approximately 9.84% from stock price forecasting for individual stocks only.

In this paper, we investigated on whether adding learning data based on parametric eigen-space method that allows highly accurate classification with a small number of video image data, based on a small amount of initial learning data, would enable re-generation. As a result, the proportion of discernible data increased in all experiments. In addition, we conducted experiments based on somewhat inaccurate and correct judgment results, and confirmed that it is important to carefully select additional data to be added to the learning data.

Strong and efficient cryptographic systems are essential for secure consumer electronic devices. In this paper, we propose and evaluate a modification of bit-flipping decoding called gradient bit-flipping decoding, a decoding method for the QC-MDPC McEliece PKC. Simulation results show that the gradient bit-flipping decoding has lower error rates than other methods when the number of errors is small.

The authors focus on a Cyclic Vector Multiplication Algorithm that can randomize a basis parameter for each encryption. On the other hand, the calculation cost of a vector multiplication depends on the basis parameter. The calculation cost should be constant for a strict security such as a countermeasure for side-channel attack. The authors propose a method for averaging the calculation cost for vector multiplication.

In the first half of 2020, due to COVID-19, our university cannot offer face-to-face classes to students. Therefore, to conduct practical education on embedded system development remotely, we built a system that allows remote control of development equipment from students' home PCs.

We propose the concept of a profiling guided optimization scheme to boost the performance of quantum simulations. The key idea of the proposed scheme is that a quantum simulator is used to collect the quantum-wise information during the simulations, and the collected profiling data are fed to a quantum compiler to guide the generation of a more efficient quantum code version achieving higher simulation speed. Our result shows our proposed scheme achieves a speed-up by a factor of 1.36 when simulating a 28-bit quantum circuit.

Consumer electronics pursue the balance of power consumption and performance, especially for today's AI consumer electronic devices. Reducing power consumption while maintaining system reliability is a significant challenge. At the same time, artificial intelligence edge computing devices, such as Artificial Intelligence of Things (AIOT) devices, are expected lower memory occupancy and consumption for Convolutional Neural Networks (CNN) computing devices. In this paper, we design a model with certain reliability under low voltage composed of two parts: protected bits and unprotected bits. The system's reliability with SRAM as memory is improved, and we analyzed the relationship between the Bit Error Rate (BER) and the recognition rate. We present a circuit model that can reduce power consumption while maintaining a high recognition rate.

For real-time image recognition, YOLOv4 is currently the most widely used object detection algorithm. However, while the technology is advancing, the problem of improving misjudgment must also be faced at the same time. This study will conduct experiments on remote sensing images, and discuss the effect of data augmentation methods that can improve the amount of data and model performance applied to remote sensing images. Mosaic is a data augmentation method proposed by YOLOv4. Mosaic mainly stitches pictures together so that the model can learn multiple pictures at once and improve training performance. However, because of the way the pictures are synthesized, it will have adverse effects on remote sensing images probably. That is, the interaction between the object and the background makes Mosaic unable to fully play its role, so this study will also explore the different effects of subsequent training and detection caused by differences in stitching methods, and how objects in remote sensing images cause models may not be successfully identified, and propose a data augmentation method to solve the above issues.

COVID-19 has brought many challenges to higher education, such as remote cheating derived from remote assessment, students are easy to obtain cheating channels and resources, and the burden of remote invigilators being large, all of which make the fairness of remote assessment highly critical question. In order to improve the problems existing in remote assessment, this paper proposes a deep learning artificial intelligence proctoring system (AIPS), which detects whether there is any abnormality in the answer through the webcam gaze and eye tracking. Behavior, combined with Long Short Term Memory (LSTM) to design a deep learning model that can be used to detect whether there is an abnormal answer in real-ter the test is over, a report can be output, including eye movement answer hot spots, test question answer order. This paper develops AIPS which is expected to assist invigilators to find candidates with abnormal answers through artificial intelligence assistance, so as to improve the fairness of online assessment.

What news is factual and fake has formed a hot topic in recent years. The current discussions concerning fake news indicate that academics and practices still need more consensuses on measuring the influencing factors. Therefore, the purpose of this study is to investigate and assess influence factors to reduce the gaps in spreading fake news by interdependence and feedback problems among dimensions and criteria to find the most significant factors. We used the DEMATEL-based Analytic Network Process (DANP) to solve these problems. Then we conducted several experts to investigate the root and cause of the spreading of fake news. These results can provide the government understanding of how to prevent strategies and fill the gaps between academics and practice.

As headlines of Internet news become out of context. This project is expected to use the GPT-2 model and use Chinese news materials for training. And use the model to generate in-text headlines instead of exaggerated headlines. This research will use web crawlers to collect news data in Taiwan and use Pearson's correlation coefficient to detect the correlation between manual ratings and machine ratings and verify whether this experimental model can produce headlines of sufficient quality through a variety of research tools.

This study proposes a computer-aided diagnosis system for identifying benign and malignant breast cancer on mammograms to assist physicians. We use MIAS and DDSM databases as source image data. The process includes image preprocessing, region of interest (ROI) extraction, and tumor classification. During image pre-processing, image contrast will be enhanced, and useless color blocks such as chest muscles will be shaved off. In the ROI extraction, the suspicious tumor area is separated from the rest of the background area. The resulting image is the input data of the subsequent classification method. The classifier uses the extreme learning machine (ELM). Finally, the proposed method will be compared with artificial neural network (ANN) convolutional neural network (CNN). Experimental results show that the classification accuracy of ELM and CNN in the DDSM database is also 68%. However, The model training time of CNN is seven times that of ELM.

Owing to the rapid development of artificial intelligence (AI) technology in recent years, numerous scholars have applied it to classify arrhythmias using electrocardiography (ECG). However, raw ECG data contain high- and low-frequency noise, which can affect the accuracy of the classification results. To address this issue, we propose using autoencoder processing to reduce noise. Furthermore, most previous studies have utilized a segmented one-dimensional time series to train the AI classification model. In contrast, our approach involves transforming the segmented time series into a two-dimensional image, which allows for better feature coverage to highlight the differences between the various types of arrhythmias. This approach can improve the accuracy of the AI classification model. Finally, we developed a portable ECG arrhythmia classification model by integrating the D1 mini development board and AD8232 ECG sensor. We worked in cooperation with the doctors of Dalin Tzu Chi Hospital to perform field tests on the model in the hospital.

Virtual reality technology is advancing quickly, and its applications are becoming increasingly diverse. In this study, virtual reality technology is combined with computer graphics rendering technology to simulate a virtual ultrasound examination for students to scan the liver and gallbladder, just like in a real-life scenario. Through this simulated ultrasound examination, students can practice ultrasound examinations in an immersive virtual environment without the need for expensive equipment and real patients. The system only requires a virtual reality device and a pillow to facilitate all virtual simulation scenarios. Thus, this system is feasible for providing a virtual ultrasound examination simulation that eliminates time and space constraints and reduces material costs. It also makes ultrasound examination simulation training more convenient and effective for medical students.

This paper presents a 12-bit 45.5Ms/s SAR ADC with TSMC 0.18µm CMOS process. With the proposed architecture, a multi-segment split capacitor DAC was used to reduce the total capacitance and the possibility of capacitor switching. To reduce the swing range of the common voltage, hybrid capacitive switching was applied to the MSB and monotonic capacitive switching was applied to the LSB. In the CMOS process, split capacitors cannot achieve accurate fractional values. Therefore, non-fractional values are also used to solve the problem of capacitor mismatch. The final measurement data shows INL ranges around -0.71 LSB to 0.64 LSB and DNL ranges around -1.0 LSB to 1.032 LSB. Total capacitance reduced by 75% relative to capacitive DAC.

With the rapid change in technology, portable electronic products are designed to extend battery life through low power consumption and small size. The low-dropout regulator (LDO) is a crucial power management block and it is widely used in electronic products. Although the conventional LDO has the ability of voltage regulation, it cannot achieve low power consumption and fast transient response at the same time. In this paper, an output capacitor-less LDO with a fast transient response was proposed based on the extract power supply ripple technique. By adding a proposed voltage buffer and low-pass filter, the transient response was further improved. Moreover, a cross-coupled technique was applied to enhance the power supply rejection ratio (PSRR). The proposed regulator was implemented by TSMC 0.18 μm CMOS process technology and simulated with a 1.8V power supply. From the results, the proposed LDO achieved a fast transient response of 0.27 μs and a high PSRR of -76 dB at 1 kHz.

This work presents a circular polarized antenna used to increase the communication distance between small unmanned ships on the sea and the control station on land. Compared with popular commercial linear polarized antennas such as monopoles, the proposed antenna can effectively resist destructive interference caused by sea reflection, especially for small unmanned boats whose antennas are very close to the sea. The field test shows that the designed antenna equipped at the control station on land can achieve a longer communication distance than the antenna in the commonly used commercial wireless remote control system.

Deep ultraviolet (DUV) communication shows great potential in future wireless communication due to its non-line-of-sight (NLOS) security, omnidirectional communication links, Solar-blind for outdoor applications, and low background noise. In this work, we demonstrated a simple method to achieve the response wavelength tunable Deep UV photodetector for NLOS Ultraviolet Communication, which could facilitate the ultraviolet optical signal demodulated/restored or opto-electrical (O-E) conversion and meets the need for wavelength division multiplexing in deep UV spectra.

Nowadays, diversified smart door access control gives people privacy and security in life. But when an emergency occurs, will the door access control system become an obstacle to escape? In this study, our door access control system with sensors can detect smoke and earthquakes to determine whether an emergency has occurred. As soon as an emergency is detected, the door lock is only unlocked, not opened, maintaining the privacy, and increasing the chance of escape and rescue. Furthermore, access control uses biometric technology to unlock the door, prevent the situation when the key is lost and easily duplicated.

In this paper, we introduce a method for generating well-balanced sequences by two types of the logistic maps over integers with a small arithmetic precision. We proposed this method last year. We think that the randomness and unpredictability of the sequences are caused by a simple combination of partial outputs of two symmetric mapping, however we cannot explain the reason in detail. Here, we show that the sequence has superior properties from results of typical randomness tests.

This is a success report of an experiment of an attack for an ECDLP of a Pollard's parallel rho method with skew Frobenius mapping for a 116bit prime order Barreto-Naehrig curve.

Domain Name System(DNS) is a crucial component for us to access the Internet. However, it is vulnerable to several attacks such as DNS Cache Poisoning Attack and DNS DDoS Attack. Because of the high security promised by the Blockchain technology, we investigate the domain name systems based on blockchain, and focus on Ethereum Name Service(ENS). In this paper, we propose an ENS-oriented light node in order to solve the Full-node problem which means each participant has to hold a large amount data.

One of our research purpose is to evaluate a symmetric key encryption scheme by using a machine learning. As the first step, this paper shows results of a known plaintext attack by using the neural network against Data Encryption Standard (DES) as the symmetric key encryption scheme. A plaintext guessing attack from ciphertexts against 16-round DES did not yield good results. On the other hand, we report interesting results in their attacks against 1- and 2-round DES.

In recent years, CAPTCHAs have been used in a variety of online services. A typical example is reCAPTCHA v2, which uses a system to select the correct answer from a set of images given in a question text. However, there is a possibility that a computer program can break through the CAPTCHA by automatically recognizing each image in the image set and recognizing the subject of the image. In this paper, we propose a CAPTCHA that uses processed images that are difficult for search engines and AI to recognize. Experiments will be conducted to verify and evaluate the proposed method's high accuracy rate for humans and low accuracy rate for Google's reverse image search and Google lens. We will discuss the issues that have been found.

This paper proposes an indirect-permission approach to address the circular protection issue and collusion problem in private key backup and recovery processes. The approach separates the management of permission and possession by escrowing only the permission of the backup to the trustees, using a randomly selected symmetric key. The trustees are prevented from possessing the secret key even after the owner requests recovery, and the approach solves both the circular protection issue and collusion problem. The proposed method provides a secure and efficient solution for backup and recovery processes, which is crucial for ensuring data security in modern information systems.

Fully Homomorphic Encryption (FHE) can directly perform desired operations on ciphertext, but the efficiency of various FHE systems is highly dependent on how effectively the homomorphic multiplication is implemented. The Number Theoretic Transform (NTT) is a widely used method for carrying out large-sized multiplication. This work aims at developing efficient hardware design of Bluestein's NTT for the latest BGV-FHE. Primes with low Hamming weight are selected to optimize the Barrett modular multiplier design and memory requirements. The experimental results demonstrate a significant improvement in the area requirement compared to the related work.

In this paper, we propose a new pseudorandom bit sequence generator, which is based on a method digging out walls. Although each of these sequences lacks enough randomness, synthesized sequences of them can pass all of the NIST statistical test suite. We also describe a conjectural relation between the number of types on the square mazes and their sizes.

With the development of quantum computers, widely used cryptosystems will be able broken in the near future. Therefore, researches on Post-Quantum Cryptography (PQC) has been actively conducted. In this paper, a method to improve the calculation of Streamlined NTRU Prime, which is one of the PQC. The authors propose to employ the Cyclic Vector Multiplication Algorithm (CVMA), which uses a normal basis called Type-II Optimal Normal Basis, with the parameters of Streamlined NTRU Prime. The processing speed of multiplication and inversion are compared with those of previous studies. As a result, the multiplication by the CVMA was 58% faster than the previous method with a certain condition. On the other hand, the inverse calculation was 50 times slower than the previous method due to the lack of optimizations.

Based on the 2-party XOR protocols, we consider an extended protocol that allows the input of a third value "indeterminate" which is neither 0 nor 1. Assuming only algebraic structures that satisfy the associative law, we classify patterns that can be constructed as boolean tables, and report on the feasibility of implementing the protocol as card protocols.

This paper proposes a deep learning multi-objective optimization (DL-MOO) method for smart manufacturing to generate wire electrical discharge machining (WEDM) parameters via an elitist genetic algorithm with the deep neural network (DNN) and the transfer learning mechanism. Given multiple objectives, the proposed method can generate WEDM manufacturing parameters to optimize the multiple objectives at the same time. Practical experiments are conducted to validate the proposed method.

Photoplethysmography (PPG) is an optical method widely used for detecting changes in blood volume in the microvascular bed of tissue. Current contact-based PPG measurement methods are mature and stable but can be inconvenient for subjects. Hence, researchers have been exploring non-contact measurement methods to generate PPG signals. Achieving a balance between PPG signal quality and measurement convenience through non-contact methods remains a challenge. In this paper, we propose a non-contact image-based PPG (iPPG) generation method that uses an optical camera to capture video from the subject's wrist and an image recognition scheme to generate iPPG. For subjects with noticeable wrist skin deformation due to pulse beating, the outline and signal quality index (SQI) of the generated iPPG were consistent with their contact-based PPG (cPPG).

A shunt is placed in hemodialysis patients. However, the vessel could be stenotic. Auscultation at the location of the stenosis is required because capturing the features of the stenosis away from the location of the stenosis is difficult. In this study, we attempted to improve detection of vascular stenosis by screening out uncertain data from shunt sounds measured at multiple locations.

The Pulse Transit Time (PTT) is often considered a reliable surrogate for noninvasive monitoring of blood pressure (BP) and its fluctuations. One of the primary limitations of PTT estimation accuracy is the inclusion of the pre-ejection period (PEP) artifact in the PTT signal. The second method, using dual-point photoplethysmography (PPG) to replace electrocardiogram (ECG), can eliminate the confounding PEP component and improve the accuracy of PTT estimation. However, few studies have utilized multi-point PPG for BP estimation, and the weak correlation between PTT and BP is often attributed to the lower accuracy of PTT. To address this issue, we propose a PTT acquisition method for Multi-point Photoplethysmography architecture to improve accuracy. Experimental results demonstrate a maximum PTT error of less than 1.3 ms, indicating high precision. Our approach is easy to implement and has the potential to improve the accuracy of PTT-based BP monitoring.

Artificial intelligence has access to every field in this era. Currently, you have access to everything, from simple tasks to quick calculations. The construction industry is one of them. Safety work, installation, and construction are also part of the drive. Demonstrating the pipeline to date does not prepare as comprehensive an assessment as it could. To this end, we benchmark performance using several cutting-edge approaches that have recently the best performance from state-of-the-art method studies, such as YOLOv5x, YOLOv6l, YOLOv7x, and YOLOv8x. The result show that the recent YOLOv8x accomplish the most effective at generating region of interest box comprehensively. Whereas some existing ap- proaches, like YOLOv5x and v7x, get the highest capacity at classification instead.

This paper introduces a room in a lavatory space with Ultraviolet-C radiation to maintain a sterile environment. Usually, the cleaning and disinfection of a lavatory room is performed by the staff, and users are unaware of which room is in a better condition if there is more than one room available in a lavatory space. In the study, a more efficient disinfection process is used and a display is installed on the door to inform users about the state of the room. With the proposed design, users could have a friendlier experience and each room in the lavatory space might be used more evenly. Furthermore, it could save much time to manage public lavatories if the state of a room updated with real-time detection could be read in the cloud.

This study presents a mental illness prediction model that combines a retrieval-based chatbot with a Deep Neural Network (DNN). The proposed DNN uses the natural language model of mental RoBERTa to extract sentence feature vectors, and a refined-attention transformer encoder that includes a refined highway and sorting layer to enhance the feature vectors further. The classification task is performed using a fully-connected layer. Our DNN achieves an accuracy of 88.98% and a Macro F1 score of 55.23% on the datasets in references [1] and [2], respectively, demonstrating superior performance compared to conventional models. Furthermore, we applied our DNN to actual doctor-patient consultations, obtaining promising results. Therefore, the DNN proposed herein is highly suitable for early detection of various mental illnesses, and can be effectively integrated with a chatbot for various application services.

This paper proposes a credit card fraud detection method based on DeepInsight and deep learning. The proposed method first uses the DeepInsight mechanism to transform non-image credit card transaction data into well-organized images. It then feeds the images into a parallel convolutional neural network (CNN) deep learning model to extract critical features hidden in the raw data for detecting credit card frauds. The European credit card transaction data are employed to evaluate the performance of the proposed method. The evaluation results are compared with those of related methods. The comparisons show that the proposed method outperforms the others in terms of the accuracy, true positive rate, true negative rate, and Matthews correlation coefficient.

The fingerprint-based indoor localization (FIL) method has good localization accuracy. It performs indoor localization by collecting beacon signals of beacon nodes at different reference points to form beacon fingerprints, where beacon nodes are pre-deployed at specific locations. To improve the localization accuracy when FIL methods are applied to new application domains, exemplified by those using new beacon signal receiving devices, it is necessary to perform localization calibration. This paper proposes a transfer learning indoor localization calibration (TL-ILC) scheme based on weight-freezing of the transfer learning concept. The proposed scheme needs to collect only a small amount of data to fast retrain the artificial neural network used by the FIL methods. Experiments show that TL-ILC can indeed perform localization calibration to achieve good localization accuracy.

Basketball analysis systems are essential tools in modern basketball, where identifying the ball handler is one of the most critical tasks. The reason for this challenge comes from the overlapping of players in basketball, which makes it easy for the analysis system to misjudge the ball handler. We found that it is easy to misjudge ball handler using traditional algorithms, such as calculating the degree of intersection over the union or calculating the coordinate distance between the player and the ball. In this paper, we propose a transformer-based object relationship finder to classify the relationship between players and the ball, which uses features of different objects, such as the use of coordinate information and skeleton information as inputs, to learn the relationship between players and the ball through selfattention. Experimental results show that our method achieves an accuracy of ball handler up to 91.2% based on a smaller dataset, surpassing the 83.9% accuracy of traditional algorithms and the 77.8% accuracy of Resnet-based convolutional neural networks.

In this paper we propose a complete end-to-end network architecture to estimate 3D hand mesh from RGB hand images. We use ResNet-50 to extract the image features, and for better regression of model parameters later, we obtain some 2D feature maps, such as 2D heatmap and mask images, through some convolutional layers. In the model parameter regression part, we use the fully connected layer for iterative regression of the model parameters.

The study conducts a psychophysical experiment to explore the subjective emotional perception of different colored light sources in a beauty aromatherapy condition. The results show that the illuminance level, chroma, and hue of different colored light sources obviously influence subjective emotional perception. In addition, the study also established a perceptual evaluation model based on the experimental data that can be used to design more suitable lighting conditions for beauty aromatherapy settings.

For coffee farmers, the yield of coffee is an issue that will affect their income. At present, the method used to estimate the coffee production is usually to estimate the production of the entire coffee garden by weighing the coffee till the harvest; however, this method does not have a way to understand the reasons of affecting coffee production in each period of the planting process. In this paper, we propose an intelligent system using a crowd counting model to estimate the yield of coffee, so that farmers can observe changes in the amount of coffee during the planting process from CCTVs remotely. Experimental results show that the proposed method can obtain estimated results with mean absolute error about 6.91.

In the incremental learning technique, the accuracy of the model is improved by increasing the number of training samples. However, there are some problems when making this training mode. First, it is not known whether the newly added training samples will affect the learning effect of the model. Second, we know the result of the model identification, but we don't know what features are used to identify this object. In this paper, we propose a training dataset curation approach to improve model accuracy with incremental learning. To select the most appropriate explainable AI (XAI) method, we use the IOU metric to measure the range calculation of real objects and model features to know the rationality of the model explanation. As we know that increase suitable training samples can improve the accuracy of the model, we added the XAI method to filter out new samples that are not suitable for the training set. We use the crack identification dataset in our experiments, and the results show our accuracy improves to 0.93.

It takes a lot of manpower to clean up beach litter. This study proposes to identify the beach trash using deep learning method and combine Arduino to control the forward direction. In the process, it is correctly judged whether it is trash, and these data are input into the main control system "Raspberry Pi" and combined with Arduino to control the motor, set the operating range of the beach garbage collection vehicle, and complete the purpose of automatic cleaning of beach garbage.

As a newly developed industry, e-commerce not only develops fast, but also conforms to the trend of the times. For example, Amazon, the leading e-commerce platform, has attracted much attention because of its fast development speed, outstanding price advantage, and rich experience. The outbreak of the COVID-19 epidemic in 2020 will have an immeasurable impact on the global economy, society, and life. Similarly, the epidemic has brought many challenges to Amazon, such as the interruption of the industrial chain, but opportunities and challenges often coexist, and Amazon also hides opportunities for development. This paper adopts literature research method, questionnaire survey method and SWOT to explore how Amazon responds to the challenges brought by the epidemic under the background of the global spread of the COVID-19 epidemic, discovers and seizes the hidden opportunities under the epidemic situation, and improves its own development. To promote the upgrading of the e-commerce industry. The research and analysis of this article not only meets the actual needs of the development of contemporary e-commerce, but also provides a certain reference for cross-border e-commerce platforms after the epidemic.

Under the COVID-19 epidemic, the original living habits of the public have been changed. The live broadcast e-commerce creates an environment in which merchants can conduct real buying and selling transactions with buyers. In particular, the real-time interaction between buyers entering the live broadcast room and anchor sellers, such as bargaining, allows buyers to experience the normalization of everyday life. Tiktok is one of the largest domestic social platforms, video platforms, and traffic pool application software. Therefore, this study focuses on the development of Tiktok's live streaming e-commerce during the COVID-19 period. This article uses the literature research method to collect relevant domestic and foreign literature and organizes and analyzes the development status and strategies of Tiktok's live streaming e-commerce during the period of new crown pneumonia, and then uses the results of literature collation to conduct a questionnaire survey on its existing problems, and then use SWOT to analyze the survey results. Its research found that Tiktok's live delivery of goods is still developing rapidly during the new crown epidemic, but there are also certain challenges.

Nowadays, many scenarios require 3D positioning. The current application of global positioning systems in indoor environments is unlikely to meet the requirement for high-precision positioning due to the influence of obstacles. Therefore, this paper proposes to use the ultra-wideband (UWB) positioning module for 3D positioning measurements in indoor environments. Based on the characteristics of UWB, the positioning error can be reduced to less than 1 m. Regarding equipment, 3D coordinates can be positioned only by using four UWB positioning modules as three anchors and one Tag. Compared with Wi-Fi and BLE positioning methods, the proposed method can achieve simple equipment deployment and better positioning accuracy.

There is a large amount of inter-frame local spatial-temporal redundancy in natural videos. This paper presents a survey on the current significant temporal redundancy removal technologies necessary for real-time AI video object detection especially for mobile platforms. Temporal redundancy removal has once made great contributions to video data compression. However, the methods used in video data compression cannot be applied to video object detection directly. The same basis is that each frame in videos are classified as key frames or non-key frames, and then different types of computations are applied for each frame category. Two major directions in using the temporal information in video object detection are developed, i.e., exploiting the data redundancy between consecutive frames to increase execution speed, and aggregating the temporal features to improve detection accuracy. This paper gives a quick but in-depth analytical overview for the state-of-the-arts in this field.

In emergency medicine, CardioPulmonary Resuscitation (CPR) is usually a necessary and urgent first aid measure for patients whose heart stops beating due to diseases or accidents. However, the success of CPR in saving more patients depends on the popularity of CPR education and training. The more popular the education and training is, the more people can become rescuers in emergency situations and perform CPR on patients to improve their survival rate. In this work, we implement a CPR quality evaluation system, in which there are four modules, they are the pose detection module, the compression frequency module, the compression depth module, the visualized feedback module, respectively. In addition, we build the visualized real-time feedback system to present the overall information of chest compression of learners, including that whether the compression pose is correct or not, the frequency of compression, the depth of compression and the position of compression.

In this paper, a home care system with location information is designed. From this we can learn from the physiological information and location information of the elders. The physiological information part includes the measurement of blood oxygen and pulse value and the medication information. The medication information part is realized through Image Recognition technology, by using Yolov3 for drug identification. The location information part uses Wi-Fi wireless transmission signals to do indoor positioning (Indoor Positioning), measures the received signal strength (RSSI) of the distance between the elder and the wireless signal sender, and judges the elder's location through the measured results.

An all-digital DC-DC boost converter using a digital voltage control delay line is proposed for voltage conversion and regulation of indoor solar cells. This chip designs a low-input voltage boost start circuit and integrates a time-to-digital converter and a digital PWM circuit. The chip is realized in a 0.18 μm CMOS process. After the IC layout, simulation results show that the input voltage ranges from 0.35 V to 0.7 V, the output voltage is 1.8 V, and the peak conversion efficiency can reach 90 %.

n this paper, the Ultra-Wideband (UWB) technology Angle of Arrival (AoA) are combined with Adaptive Offset Cancellation (AOC) method to improve indoor positioning performance for Non-Line-of-Sight (NLOS) environments. Experimental results show that the positioning error can be greatly reduced by the proposed MAOC (Moving Average filter AOC) for the test points (TPs).

This study designs an Internet of Things (IoT) system for LCD Module (LCM) aging detection and analysis, aiming to achieve the detection and analysis functions for the LCM manufacturing process and provide intelligent quality inspection of LCM. In this system, a custom- Compound LCM Drive Measurement Circuit Board (CLDMCB) is used to drive different types of LCM. For the analysis of LCM aging status, the CLDMCB is used to measure the voltage and current values of the LCM at specific screens, and the detected voltage and current data are uploaded to the cloud platform. Finally, the voltage and current values are presented with data visualization, and defect detection and analysis of the LCM are performed.

The proposed system's framework applies the You Only Look Once (YOLO) method to perform preliminary testing and verification of Synthetic Aperture Radar (SAR) images from the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database. The results show that YOLO3 and YOLO4 can identify MSTAR data well. However, for SAR images, YOLO3 performs better: the mAP is relatively high, and the performance is stable and consistent, i.e., lower standard deviation.

This work presents a cost-efficient hardware architecture design of 2-D sliding discrete Fourier transform (SDFT). The proposed design requires the lowest eight real adders and six real multipliers in hardware resource, compared with Park's method reduced by 11.1% and 25%, respectively. In the FPGA implementation, the proposed hardware accelerator is operated at 47.47 MHz, and then it is very suitable for time-frequency analysis in real time.

By adopting the Online to Offline (O2O) commerce strategy, this paper designs and implements the Bluetooth Low Energy (BLE) indoor positioning technique for an Android APP to let the customer browse online and promote the offline shopping conversely. This APP supports Bluetooth indoor positioning, multi-floor route planning, push message notification and various feedback mechanisms. Hence, it is able to locate current position of the customer in the department store and navigate the customer to each retailer for buying the commodities ordered online through the shortest path.

The purpose of a secure communications is to enable the authentic destination station successfully receive the information from the source. At the meantime, a secure system protects from the eavesdroppers to interpret this information. Traditionally, the encryption, authorization and authentication schemes could be employed for protecting the eavesdroppers. However, the new and unanticipated algorithmic attacks have been conjectured developed. According to Shannon's perfect secrecy, it made the limited assumption that an attacker could access to an error-free cryptogram. Hence, it is expected that the secure communication could be achieved based on the physical-layer security. Besides, physical-layer security coding is not only to obtain a critical level of security, but also to achieve an arbitrarily low probability of decoding error. For a security concern, increasing the secrecy capacity leads a higher level of security ability. However, it might suffer the system performance without bandwidth efficiency. Hence, the tradeoff between the security level and coding rate is important to lead a secure communication.

Orthomode transducers (OMT) are a key passive microwave device in satellite and wireless communication systems. Their function is to combine or separate two orthogonal signals received and transmitted by dual-polarization. The main structure of the orthomode transducer considered in this paper is the turnstile junction. The design of the OMT combines a turnstile junction, E-plane bend, H-plane bend, and Y-junction, and it exhibits good return loss, high isolation, and low cross-polarization. We discuss its tolerance (< 0.05 mm) by analyzing misalignment between its top, middle, and bottom sections. The transducer does not exhibit significant spikes in response to high order model reflection but an offset of 0.1 mm results in obvious spikes.

VR can be used for creating situated learning environment, so most of research focused on applying VR on teaching and learning. Nevertheless, learning new information technology, including VR, is the educational target in senior high school. To engage the learners in learning information technology is an important educational task. Therefore, this study developed a systematic learning modular for VR technology based on learning by doing strategy, conducted a one-semester learning experiment in a senior high school, and investigated the learners' perception towards VR technology learning. The results revealed that the learners showed high motivation and perception towards VR technology learning. However, some of learning dilemmas should be reduced in the future.

Fifth generation (5G) New Radio (NR), was designed to offer more flexibility to satisfy new service requirements. However, this implies the increasing complexity of maintaining and configuring cellular network. Meanwhile, deep learning has proven successful in a variety of tasks, such as natural language processing and computer vision, in particular, which is proven to have a performance that scales with the total amount of training data and with faster inference time than traditional optimization algorithm. In NR, beam management is a procedure that chooses the beams to combat path loss attenuation according to signal strength, which is more critical in future millimeter wave (mmWave) communication system. However, in the mobile communication, the misaligned beam issue is one of the serious problems to be solved. The beam aligns between transmitter and receiver does not match will cause signal-to-interference-plus-noise ratio (SINR) degradation which will directly influence the users' experience. In this paper, we present an adaptive antenna pattern management scheme based on distributed reinforcement learning (RL) to solve misaligned beam management problem. More specifically, using distributed RL algorithm, the agent learns over time the best combination of antenna patterns among base stations (BSs) to maximize the overall SINR.

Machine learning method is efficient and effective in detecting DDoS attacks, but it all begins from identifying and selecting their important features. This paper presents an implementation of feature selection for DDoS detection based on Random Forest method. In our implementation, we use a LOIC software flood DDoS traffics to a target computer, then control the target to extract the features from the captured IP packets, and finally calculate their Gini feature importance and ranking for subsequent feature selection.

Modern network administration desires to have early detection of DDoS traffic before damages occur. Nevertheless, as Internet traffic grows over years, it becomes more challenging to detect DDoS traffic in an efficient and effective manner. The survey of existing literature shows that random forest classifier, when applied to DDoS detection, yields great performance at low learning costs. This work is devoted to a case study of implementing and using random forest classifier to detect DDoS traffic generated by a well-known DDoS tool named HULK. Our result indicates that when used with a good feature selection mechanism, random forest classifier can achieve a high detection accuracy with fast training time.

This study proposes an intelligent mask detection system. During the continuous epidemic situation, it is necessary to detect whether visitors are wearing masks. When the body temperature sensor sounds, if the face detection system detects a face, it indicates that the visitor is not wearing a mask as required; otherwise, it means that the face is wearing a mask. After many live simulation tests by the students of our laboratory, it was found that this research can indeed achieve the expected goal with a 98.6 % recognition rate and effectively remind visitors to wear their masks.

The design of an asymmetric T/R switch for the upper Ka band is presented. In particular, how to compensate for the effect of the parasitics to achieve proper matching in the millimeter-wave regime is described. To be specific, a shunt short stub is added at the antenna port to resonate with the unwanted parasitic shunt capacitance of the TX path to improve the return losses. The proposed asymmetric T/R switch is implemented in WIN 0.15-um GaAs pHEMT technology. The chip area is 1x1 mm2. Measurement results show that, from 35 to 40 GHz, the insertion loss is less than 0.9 dB and the TX-to-RX isolation is greater than 15.0 dB in the TX mode.

In this work, a 28-GHz millimeter-wave identification (MMID) frontend implemented in 90-nm CMOS for wireless sensors is presented. The MMID frontend consists of a rectifier and a modulator. The feature of the proposed MMID frontend is that the functions of modulation and impedance matching are combined in one circuit. In dormant mode, the measured input return loss (IRL) of the MMID frontend is greater than 10 dB from 26.5 to 34.1 GHz. At 28 GHz, the measured IRL is 1.8 dB and 13.3 dB for the modulated transistors being on and off, respectively, corresponding to an on-off ratio of 11.5 dB for the backscattered power. With -2-dBm input power at 28 GHz, the measured DC output voltage reaches 1.2 V, which is high enough for 90-nm CMOS circuits to function.

In recent years, the use of wireless signals for indoor localization and activity recognition has received considerable attention due to its wide range of applications, such as home automation, healthcare, and security. In this paper, we present a novel method that combines received signal strength (RSS) and channel state information (CSI) to perform accurate indoor localization and activity recognition. Our proposed method exploits the unique characteristics of RSS and CSI to achieve robust results, enabling accurate localization and recognition of human activities in indoor environments. To further enhance the performance of our approach, we incorporate radio tomographic imaging (RTI) and support vector machine (SVM) techniques into our framework. We conduct a series of experiments in an indoor environment to evaluate the effectiveness of our proposed method, and the obtained results demonstrate its high accuracy and reliability. The proposed system has great potential to facilitate the development of future human-centric Metaverse applications.

Air pollution has always been an important issue of concern, especially the emissions from transportation. Because the main primary sources of pollutant emissions are PM2.5, NOx and CO, most of them come from "vehicles". According to statistics, the total number of registered motor vehicles in Taiwan is as high as 22.844 million, which means that each citizen owns one motorcycle on average. Therefore, encouraging the use of electric vehicles, or even sharing them, may help to improve the pollution and congestion caused by urban traffic.

The carbon reduction of transportation is very important. This work determines the key factors influencing intelligent transport tools for future design. The experts' data collection was 15 (sea transportation) and 12 (air transportation). The two-dimensional mean value matrix (TDMVM) tool is used to analyse the importance, energy savings and carbon reduction. The results showed that sea and air intelligent transportation could efficiently solve energy savings and carbon emissions.

Nowadays, the global climate has caused a food crisis due to drastic changes. Although marine resources are the second-largest food source for human beings, they also face problems such as overfishing and sea pollution. Therefore, global aquaculture fisheries are increasingly in demand in the international market, and now with the continuous advancement of breeding technology, the increase in production and high-quality aquatic products. This article explores a production and marketing platform model based on the AIoT and blockchain system for intelligent aquaculture. As a result, this article analyzes the comprehensive technical support required for introducing blockchain technology into AIoT technology and assists the architecture of this article in verifying its feasibility.

Outlying islands, in Penghu County, for example, have several problems in public transportation. For local residents, the county residents enjoy free rides, which makes it difficult to reduce the loss of public transport in the urban area. The public also generally prefers private transport, etc.; In addition, both on and off-island transportation are faced with the problems of unsynchronized information on various platforms, difficulty for people to read the information, and little online information sharing on the use of buses on the island. Therefore, it is important to promote public transportation in the outlying islands to improve the seamlessness of the urban bus system, or even to meet the needs of the public under the premise of low carbon and environmental protection, so as to ultimately achieve a higher willingness to ride the bus.

Understanding speech from a silent video is possible through a lip-reading technique. Previous works in developing lip-reading models using deep learning mainly focused on improving model accuracy and increasing the model complexity. However, for limited resource devices, computational complexity is also an essential factor to be considered for evaluating the lip-reading model in addition to the model's classification accuracy, with some acceptable trade-offs between them. Our work mainly focused on developing a low-complexity and lightweight model for the lip-reading task. This paper evaluates two lightweight lip-reading models using computationally inexpensive feature extractors and classifiers. We trained and evaluated our lightweight architectures on the LRW dataset, a popular public large-scale lip-reading dataset. The experiment result suggests that with much lower computational complexity and parameter size, our lightweight architecture is comparable with model references for recognizing word-level English language lip-reading tasks.

The goal of the Multi-Target Multi-Camera (MTMC) pedestrian tracking task is to simultaneously track multiple target individuals using multiple cameras. Current methods mostly use exponential moving averages to store features and perform multi-camera person matching using these features. Apparently, it will cause the issue of poor long-term feature storage, in this paper, we propose a new method to address the issue, which often leads to ID-switching when individuals change clothes or when lighting conditions change significantly, and also improve the ID-switching problem that occurs during single-camera tracking. To evaluate our method, we created our owns dataset. The dataset was included approximately 40000 frames from 1080p, 30fps videos, whitch were recorded by these cameras. Experimental results show that our method outperforms existing methods in both single-camera and multi-camera tracking, with single-camera tracking improved by 7.06% and multi-camera tracking improved by 12.51%.

This paper proposes a novel region-of-interest (ROI) detection method that utilizes a graph convolutional network (GCN) and information from H.266/VVC video. The GCN takes various features, including the geometric features, coding mode, motion information, and quantization parameter of the coding unit (CU), as input and produces an output that indicates whether the CU belongs to the ROI. By combining the GCN model with the encoded video, the proposed approach significantly reduces the computation time, with an average time saving of 60.88% compared to that of YOLOv7. This approach provides a fast ROI detection method and facilitates post-processing tasks such as the quality enhancement of ROIs for H.266/VVC video.

Two proposed divide-by-4/5 dual-modulus prescaler (DMP) based on true single-phase clock (TSPC) and extended TSPC (E-TSPC) structure for the 5G applications is designed by TSMC 180nm CMOS technology. The TSPC DMP achieves a maximum operating frequency of 4.9GHz with 4.4GHz bandwidth and consumes 0.72mW at 1.8V. The maximum operating frequency of E-TSPC DMP is up to 7.6GHz with 5.3GHz bandwidth and consumes 1.29mW.

This paper presents the synthesis result of extreme learning machine (ELM), a machine learning technique, to detect age-related macular degeneration (AMD), an eye disease prevalent in elderly people. The model is trained with optical coherence tomography (OCT) images and a desirable clock period, area, and power is obtained. This is the first of its kind synthesis result of ELM implementation for AMD detection on OCT images.

Deep Neural Networks (DNN) find its prominent presence in the AI world. The properties of its algorithms are very well exploited to make its computation power faster and more efficient. One such adaptation requires reduction in the bit width of the operations for DNN. This paper deals about a DNN architecture design with reconfigurable bitwidth accelerator without affecting the accuracy. This proposed architecture is a modified version of Bit Fusion architecture dealing with dynamic bit-level decomposition for accelerating complex DNN computations. The design make computations with less power consumption, and it is more suitable for the trade-offs among the latency, power, and temperature. The micro architecture design of the modified Bit Fusion using RTL simulations is carried out to evaluate the functionality.

Image dehazing is a post processing technique to remove the haze in an input hazy image. Since the haze removal poses a challenging problem. In this paper, a stripe-based airlight estimation (SBAE) algorithm is proposed for low complexity image dehazing. At first, SBAE transforms RGB of a hazy image to its HSV. The depth information of each pixel can be estimated from HSV. Every maximum depth of each 5x5 patch in four stripes is collected. The airlight is estimated as an average of their corresponding pixel intensities in a hazy image. Experiment results show that the computation saving ratio (CSR) is 99% with a minor error rate of only 1.55%. SBAE greatly reduces the computation burdens while keeping the accuracy of airlight estimation. SBAE provides the algorithm for low complexity image dehazing .

For high speed single-end parallel link system, its data rate is limited to simultaneous switching noise (SSN) by interface circuits. To reduce SSN, this paper proposes an orderly-drive parallel interface circuit. The parallel interface circuit transmits data at different phases of data period instead of simultaneously transmit data at the same phase. The current spike at power supply is averaged to all the data period thereby highly reducing SSN. To verify the proposed SSN suppression scheme, a single-ended driver and an 8-channels parallel link system are realized in TSMC 90nm CMOS technology. As compare to conventional parallel link system, the proposed design reveals a 75% of average SSN reduction and at all different data rates.

In this paper, we propose a node-based epidemic model for a countermeasure strategy against future malware evolution. The proposed epidemic model constructs a countermeasure group to suppress the infection spreading of malware. The countermeasure group aims to discover and repair software vulnerabilities through distributed machine learning using the computing resources of hosts belonging to the countermeasure group. The proposed epidemic model gives infection dynamics of malware under the situation where the countermeasure group works, considering network structures. We discuss the infection trend under this situation through numerical calculation.

We propose a field monitoring system whereby drones search for and collect information on a crop field. In the proposed system, field monitoring starts with most of the drones searching for objects. Over time, the drones gradually change their role to collect information on the found objects. As there is no communication infrastructure in the field, communication is achieved through a delay-tolerant networking technology called store-carry-forward routing. In store-carry-forward routing, drones store received messages in a buffer. When drones encounter one another, if a drone has a stored message, it forwards it to the encountered drone. Through simulation experiments, we show that the proposed system searches for objects and collects information efficiently.

In unmanned aerial vehicle~(UAV) relay networks, the throughput performance is influenced by wind perturbations. We consider a two-hop full-duplex UAV relay network using maximum ratio combining multiple-input multiple-output (MIMO) data transmission and propose an adaptive antenna selection scheme to improve the robustness against these perturbations. As the links with shorter distances are more influenced by the perturbations, we adjust beamwidths of the source-relay and relay-destination links by selecting antenna elements of the nodes according to the distances between neighboring nodes.

Wave source localization or emitter localization aims at estimating positions of unknown transmitters of radio signals. A maximum likelihood wave source localization scheme (MLE scheme) using unmanned aerial vehicles (UAVs) was proposed in non-line-of-sight (NLOS) environments. We propose a new wave source localization scheme by extending the MLE scheme. The proposed scheme first obtains a tentative wave source location using the MLE scheme, and then identifies whether each path between the tentative location and a measurement point is in LOS or NLOS. Finally, the tentative location is corrected with the LOS/NLOS identification. We evaluate the performance of the proposed scheme with simulation experiments and show that the average estimation error of the proposed scheme is improved by more than 30% compared to the MLE scheme.

As online services have become increasingly popular and convenient in recent years, the number of phishing scams has increased. To prevent phishing scams, deep learning-based phishing detection methods have been proposed and shown to be effective. Recently however, backdoor attacks, in which deep learning intentionally induces false positives by falsifying training data, have appeared. In this paper, we propose a countermeasure against backdoor attacks that is suitable for deep learning-based phishing site detectors. In our countermeasure, the values of the layer before the output layers of the neural network are used as feature vectors, and the values of these feature vectors are monitored. The number of trigger data is much smaller than the number of normal data and can be detected as outliers. To detect outliers, our countermeasure uses HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise), which is a density-based clustering method. The feature vectors are clustered using HDBSCAN, and data suspected to be triggers are classified as noise. Through experiments with poisoned data, we show the effectiveness of our countermeasure against backdoor attacks.

In this study, a directional beam antenna is proposed to relay transmissions between flighting drones. The proposed antenna maintains the communication link by varying the beam width when the drones are tilted by wind or other factors. Its structure is based on a small patch antenna, and the beam width and gain are changed by overlapping no-fed elements. In this study, the beam width and gain are analyzed by changing the number of no-feed elements by computer simulation, and the optimum application of the beam width to the tilt of the drone is described. Moreover, a switching circuit is proposed for dynamic beam width control.

Previously, we developed pear pollination technology using drones. Drones fly in pear fields, detect flowers for pollination, and then spray pollen with high precision. To achieve these tasks, drone flight accuracy by positioning is required. Conventional outdoor drones fly using a global navigation satellite system (GNSS), which receives signals from satellites and estimates their position. However, conventional positioning methods have an error of several meters to tens of meters from the indicated route. When flying drones in a pear field, as they need to fly through narrow spaces between trees with an accuracy of tens of centimeters or less, flight control is critical. In this study, in addition to GNSS position information, real time kinematic (RTK), which performs positioning correction at the ground station, was used to achieve positioning within a few centimeters in the pear field. We examined the principle of positioning and developed a positioning system using RTK and the flight experiment of a drone.

This paper proposes a deep learning-based monitoring system for continuous ambulatory peritoneal dialysis called iCAPD. iCAPD comprises a load cell module, a heating and thermal preservation module, a fluid turbidity test chart, a camera module, an Edge-AI computing module, and a display module. iCAPD can heat and preserve the temperature of the dialysate bag at 37℃ within 20 minutes. Furthermore, it can detect the dialysate infusion volume by sensing the weight change of the dialysate bag. Moreover, iCAPD uses deep learning-based technology to recognize the turbidity of drainage fluid from the abdominal cavity. The experimental results show that the turbidity recognition performance of iCAPD can achieve the purpose of assisting the assessing the severity of associated peritoneal symptoms (e.g., peritonitis).

Bioelectrical Impedance Sensor Analysis (BISA) measures body composition by passing a small amount of electrical current through a subject's limbs. The water in the body increases, the electrical resistance decreases, and the measured body fat decreases. The Bioelectrical Impedance Sensor Analysis curve can observe the number of corresponding heartbeats. The vertical axis on the distribution graph represents the impedance value of the body. This value is between 33.5~35.5Ω measured on a normal person.

Acupuncture medicine has been proven effective by many studies and is popular in the market. However, in busy medical work, patients often take the needle home because of missed needle removal. In order to solve such incidents, this research introduces Image recognition and IoT sensing technology and design & implement a device that can automatically calculate and compare the number of needles used and the number of recycled needles. The system will send an error alert when the two values are not equal. After three days of actual verification and testing, users (doctors and nurses) are satisfied, and the rate of missed needles during the experiment is 0%, which proves that this product is indeed feasible and effective

Auscultation is one of the fetal monitoring methods. Since the acoustic signal acquired from the pregnant woman's abdomen contains a lot of noise, it is necessary to perform noise reduction. Software implementation of the noise reduction is not effective for real-time monitoring because the processing time is longer than the recording time of the heart sound. In this work, fractal dimension analysis, which accounts for the majority of the heartbeat detection process, was implemented as hardware on an FPGA to realize constant real-time monitoring.

This paper proposes a fitness assistance system, called A2Fitness, which integrates with an AI subsystem and an AR subsystem. By using A2Fitness, the elderly can wear an AR smart glasses at anytime and anywhere to instantly see the images and correctness of their physical fitness movements. Therefore, A2Fitness is quite suitable as a physical fitness tool for the elderly to effectively promote their own health.

The E13B font recognition method achieved fast and accurate recognition but had certain drawbacks. One of which was the lack of confidence information for character prediction. When the decision tree encounters an issue, a closer answer cannot be recovered, leading to the failure of the entire check recognition. Although this event is highly improbable, it could still occur. Therefore, we needed to continue optimizing the recognition process, and this study proposed a novel framework to address this issue. This study incorporated a model trained by a lightweight neural network to predict the result when the decision tree encounters an anomaly. This resolves the previous flaw of failing to provide a prediction when the tree encounters a problem. The trained model can also meet the computational requirements of embedded devices and complete the recognition of a character within 1.8 ms.

When muscles contract or relax, they are usually clinically examined using an electromyography (EMG) machine. Currently, diagnosis requires professional medical personnel to use invasive needle electrodes to examine the muscles. During the examination, the physician will request the patient to perform relaxation and exertion movements, and the EMG machine is used to diagnose the patient's muscle condition in real-time. In addition, the inability of EMG machines to store real-time data makes it difficult to conduct subsequent tracking and analysis. Therefore, this paper designs a muscle condition measurement system that replaces electrode needles with adhesive patches, which is a non-invasive way to perform measurements. The advantages of this system allow patients to be more relaxed during the measurement of muscle signals and avoid misjudgments due to nervousness during measurement. Moreover, non-invasive devices can reduce the risk of infectious diseases. Additionally, this system provides signal recording and storage functions to facilitate subsequent data analysis. In the experimental part of this paper, the K-Means algorithm was adopted to initially find the baseline of muscles among the subjects for further research.

The field of wearable electronic devices is rapidly expanding, with a wide range of applications such as personalized health monitoring, soft robots, and human motion detection. To create textile-based electronic devices, it is essential to study electronic devices that are based on fibers. This study presents a simple and controllable process for creating a new stretchable fiber strain sensor using silver and polyurethane. The newly developed fiber strain sensors have several benefits, including a low initial electrical resistance of 10 Ω/cm, a wide range of strain sensing, high sensitivity, and high stability against repeated stretching and releasing in only one-process adsorption and reduction. The adsorption of continuous silver particles on the fiber surface is demonstrated using electric resistance variances, micro-images, and thermogravimetric analysis. These results provide valuable insights into the development of wearable applications for the next generation.

Pedestrian trajectory prediction is one of the important research topics for autonomous vehicles. Many factors such as Pedestrian's actions and waypoints in the scene can affect the pedestrian's routes. Currently, most researches focus on predicting trajectories in Bird's Eye View (BEV), which limits the application in monocular cameras. In this paper, we propose to combine BEV trajectory prediction with deterministic tracking method and create a new egocentric algorithm for predicting pedestrian trajectories. Experiment results showed that our method can accurately predict pedestrian trajectories in egocentric scenes.

People used to think that the 21st century was the most civilized century and there would be no large-scale wars, yet the Ukrainian-Russian War reminded us how close wars are and how fragile life is. To save lives and reduce casualties, we developed an object detector that can identify 4 different types of vehicles: tank, missile vehicle, combat vehicle, and civilian vehicle. Our dataset includes 4,000 images with 1,000 per category. Experiment results show that our detector can achieve high accuracy and can be applied in early warning system for citizens.

The development of automatic driving technology and Advanced Driver Assistance Systems (ADAS) has become more and more mature, and has led to the development of 3D object detection technologies and discussions. Most of the researches are focus on the applications of small cars, but not common in large vehicles, such as buses and trucks, although such applications are more important for them. Therefore, this study mainly uses the bus data to verify the feasibility of the existed model. Based on our experimental results, we found that the model trained by KITTI dataset can predict acceptable result refers to the input images from the bus viewpoint. As a result, we can expect that the ADAS for large vehicles can be implemented based on the sample model structure, and its accuracy is able to be improved by fine-tuning based on more large vehicle viewpoint data.

Quantization is one of the optimization methods for developing deep learning models for edge devices. Through converting the floating-point into 8bit integer or even lower bitwidth, the model's storage size can be reduced. As the rounding error exists during the quantization process, the model performance decreases. As a result, a method that can recover model performance is needed. In this research, a compensation method for improving the performance of quantized deep learning models is proposed, which make the quantized model can achieve equal or even better performance compared to the original floating-point model.

Obtaining high-resolution depth map from radar and lidar for training depth estimation network is a common approach in recent years, although the expensive price and high computational complexity cause it hard to be apply in real-world applications. In this research, we propose a novel concept that only requires low resolution depth map for training, but able to reach similar performance comparing to high-resolution depth map-based approaches. Refers to the experimental results, which evaluated on KITTI dataset, our proposed concept decreases 7% and 2% on the mAP of car and pedestrian comparing to the original approach in CaDDN, although it achieves similar performance on visualization result.

In the field of autonomous driving, many models based on deep learning methods have been constructed to solve computer vision tasks related to this domain, such as object detection, semantic segmentation, and depth estimation. Each model can provide different information about the surrounding environment of a vehicle to assist in driving. However, for applying in the real world, more detailed information about the surrounding environment is essentially required. In this study, we propose a model based on the concept of multi-task learning. This model is an encoder-decoder architecture mainly consisting of an encoder using the hard-parameter sharing technique and three decoders for individual tasks. Therefore, this model can handle object detection, semantic segmentation, and depth estimation at the same time. Our proposed multi-task model has been verified to perform well on the public dataset Cityscapes and has higher generalizability than other models.

Federated learning is a new machine learning paradigm that trains an algorithm across multiple edge devices without sharing data. However, traditional deployment of federated learning encounters difficulties for efficient environment setup, resource allocation and maintenance. To resolve this, in this study rebuilt the federated learning algorithm into a containerized software structure based on Kubernetes and Docker. This framework significantly reduced the construction time and improves the efficiency of resource utilization. Moreover, we proposed removing gradient computation of some fully connected layers in the backward pass to shorten the training time. Compared with the classic federated learning algorithm FedAvg initiated by Google, the proposed lightweight federated learning algorithm could reduce the overall training time by 50% or more for large batch size and heavier local training without sacrificing accuracy.

Machine unlearning is an act of eliminating a user's data from a machine learning model. Machine unlearning, however, causes a new problem of privacy leakage because an attacker may infer the deleted data by comparing the difference between the original model and the unlearned model. In this study, we develop a new privacy-preserving mechanism for machine unlearning that is universal to all machine learning models. The shadow model, typically used to train attackers in a black-box scenario, is employed as a defense means to protect the deleted data in a white-box scenario. We produce a machine learning model by a linear combination of the unlearned model and shadow models. This approach confuses the attacker while retains the accuracy of the original model. The proposed system provides a flexible trade-off between the accuracies of the adversarial attack and the protected model. Experimental results show that the proposed method could reduce the accuracy of the attack by more than 13%.

In recent years, mobile music games have become a popular form of entertainment, but gameplay is limited to the operation on the screen. The use of three-axis gyroscope sensing technology in mobile games is very common, but not in music games. This study proposes a motion-sensing music rhythm game, using three-axis gyroscope sensing to increase the judgement of operation methods and thereby increase the game play. Players must hold the mobile phone with both hands, and as specific beat notes appear in the game interface from the center to the circle edge, they are required to rotate the body of the mobile phone to correspond to the correct area of judgement. The judgement area is responsible for detecting the type of note and determining the direction in which the user should rotate. Through this game, it is expected to bring new gaming experiences to players.

The implementation of Gated Recurrent Neural Networks (GRU) to generate background music (BGM) combines deep learning technology with music that is used for the visual content of a commercial or educational. Indeed, this BGM is necessary to enhance the intended message expressed to the other audience. This work aimed to provide the model network of GRU to generate multi-label genres of music by using the open source of GTZAN to evaluate the new BGM. Our GRU networks can solve the vanishing gradient problem by utilizing both the reset gate and the update gate on the network. In the results, we achieved a new BGM that synchronized with the human mood which made more variety of sounds.

This paper is about predicting the winner of League of Legends (LoL) games through machine learning. The dataset used features the game statistics of the first 10 min of diamond ranked games. This paper aims to determine the top 3 most accurate models from ANN, LR, SVM, kNN, LDA, and NB. After training and testing the models with the dataset, LR, LDA, and SVM were found to be the most accurate with all 3 being around 72% with NB close behind.

Violent incidents have happened regularly since ancient times and usually cannot be rescued immediately, hence causing tragedies. Therefore, we expect to study a violence detection and alarm system based on human body pose recognition of images captured from public CCTVs. The system uses OpenPose for real-time skeleton recognition to detect abnormal poses or abnormal speeds, based on the angles or speeds of movement between the various joints of the limbs, to determine whether it is violent behavior or not. If a possible violent event is detected, the system will trigger an alarm and notify the police or security guard to deter the attacker. Victims will be able to receive assistance as soon as possible through this recognition system, even if they have obstacles for calling help themselves. Experimental results show that the target violence behaviors can be detected successfully.

Software testing has become a key factor in software development processes in the last decades. Ideally, it should check programming errors, assure conformance, and prevent weaknesses that attacks can exploit. However, at the same time, the ability to increase the coverage of testing approaches regarding expected features or source code also turned out to be quite a challenge. The present work considers tests applied to evaluate interactive middleware modules for digital TV receivers. To enhance their resulting quality assessment, we have developed a methodology that verifies the current test coverage provided by a test suite and performs an exploratory study to expand it. Finally, we have conducted experiments considering test suites for commercial versions of the middleware Ginga-D used in Brazil. As a result, new error conditions and weaknesses were identified, which can effectively improve middleware modules.

A modified direct binary search (DBS) algorithm which minimizes the flickers on the stationary regions has been proposed for electronic paper (ePaper) in our previous work. Acceptable video quality can be achieved in one iteration, but the high time complexity of DBS algorithm still leads to the difficulty of implementing it on an ePaper system. In general, the ePaper is equipped with a low-power microcontroller. It is infeasible to perform real-time video halftoning on such a hardware platform. This paper presents a novel hardware architecture based on a parallel video halftoning algorithm, for electronic papers. This halftoning engine performs DBS for four pixels in a macro block or in different macro blocks parallelly.

Electronic paper (ePaper) is typically equipped with a thin-film-transistor (TFT) backplane as a liquid crystal display (LCD), but their driving methods are different. The polarity of the driving voltage on LCD is interchanged frame by frame or pixel by pixel for maintaining the reliability of the liquid crystal. However, the pixel state of the ePaper depends on the polarity of the supply voltage, and the gray tone of the pixel is achieved by pulse width modulation. This paper aims at presenting a new driving system by using the commercial TFT backplane with the drive chips for LCD to integrate the entire ePaper system. Thus the ePaper system could share the same TFT backplane as well as the driver chips for LCD. It significantly reduces the production process and cost of making ePaper modules.

Due to the bistable property of the electronic paper (ePaper), its driving sequence starts from erasing the previous image and then the pixels are driven to the target state according to the next image. The timing controller (TCON) of ePaper typically contains two image buffers to store the previous and the next image data. This paper presents a new TCON for ePaper which contains only an image buffer. Unlike the traditional ePaper TCON, which receives the next image data and stores it in an image buffer inside the TCON, the new TCON continuously receives the data of the next image when the image update period and the image buffer for the next image is eliminated. This new TCON leverages the property of the video decoder which continuously sends out the image data in video rate. The TCON compares the incoming image data with the previous image data to sequentially send out the driving signals. With this new architecture, the required memory buffer could be significantly reduced.

In this paper, a denoising method of bandlimited graph signal without using graph Fourier transform (GFT) basis is presented. First, the bandlimitedness prior in spectral domain is used to develop a denoising method which needs to know the orthonormal basis of GFT for computing the denoising matrix. Second, the denoising matrix of bandlimitedness-based method is approximated by the transfer matrix of graph filter whose optimal coefficients are determined by minimizing the Frobenius norm of error matrix. Third, the approximate denoising graph filter is implemented in vertex domain, so the proposed method does not need to know the orthonormal basis of GFT. Finally, the temperature data at USA is used to validate that proposed GFT free method can provide almost the same performance as the conventional method.

In this paper, a decentralized implementation of heat kernel smoothing (HKS) is presented by using a rational graph filtering method. First, the basics of graph signal processing (GSP) are briefly described. Then, the centralized implementation of HKS using matrix exponential is studied. Next, the Taylor series expansion of exponential function and Prony method are applied to determine the filter coefficients of a rational graph filter. Because only linear equations need to be solved, the proposed design method is easy to be used. Finally, the performance of the proposed HKS method is evaluated by using the irregular data captured from the sensor network.

This paper presents an Infrared (IR) image interface circuit on FPGA. By using the Non-Uniform Correction (NUC), the process variation in the IR sensor could be cancelled. With the Histogram Equalization (HE), the distribution of the pixels is extended to the range of the full color ramp such that the contrast of the image could be improved. Besides, the ping-pong processing with 2 SRAMs is proposed for the real-time operation. The experiment is executed on DE2-115 FPGA board with the gate count of 12758. Based on the experiment results, the quality of the image is improved.

This study obtained a wide range of images of the lighting environment in the illumination control space, and used illuminance correction technology to achieve the actual illuminance information provided by the lighting system, forming a closed-loop lighting control system to drive each luminaire and corresponding EN 12464-1 reading light standard. Under the condition that all luminaires are driven at 80%, the average illuminance of the nine desktop center points is 1011 lux, the minimum illuminance is 920 lux, and the uniformity is 0.9099. If intelligent control is achieved through digital controllers and illuminance, and part of the luminaire output power is reduced according to the corresponding illuminance relationship, the overall driving power of all luminaires can be reduced to 72.5%, which is 298W, and the intelligent energy-saving amplitude is 6.87%. The average illuminance of the nine desktop center points is 979 lux, the minimum illuminance is 943 lux, and the calculated uniformity is 0.9632, and the uniformity enhancement amplitude is 0.0533 (5.33%), achieving appropriate lighting quality and energy savings through intelligent control in the field.

This study improves the accuracy of image recognition technology to over 90% by optimizing the YOLOv3 training model and verifying the results through recorded images from the experimental field - Baoshan Library. The data provides highly reliable personnel movement information. By analyzing the data from the camera system and grayscale images with personnel recognition technology, and combining it with the control system interface developed for the experimental field, the continuous image lighting system control verification is achieved. The system can detect and turn on the lights within 1.5 to 1.8 seconds when a person appears within the detection range of the system. When the person disappears from the detection system, the system can turn off the lights within 1.4 to 1.9 seconds.

The article proposes a non-contact method for negative emotion recognition. Our method divides emotions into positive and negative categories and uses facial expression features and heart rate as indicators to identify negative emotions in indoor spaces. This method uses a combination of cameras and millimeter-wave radar as sensing devices, which enhances its adaptability to the environment and improves emotion recognition accuracy. This method can be applied to office and commercial spaces to assess employees' quality of work and psychological well-being from multiple angles and can be used for early analysis of depressive tendencies and stress detection.

Depth camera is a field-of-view (FoV) based distance sensor and has been widely used in commercial entertainments such as augmented reality, industrial field for object 3D modeling, as well as intelligence vehicle obstacle avoidance. No doubt, the depth camera measurement accuracy definitely affects the associated application performance and therefore the noise behavior should be properly modeled. The measurement noise of the depth cameras depends on various factors, which can be difficult to model in practice. In this short note, three different depth noise models are presented based on the pin-hole model of the camera. The goal is to match the practical depth camera noise distributions as close as possible and to provide a simulation sketch for further noise analysis and possible improvement of the depth measurements.

A depth camera is a type of camera that uses infrared or structure light to create a 3D depth image of an object being captured. Unlike a traditional camera, which can only capture a 2D image, a depth camera is able to determine the distance between objects within the field-of-view (FoV). As a result, depth cameras have become the popular consumer devices and are often used in applications when the accurate measurement of distance and depth information is important. However, measurement noise is not avoidable and it definitely contaminates depth information. Therefore, this short note presents a simple depth point cloud sharpening scheme to enhance the measurement accuracy, which can give significant contribution for 3D scanning and 3D object reconstruction uses.

Object detection is an important task in computer vision. In earlier studies, object detection mostly works on single-modal. But recently, in order to improve detection efficiency, researchers start adding more modals for improving the robustness of detection result. In this paper, we assume that all input modals which can convert to image style are able to complete object detection task through a unified model. We choose YOLOv7 as our baseline model and use RGB and infrared images from the public dataset Camel to test our hypothesis. Based on our experiment results, training RGB and IR images together obtains better performance than training individually, which shows that all image style modalities are possibly handled by a unified model and even achieve better results.

In this study, we integrated the Polynomial Regression module with a Variational Autoencoder (PolyReg-VAE) model to foster the inference of transfer curves before further deriving the electrical properties of a-IGZO TFT. This approach is able to predict the electrical characteristics of a-IGZO-TFT with different process parameters in a short time, reducing the cost and the time of developing new technology.

A performance comparison was conducted between T-type FinFETs (T-shaped Fin Field-effect-transistors) and I-type FinFETs (I-shaped Fin Field-effect-transistors) under identical conditions. After optimizing the structure of the T-type FinFET, it was found to have a higher drive current and better short channel effect than the I-type FinFET in the same scenario.

In this paper, the fabrication of highly sensitivity MSM structure humidity sensor were based on simple one-dimensional zinc oxide nanorod (1-D ZnO NRs) were synthesized with a low cost aqueous solution hydrothermal method for grown on glass substrate. The MSM structure 1-D ZnO NRs humidity sensor devices was labeled for MSM@1-D ZnO. The physical property analysis displayed the surface morphology of the MSM@1-D ZnO. The electrical characteristics analysis results showed that MSM@1-D ZnO demonstrate highly response and sensitivity under different Relative Humidity (RH%) conditions. This paper verify that 1-D ZnO NRs was a potential option for humidity sensor research

Virtualization technologies are still growing bigger and faster. Despite the greatness of its advancement, the costume industry is still very accessible when it comes to real trials. Off-the-shelf stuff are inadequate details for the desired individual to assess its in-depth utility for each garment trying on for a second, including custom stuff are much harder to try out right away. To this end, 2D image-based 3D reconstruction inclusive of touchable-virtualized space is accessible easier to stuff details for mans' decision making in purchasing. We establish the overall end-to-end pipeline from reconstruction until visualization for one instance to be triable on its stuff for a moment. As an expectation, our proposed approach can bring objects into the experimental area and use them immediately without any obstacle.

Energy saving and carbon reduction have become the most concerned issues of enterprises in recent years. 5G base stations are also expected to be equipped with energy-saving technologies for 5G ESG benefits and a green future. To this end, we develop an energy-saving control mechanism on the management platform and control the setting parameters of the system. The designed mechanism is used to avoid the ping-pong effect of the base station switching caused by the capacity differences between the two cells. When the traffic load is low for a while, the cell of 3.5GHz will switch to sleep mode and only provide users with the cell of 2100MHz for transmission to reduce the power consumption of the 5G base station. Conversely, when the traffic demand continues growing, the cell of 3.5GHz will wake up to support the capacity of the base station. The measurement results show that the base station can significantly reduce energy consumption.

For the coming of IoT era with huge number of smart IoT devices, simultaneous localization of multiple users is becoming an important research topic for many application of indoor scenarios such as health care, manufacturing automation and robotics. This paper will propose a simultaneous localization method for Wireless Environments using leaky coaxial cables (LCX) using a probabilistic neural network (PNN). The PNN learns the relationship between the locations information of multiple users and their time of arrivals (TOAs) at both sides of LCX and then predict the new location results with input values of TOAs. The experiment results show that the proposal has a promising performance in simultaneous localization for multiple users.

In this paper, a speedup method is presented to accelerate the convergence of iterative computation of pagerank centrality which can be used to identify the important or central nodes of complex networks based on the linkage structures of their graphical representation. Because pagerank computing is still a challenging task for a large graph, this paper aims to study the fast iterative methods for computing original pagerank and personalized pagerank in a distributed manner. As a result, the convergence speeds of iterative computing methods can be improved at the cost of increasing of few computation load and memory requirement at each node. A numerical example of node centrality of karate-club social network is used to demonstrate the effectiveness of the proposed method.

In this paper, the influential station identification of Taipei metro network using Laplacian centrality is presented. First, the Taipei metro system network is represented by an un-weighted undirected graph without containing self-loop. Then, the energy of graph Laplacian matrix is used to define the Laplacian centrality measure and some useful properties are studied. Next, the relation between vertex degree sequence and Laplacian centrality is established such that the Laplacian centrality can be easily computed by using degree sequence. Finally, the experimental results demonstrate that the Laplacian centrality can provide reasonable influential stations of Taipei metro network and require less computational complexity than classical centrality methods.

This study introduces RandLASAMP-Net as an advanced version of RandLA-Net, designed to facilitate efficient per-point semantics inference for large-scale 3D point clouds. Existing algorithms encounter difficulties in processing large point clouds due to costly sampling or pre/post-processing necessities. To enhance the feature aggregation module of RandLA-Net, which relies on a rudimentary attention mechanism, we propose the integration of a self-attention mechanism. This mechanism enables the model to focus on different regions of the point cloud, enhancing its capacity to extract significant features from small and unstructured point sets. Furthermore, we introduce a multi-scale pyramid module that adapts to the UNet-shaped architecture for the point cloud to better utilize encoding features. Our experimental results demonstrate that these enhancements improve the accuracy of semantic segmentation while maintaining computational and memory efficiency for large-scale point clouds. Our goal is to develop a reliable model that can efficiently handle large-scale point clouds with high accuracy.

H.266/Versatile Video Coding (VVC), the latest video coding standard, provides higher coding efficiency than its predecessor, H.265/High Efficiency Video Coding (HEVC). The quadtree and nested multi-type tree (QT/MTT) partition structure in H.266/VVC greatly increases the encoding time. This paper proposes a fast QT/MTT early termination algorithm to bypass the coding unit (CU) depths that occur less frequently or skip the ternary tree (TT) partition based on the statistics of conditional probabilities. The fast MTT decision algorithm is also proposed based on Pearson correlation analysis, which utilizes the rate-distortion costs of the spatial and temporal coding tree units (CTUs). The experimental results show that the proposed method can effectively achieve encoding time reduction and maintain the bit rate.

Due to the mobility of drones and their limited power supply, the network topology frequently changes, making routing more complex. To address this issue, an energy-aware routing protocol is proposed to extend the lifespan of the unmanned aerial vehicle (UAV). Because vehicles have less power constraints, the proposed algorithm prioritizes vehicles for packet transmission when multiple paths exist, and both UAV networks and vehicular communication networks can aid in packet forwarding to improve the lifespan of UAV networks.

Fifth generation (5G) New Radio (NR), was developed to offer more flexibility to meet new service requirements. Meanwhile, machine learning (ML) has proven successful in a variety of tasks, such as natural language processing, computer vision, and pattern recognition, in particular, which is proven to have a performance that scales with the total amount of available data. In NR, the capability to locate users is still one of the critical obstacles when mobile operator is planning and optimizing the cellular networks. Developing the technique to distinguish indoor from outdoor users' traffic pattern can achieve higher efficiency in terms of resource management and which results in larger economic benefit. In this paper, we present a pattern classifier based on decision tree to solve the indoor/outdoor classification problem. More specifically, rules for classification of indoor/outdoor users are generated by repeatedly splitting the features from cellular network key performance indicators (KPIs) which utilize the measurement criteria of entropy from information theory community.

Among the causes of traffic accidents in Taiwan in 2021, "not paying attention to conditions ahead of the vehicle" ranked among the top two. It is a big challenge to assist drivers in judging road conditions in advance to reduce the occurrence of traffic accidents. This project uses the DR(eve)VE dataset to capture video data of driving conditions in different environments and applies the YOLO v4 algorithm to detect real-time road objects. The detection results are then imported into the Deep Sort algorithm to track the trajectory of the objects. Finally, the optimal safe distance between the vehicle and the detected objects is calculated. If the objects are within the safe distance, the driver will be alerted to avoid potential car accidents. By integrating the YOLO v4 and Deep Sort algorithms, this project aims to assist drivers in judging road conditions so that they can arrive home safely.

Named Entity Recognition is a fundamental task in information extraction, which locates and classifies defined named entities in unstructured text. Chinese NER is more difficult than English NER. Since there are no separators between Chinese characters, incorrectly segmented entity boundaries will cause error propagation in NER. In this study, named entity recognition is constructed and applied in the Chinese medical domain, where Chinese medical datasets are labeled in BIO format. The Chinese HealthNER Corpus contains 33,896 sentences, of which 2531 sentences are divided into the validation set and 3204 sentences are divided into the test set. This study uses PyTorch Embedding + BiLSTM + CRF, RoBERTa + BiLSTM + CRF, BERT Classifier and BERT + BiLSTM + CRF for training and compares their model performance. Finally, the BERT + BiLSTM + CRF achieves the best prediction performance with precision of 91.30%, recall of 89.46%, and F1 score of 90.53%.

Eye-tracking is a potential tool for assessing laparoscopic surgery skills, but the effects of laparoscopic training on eye movement measurements are not fully understood. To investigate this, we recruited 9 graduate students with no prior laparoscopic skills and used a repeated-measures ANOVA to analyze task performance and eye movement measures. Results showed that saccade duration, amplitude, velocity, and peak velocity increased while the number of fixations and saccades decreased significantly with training. These findings provide insights into developing an objective eye movement-based system for assessing laparoscopic surgical skills.

Advance with times, it has been a modern trend in education for parents to participate in school education affairs. Parents are no longer bystanders in education. It makes parent-teacher communication gradually become frequent and unavoidable, especially in the first stage of formal basic education (elementary school). Parent-teacher interaction will greatly affect children's growth and learning development. In this paper, it tends to build an automated interactive system by using chat robot technology to help parents and teachers communicate effectively. Parents can use this proposed system to understand their children's campus life and participate in teachers' class management, such as inquiring about children's homework, academic performance, learning status, school activities, and so on. This way for automatic parent-teacher interaction not only reduces the burden on teachers in communication, but also can solve parents' questions instantly and conveniently. So that, both parents and teachers can get closer and strengthen each other's consensus on education.

Applying the Internet of Things (IoT) is a way to link things through the Internet, thus creating work automation advantages. Therefore, program development regarding the IoT is becoming a field that attracts researchers and teachers to explore. However, low-level programming tools are often used in the learning process, which are not always user-friendly for beginners. Therefore, this study proposes a set of visual environmental analysis systems of self-directed learning courses to write instructions by dragging and dropping building blocks to reduce beginners' learning obstacles. Additionally, this system uses the concept of self-directed learning in the study design. As such, we expect it to assist students in adjusting their learning speed to make learning personalized and reduce the situation of different states during unified learning. Moreover, this system also records all data detected. It presents the charts drawn after data analysis through web pages to increase students' mastery of environmental changes at various time points. This study designed a complete experiment to explore the differences in the learning effects of this system on students.

Given modern lifestyles, where people primarily eat out, obesity and suboptimal health due to unbalanced diets have been increasing. This study aims to develop an easy-to-use food image detection system, integrated with a website and chatbot, to help people improve their dietary intake and enhance their overall health.

The visually impaired people often use guide sticks, guide dogs, or ask people to know their direction when walking outside. However, there are many guide bricks not only fail to help them, but also guide them to obstacles and wrong directions. Therefore, we apply deep learning algorithms for image segmentation and recognition of different objects, and automatically estimate the distance with low-cost, lightweight embedded devices. We also provide voice service to help visually impaired people understand the surrounding obstacles, so that they can reduce the risk of collision and find the correct and safe direction. Our approach assists the environment perception and guidance of the visually impaired people.

The electronic nose imitates human smell mechanism and has been used in many diverse fields for odor detection. We apply several gas sensors with Arduino platform and design the multivariate analysis pipeline from time series sensor data. We also combine augmented reality technology which is superimposing virtual objects of the predicted results on the captured image in real-time. Based on the experimental results on our collected odor sensor data, the performance of our approach shows 86% F1 measurement. The results show that our analysis processes including feature engineering, dimensional reduction and machine learning methods that has good discriminative ability for odor classification.

Typically, users can pay higher transaction fee to reduce confirmation time while submitting a transaction in Ethereum. However, the rising transaction fee becomes a major issue. Therefore, a fine-grained gas price suggestion tool is expected to various decentralized applications (dApps) with different Quality of Services (QoS) requirements. In this paper, we proposed a machine learning based gas price prediction approach for optimizing transaction fee in Ethereum. The proposed approach takes features of dApps into consideration. The experimental results show that the proposed approach can provide fine-grained transaction fee estimation for dApps with various QoS requirements.

Fans of live music support artists on stage with cheers and applause. Before the COVID-19 pandemic, live music performances were characterized by audience cheers, and artists could feel the fans' excitement through these cheers. However, with the spread of COVID-19, it has become difficult for fans to cheer, and such sharing has become challenging. In this paper, we propose a sensor worn on the wrist to record fans' excitement using an accelerometer. Our method allows artists and fans to share excitement, momentum, and status during online live performances during the COVID-19 pandemic. We created a prototype implementing this proposed method and built an analysis platform to collect data securely in the cloud.

With the progress of technology, most of the entertainment of the younger generation is related to electronic products. People gradually forget traditional games. Red Light, Green Light has ample liquidity that people worldwide know it. Therefore, to bring traditional games back to children's childhood, we use IoT devices to develop a red Light, green light game system. This system comprises three modules: an identification module, a distance measurement module, and a notification module. In the identification module, we use the Raspberry Pi equipped with the MediaPipe Pose to detect 33 pose nodes of the human body. Then the model reads the camera image through OpenCV and marks the players' skeleton position. Finally, we compare the players' skeleton positions in the front and rear images to determine whether the player has moved. The ultrasonic sensor "HC-SR04" is used in the distance measurement module to measure the distance. The change of the front and rear distances is used to identify whether a player is passing through the finish line. The notification module places the "TM1637" four-digit display and the speaker. It will display the remaining time and play the sounds needed for the game during the gameplay. The proposed game system allows adults to relive their childhood and introduces traditional games to younger generations.

Gamification has been used for a variety of purposes, most particularly increasing user interaction and engagement. Gamification is being used to increase the efficiency of data collection by answering questionnaires, which can effectively minimize loss of information and expenses. The simple steps initiate with changing the format of the boring questions to match the game or including pertinent questions in activities while playing games. Respondents will enjoy themselves and will not feel embarrassed answering questions. Gamification can also assist with data collection and processing. This study reviews an overview of gamification and user motivations for playing games, and the researchers discuss game design elements and the accuracy rate of the gamification survey

This academic article explores the relationship between D.I.S.C personality types, avatar selection, and in-game decision-making in video games. The authors investigate how different personality types are associated with specific avatar choices and how this influences decision-making strategies during gameplay. The article describes the D.I.S.C personality assessment and its four categories: Dominance, Influence, Steadiness, and Conscientiousness. The authors discuss how avatars in video games can be customized to match players' preferences and personalities and how this can impact their gaming experience. The study recruited 431 participants who completed an online survey assessing their D.I.S.C personality type, avatar preferences, and in-game decision-making. The survey data were analyzed to identify significant differences between the personality types. The authors suggest that understanding the correlation between D.I.S.C personality types, avatar selection, and decision-making in video games can provide valuable insights into the psychology of video gameplay, leading to developing approaches to enhance player engagement and performance.

The distributed management of metaverse is necessary since the services of different companies lack standardization and can be subject to unilateral decisions regarding their terms of use and termination. On the other hand, the connection between the virtual and the real world becomes increasingly important with the proliferation of the metaverse. In this paper, we propose a system for expanding and standardizing metaverse functionality under a distributed configuration, where IPFS is used for distributed content management, enabling the incorporation of digital content and IoT devices into the metaverse. We demonstrate that the distributed file system realizes scalable sharing of the world data.

Wireless 360-degree video is an important application in 5G and beyond. However, 360-degree video consumes huge bandwidth for video transmission so as to deliver the contents of all angles to the user. To reduce bandwidth consumption, tile-based 360-degree video streaming is a promising technique. In this paper, we address two key challenges when designing a tile-based 360-degree video streaming system: (1) How to determine which tiles to transmit and their corresponding quality levels under various bandwidth conditions? (2) How to evaluate the user experience of watching a 360-degree video? We develop a tile selection algorithm for 360-degree video applications which can adapt to bandwidth condition while maintaining low Motion-to-Photon (MTP) latency to reduce cybersickness. Simulation results show the effectiveness of the proposed algorithm.

The pictorial space described in Chinese scroll paintings is based on parallel projection and thus can be extended horizontally. As a result, when displayed in secure boxes in museums or galleries, one cannot view with the intended collaboration of the hand and the eye, and the perceptual experience is lacking. This study presents an interactive device that restores the authentic viewing experience with a tangible handscroll and interactive digital content, allowing viewers to enter the façade and visually experience the space in perspective.

Many novel image retargeting algorithms have been proposed to adjust the size of images to suit different display devices while minimizing perceptual distortion. Assessing the quality of retargeted images has become an important task for developing such schemes. In this study, we propose an image retargeting quality assessment method based on the importance of objects in an image. We utilize semantic segmentation to classify pixels and assign them different importance values representing the sensitivity of human eyes to distortion. A visual saliency map is created to better match the subjective perception of humans and is then used in the "Aspect Ratio Similarity" measurement to improve its accuracy. Since human eyes tend to be more sensitive to the information loss in images without prominent foreground objects, we introduce an information loss adjustment strategy for such images. The experimental results demonstrate that the proposed method is effective in evaluating image retargeting algorithms and outperforms existing quality assessment methods.

Pitch estimation is to estimate the fundamental frequency and the midi number and plays a critical role in music signal analysis and vocal signal processing. In this work, we proposed a new architecture based on a learning-based enhancement preprocessor and a combination of several traditional and deep learning pitch estimation methods to achieve better pitch estimation performance in both noisy and clean scenarios. We test 17 different types of noise and 4 SNRdb noise levels. The results show that the proposed pitch estimation can perform better in both noisy and clean scenarios with short response time.

Computer audition based approaches for diagnosing COVID-19 can provide a low-cost, convenient, and real-time solution for combating the ongoing global pandemic. In this contribution, we present an investigation on data augmentation and multiple instance learning methods for diagnosis of COVID-19 from speech and cough sound data. We firstly introduce a novel deep convolutional neural network pre-trained on large scale audio data set, i. e., AudioSet. Moreover, we use a multiple instance learning paradigm to address the training difficulties caused by the varied length of the audio instances. Experimental results demonstrate the efficiency of the proposed methods, which can reach a best performance at 75.9 % of the unweighted average recall, surpassing the official baseline single best by 3.0 % and baseline fusion best by 2.0 %.

Due to the development of digital learning, instructional videos become a popular method for self-study. To reduce the time cost of producing instructional videos, a system for generating instructional videos automatically has been proposed. However, the types of explanatory math animation (EMA) generated from the system are limited, and the animation template design is also time-consuming. In this paper, we propose the technology for generating EMA by using the solution steps analyzed from the step-by-step solver. This framework can greatly reduce the time cost of the animation template design and can generate more various EMAs.

This paper presents an optimal contrast enhancement method to improve the visual quality of low contrast images. Two parts are involved in the proposed method: linear image fusion and patch-based contrast quality index. The linear image fusion combines the input image and its histogram equalized image with a weighting factor. The weight is then optimized by the patch-based contrast quality index. The proposed method is justified by four examples and compared with two contrast enhancement methods based on histogram equalization. The results show that our method is superior to the comparison methods and gives promising enhanced images.

Data-selective adaptive algorithms are able to reduce unnecessary weight updating for system identification problems. However, the presence of impulse noise can impede their data selection schemes. We propose a spareness-aware data-selective least mean square (DS-LMS) algorithm to improve the data selection scheme for sparse system identification under impulse noisy environments. Numerical experiments demonstrate the efficacy of the proposed approach.

In private communications or even in television broadcasting, we can sometimes just see a picture of the speaker but not the face video of the person. This may due to the inconvenience of the speaker to make appearance in front of the camera or limited channel capacity for sending out the video. This paper gives our deep learning approach and an attractive demonstration to make the face picture talk as a usual person in video, which potentially allows many industrial and consumer applications.

In this paper, we propose a prediction method for shoot events using multimodal features. Soccer event prediction has conventionally been conducted based on statistical analysis by using an enormous quantity of tracking data. In contrast, with the development of the machine learning technology, an approach that uses advanced semantic information extracted from various data such as video, audio, and text has achieved high accuracy in many tasks of soccer analysis. Therefore, it is expected to improve the prediction performance of shoot events by cooperatively using multiple modalities instead of tracking data. Specifically, the proposed method predicts shoot events by considering spatio-temporal relations of a graph constructed based on audio and visual features extracted from soccer videos. Through the experiments, we verify the effectiveness of our method.

We present a novel binaural audio generation method with data augmentation from 360° videos. Visually informed binaural audio generation requires ground truth pairs of video and binaural audio. However, collecting diverse ground truth requires a lot of effort, and low data diversity reduces the generalization performance of the model. Our method introduces the data generation from 360° videos to solve the low diversity of ground truth. Experimental results show that our method improves the generalization performance of the binaural audio generation model and that 360° video is effective in generating video and pseudo-binaural audio pairs.

Due to the importance of factory automation in industries such as manufacturing, which reduces labor costs, repetitive jobs and overcomes dangerous processing environments, this study focuses on developing a 3D visual object recognition and posture estimation method to achieve an automated grasping system. In this study, the object point cloud data is obtained by Ensenso 3D camera, where the 3D object recognition is trained based on the PointNet++ neural network model architecture. For the mechanical arm posture part, an optimal grasping database of the object is established to achieve the tasks. After the point cloud preprocessing (filtering, segmentation, centroid transfer), the iterative closest point algorithm is used to match the target object and then it obtains the rotation and translation matrices to calculate the target object grasping posture.

This paper utilized an Autonomous Mobile Robot (AMR) equipped with dual Lidars to construct a map. After building the map, several path planning methods were compared for navigation. The goal was to investigate a path planning method based on Voronoi global planner to find a short movement time with the good stability to reach the destination. The purpose was to enable AMR to find suitable map building and navigation methods in different environments and requirements.

In the realm of Intelligent Tutoring Systems (ITS), Knowledge Tracing (KT) techniques play a vital role in tracking and assessing student progress and understanding of a subject. However, in practice various data classes are generally collected in a way of underrepresentation, leading to the KT performance degradation. In this work, we propose a data deduplication technique to balance the inputs to improve the KT performance. Our experimental results confirm the efficacy of the proposed scheme in addressing imbalanced data and improving KT performance.

In the era of AI booming, object detection is an essential technology for computer vision tasks and widely adopted in autonomous driving. We propose a method to enhance object detection accuracy by adding virtual objects to real scenes through augmented reality, thereby quickly generating a large amount of data to facilitate model training. In addition, Augmented Reality (AR) can create data for rare scenarios in real worlds, such as a car flipping over on the road or a cargo overturned, which can alleviate the long-tail problem of AI models. Furthermore, our tool can generate both 2D and 3D bounding boxes directly. To verify our method, we performed transfer learning on YOLOv7 pre-trained model using 30,766 AR synthesized images of 4 traffic-related classes: Person, Car, Bicycle and Motorcycle. The new detector was evaluated on the COCO dataset. Experiments showed that our method can increase the detector accuracy as well its ability of detecting small objects.

Recently, deep reinforcement learning (DRL) methods attract much attention for solving job-shop scheduling problem (JSP), a NP-hard optimization problem. One of DRL methods is based on priority dispatching rules (PDRs), which is easy to be implemented, to dispatch operations to machines. In this paper, we propose a graph neural network (GNN) to enhance Luo's method to choose a PDR to dispatch. With GNN, our method, trained with small JSP problems, also performs well in large JSP problems. Our experiments show that our method outperforms PDR methods and most of other DRL methods, particularly for large JSP problems.

In recent years, convolutional neural networks (CNNs) have been proven and widely applied in the field of image recognition, including anomaly detection in manufacturing sites, and object detection in autonomous driving. However, the parameter obtained from the CNN increases exponentially with the depth of the network. Therefore, it is difficult to deploy the model in environments with limited computing resources. This study proposes a compression method for CNN by combining Simplified Swarm Optimization with structured pruning. Our method can compress the pruned VGG16 model to 6.7 times without sacrificing accuracy. The more important is, our method uses a heuristic approach to find the optimal pruning scheme without the need for repeated experimental verification.

Embedded AI on the edge is a growing trend, but specialized software tools and frameworks are required to develop embedded AI applications. Especially, training of AI models requires a command-line interface to execute different tasks of the development process. To reduce the development time of a custom trained model, an AI training user interface is proposed. This user interface is based on TensorFlow Lite Model Maker and designed for the use on a Coral Edge TPU. It is observed that using this user interface, the time effort to train a custom AI model is reduced by at least 40%.

One effective approach to improving the performance of cross-modal retrieval is to fine-tune a pre-trained cross-modal model. However, conventional fine-tuning approaches usually require plenty of computational resources. To alleviate such a requirement, we propose a parameter-efficient tuning method of a pre-trained model via prompt learning for cross-modal retrieval. Obtaining inspiration from the prompt learning technique in natural language processing, our method constructs a multi-dimensional vector as a prompt for the cross-modal retrieval, and the prompt with a few parameters is optimized to achieve better retrieval performance. We conducted experiments on the open dataset, and the results verify that our proposed method is effective and parameter-efficient.

A proficiency-level estimation method using heterogeneous features by supervised fractional-order embedding multi-view canonical correlation analysis via ordinal label dequantization (SFEMCCA-OLD) is presented in this paper. The proposed method constructs a common latent space by SFEMCCA-OLD, and estimates proficiency-level using projected features to the common latent space. Then we can utilize new features that have the powerful representation ability since SFEMCCA-OLD enables the construction of the common latent space without dimensionality limitation. Consequently, the performance improvement of the proficiency-level estimation becomes feasible.

The Internet of Things (IoTs) have become ubiquitous in all aspects of public needs today. The application of IoT technology is crucial for promoting energy-saving behavior. This paper presents a control system used for smart home devices based on hand gesture control and Android applications. The hardware design uses the ESP32 microcontroller, which has an in-chip Wi-Fi module, making it very suitable for creating IoT application systems. The software was designed using the Arduino IDE, and the application display design was developed using Kodular Creator, which is used to control the lamps and switches. The experiment results show that the electrical equipment can be activated using the Android application with gesture sensors by waving down, up, left, and right.

Identifying the service to be connected from an IP flow is beneficial for various applications, and a method of identifying the service based on the SNIs occurring in the flow was proposed. It was also shown that the identification accuracy could be improved by excluding some SNIs from those for identification. However, no discussion has been published on the probability of SNIs occurring in actual service accesses. In this paper, we compare the probability of occurrence of each SNI in each service and show that there are redundant SNIs whose probabilities are the same, and there is one SNI that occurs in all services with similar probabilities and is not useful for service identification. We then show that identification accuracy can be improved by excluding the SNI that occurs in all services from the service identification data.

ESG disclosures are attracting attention. Narrative disclosure allows companies to overcome information asymmetries by presenting more information. This study uses ESG key metrics provided in the textual data of sustainable crowdfunding projects to construct predictive models for the success of the sustainable crowdfunding project. This study uses exposed ESG-related textual data to predict crowdfunding project success. It contributes to the research on the deployment of ESG performance and signaling theory.

Traffic flow is one of the most important information for traffic management. Traditionally, the data were obtained only from intersection monitors, which lack a macroscopic view of the entire roadway. Recently, the applications of unmanned aerial vehicles (UAV) have been widely applied to many fields and have become popular. Due to the programmable path planning and 3D movement characteristics of UAVs, we integrate edge computing for image recognition processing with UAVs to perform traffic flow analysis. This study successfully developed a prototype system to analyze the road segment video that is recorded from UAV-mounted cameras. A deep learning technique will be used to perform vehicle identification and tracking tasks. The average vehicle speed and vehicle flow can then be determined. In addition, violation event detection (including speeding, illegal parking, etc.) can also be reported. The system will automatically produce the diagnosis report. It can greatly reduce the burden of traditional manual image viewing, and the analyzed results can be used for traffic management units to improve traffic strategies.

The realization of ISO-15118-based electric vehicle (EV) charging communication makes complete the economic profit brought by EVs. In the standardized timeout rule, the latency of message exchange between the EV and charging station shall be lower than the threshold. We implement ISO-15118-based communication and measure the latency under different numbers of coexistent access point users. Experiment results present the user coexistence's effect on latency. We analyze the reason for potential timeout occurrence and provide suggestions about ISO-15118-based communication in the future.

Graphs are non-Euclidean data structures used to model interaction relationships between abstract objects. When additional structures, such as graph orientation, weighted vertices, or weight functions on edges, are imposed on the graph, additional graph properties can be applied to discover hidden patterns within the structure of the graph. In this paper, we discuss possible granular structures for graphs, with a focus on granular structure for edge-weighted digraphs derived from pairwise comparison problems. These structures have been applied in various fields such as recommender systems, graph databases, and social network analysis.

As the number of devices connected to the network increases, the management becomes more complicated and the possibility of security problems increases. In this paper, we developed a security diagnosis system that can verify whether there are any problems with the devices in the wireless LAN or the settings of the wireless LAN itself. Our system has functions such as verification of wireless LAN settings, remote login verification, and Wi-Fi vulnerability verification. Through experimental evaluations, we confirmed the basic performance of our system.

This research proposes a method for generating or modifying 3D objects using a human's voice as input for a machine learning model. The input voice is translated into text using a Google API and processed by a BERT-based text encoder to understand the meaning of each word and generate the desired 3D object. The method is designed to make it easier for people with no design knowledge to create 3D objects and contribute to the development of the Metaverse, a virtual reality beyond reality. The shape generation model has been trained on a dataset of 15,038 shapes with 75,344 natural language descriptions, and the system includes a shape auto-encoder to generate different styles of 3D shapes. The proposed method is the first to use speech recognition for the generation of 3D objects in the Metaverse and aims to generate higher quality and faster results than existing methods using text for 3D object generation.

Geneticists face significant challenges in identifying susceptibility genes for complex diseases, and one of the most important machine learning methods used in gene-gene interaction studies is multifactor dimensionality reduction (MDR). However, MDR's method involves searching for all gene combinations to find the one with the highest accuracy, which can be computationally expensive when dealing with large datasets or numerous combinations. To address this issue, we propose an improved method that uses the grey wolf optimizer to reduce computational time in MDR for oral cancer risk prediction, resulting in more effective detection of gene-gene interactions.

Crowd counting is a computer vision task that focuses on accurately estimating the number of people present in a given scene. In recent years, deep learning techniques based on convolutional neural networks (CNNs), have shown remarkable success in various computer vision tasks, including crowd counting. In the field of crowd counting, large-scale changes have always been a great challenge. To resolve this problem, previous work used multiple branches to obtain information at different scales and combined it. However, purely combining multi-branch features cannot effectively utilize multi-scale information. In this work, we improve the previous multi-branch architecture, which can reasonably select the appropriate scale information. Furthermore, we test our model on the ShanghaiTech dataset and demonstrate the competitive performance of our method.

Segmentation of high-resolution remote sensing images is a challenging task. Some recent studies have used complex convolutional neural networks to address this problem, and they can only extract multi-scale features and improve object boundary discrimination. However, these neural networks are difficult to train due to high complexity. In this paper, we propose two new attention modules, Local Channel Spatial Attention and 16-piece Local Channel Spatial Attention, which can focus on improving the feature representation of each local area. In order to effectively utilize multi-scale features, we also propose a Twice Decoder Module, which is integrated into our model to improve the discrimination ability of target boundary. Our proposed method is verified through extensive experiments and experimental results show that our model can better improve the overall accuracy compared with other state-of-the-art methods.

Facial expression recognition in the wild could be affected by irregular occlusions, inconsistent face angles and varying light levels on facial images. In this work, we proposed a model containing facial attention module, residual in residual dense block and vision transformer to tackle with this problem. The facial attention module can assist the model to extract more useful features. Residual in residual dense block and vision transformer can help the model extract more detailed features. In addition to, the number of images of different facial expressions in the dataset is also very different. We proposed a new loss function to deal with this issue. Experimental results on two datasets in the wild show that our model is indeed effective. We also created a new facial expression dataset called Fairfaceplus, which is an extension of FairFace, with annotation of expression categories to the original labels. The code of our proposed method will be available on GitHub https://github.com/dreampledge/fairfaceplus.

Anomaly detection is an open and challenging problem that aims to detect anomaly in future samples. In this study, we explore a simple but effective solution that utilizes multi-level feature combination in a student-teacher network to improve the prediction result. Our approach combines low-level, middle-level, and high-level features extracted from ResNet18 to capture a range of features from different layers of the network. Through the use of a student-teacher network, we select the best possible generated features from ResNet18 to enhance the prediction performance. Our results demonstrate that combining features from different levels of the network enhances the model's ability to learn and recognize anomalous patterns, and thus improves the accuracy of anomaly detection. Our proposed student-teacher network with ResNet18 backbone achieves a prediction score of 92.80% and 96.90% for Image AUC and Pixel AUC respectively.

Alzheimer's disease image analysis is a challenging task due to the limited availability of labeled data. Labeling data requires precision and patience and is typically done manually by medical experts. However, human labeling is expensive, time-consuming, and prone to human errors. To address this issue, we propose an approach that automates the selection and labeling process using a deep learning model. Specifically, we employ YOLOv7 to reduce time and errors. In this work, we implement our proposed idea on 2D images from the Alzheimer's dataset. This work focuses on labeling the hippocampus region from the coronal view of brain MRI images. The results of our experiments show that our approach yields result that are quite close to the ground truth labels produced by experts. The mean average precision (mAP) compare with ground truth is about 2% below, while the time consumption is reduced by 80%.

The detection of lumbar disc herniation (LDH) through magnetic resonance imaging (MRI) poses a challenge due to the various shapes, sizes, angles, and regions associated with bulges, protrusions, extrusions, and sequestrations. One potential solution is using deep learning methods to identify lumbar abnormalities in MRI images automatically. The YOU ONLY LOOK ONCE (YOLO) model series has gained popularity for training deep learning algorithms for real-time biomedical image detection. This study aims to assess the performance of the latest YOLOv7 in detecting LDH across different regions of the lumbar intervertebral disc. The analysis revealed that YOLOv7 exhibits a poor performance and low detection rate of LDH across the L1-L2, L2-L3, L3-L4, L4-L5, and L5-S1 regions.

Alzheimer's disease is the most common type of dementia that causes many of the functions of the human brain to be severely weakened. To date, there has not been a cure for Alzheimer's disease. Therefore, early diagnosis is needed using MRI images with the help of a classification program. Deep learning using the Convolutional Neural Network (CNN) method is receiving increasing attention because of its excellent performance. Its architecture can be modified according to user needs based on the data to be processed and the approach for classifying, detecting, and segmenting visual objects. In this paper, we offer a classification of Alzheimer's disease using one of the architectures on CNN, namely Visual Geometry Group-19 (VGG-19), with a sagittal view of MRI images with an image size of 229 x 229 pixels. The classification accuracy of the described method is 94% for the validation set.

Breast cancer is one of the most common cancers among women in our country. Magnetic resonance imaging (MRI) has become a widely adopted medical imaging technology in the healthcare industry. It can help medical professionals accurately determine the size and scope of breast cancer. This study proposes a 3D Dense U-Net-ASPP model using a private dataset. The model addresses the issue of large data volume in image segmentation preprocessing and reduces the problem of feature loss by combining Atrous Spatial Pyramid Pooling (ASPP) and Attention Gate. Furthermore, the model's performance is improved by adjusting the loss function through a combination of Focal Loss and Dice Loss.

Masked faces can pose a challenge for security systems, making it easier for individuals to disguise themselves and evade identification. To address this issue, developing a system that can recognize masked faces can help in identifying potential security threats and preventing criminal activities. Therefore, this research proposes a deep learning-based approach, namely FaceNet-based conditional generative adversarial network (FaceNet-cGAN), to detect and remove masks, which would generate a high-quality, unmasked facial image and improve the accuracy of facial recognition systems. The proposed FaceNet-cGAN approach demonstrated superior performance compared to other existing methods, as evidenced by the obtained SSIM score of 92.2% and PSNR value of 29.97 dB.

Mice have been an important reference for the drug test. Currently, it is only possible to collect the feedbacks from mice through non-verbal methods. To analysis the behavior of mice with machine learning, there are usually two major challenges. The first one is individual variation on facial expression or behaviors, which might require a huge amount of data set to overcome. The second one is how to obtain reliable labels, which are fundamental to train a robust machine learning model. This study aimed on the analysis of different classification architectures along with the effects of training samples and features to reduce the impact of the above two challenges.

Underwater image restoration has gained more and more attention recently due to its several applications in marine environmental surveillance-related tasks. In this paper, a novel unsupervised GAN (generative adversarial network)-based deep learning framework for single underwater image restoration is proposed. Without needing paired training images, we introduce contrastive learning with feature and style reconstruction loss functions in our unsupervised GAN-based structure to learn an image generator for translating underwater images to the corresponding in-air images. Extensive experiments have shown that the proposed method outperforms (or is comparable with) the state-of-the-art deep learning-based methods relying on paired/unpaired training data quantitatively and qualitatively.

This paper presents a design and implementation of an automated guided vehicle (AGV) that addresses the dynamic obstacle avoidance problem in navigation. We employ the TEB path planner, which is based on laser range sensors, as the algorithm for dynamic obstacle avoidance. In the development process, the Stage simulation is previously conducted to obtain suitable parameters. The environment used in the Stage simulation is a flat map of the real vehicle verification site, and the robot specifications and sensors used are the same as those of the actual vehicle. Finally, we apply the parameters obtained from the simulation to the real AGV for validation.

Due to the rise of Industry 4.0, many industries have begun to replace traditional human labor with automated equipment to achieve greater productivity and more accurate inspection results. Due to its advantages of high stability, high accuracy, and low repetition error, many manufacturing industries have chosen to use robotic arms to replace human labor in production and inspection tasks to obtain faster and more accurate work results. Based on the above advantages, this experiment combines a seven-axis robot arm with 2D LiDAR to scan the object to be inspected, and the algorithm is used to analyze, calculate and plan the inspection path of the arm so that the robot arm can automatically scan the surface of the object to be inspected in an equal speed, smooth, and reproducible way.

The main goal of our paper is about using smartphones to detect road damage conditions. The proposed system contains three parts, including mobile application, road damage detection, and cloud computation. The mobile application was developed to send photos captured by the user to the cloud server. The cloud server would perform AI algorithms to detect and classify road damage conditions and send the data to the database. Moreover, this application would show the statistical results to the user through a simple user interface. The developed system would benefit road damage detection in a simple and low-cost way.

This paper introduces a parametric method for optimizing the wire electric discharge machining (WEDM) process applied to steel machining. The proposed approach employs soft-computing methods to determine the optimal process parameters for achieving maximum material removal rate and minimum surface roughness. This work conducts an analytical investigation to evaluate the feasibility of the proposed approach.