Observing the Effects of Image Quality on Object Detection Using YOLOv5
Kshitij Salil Malvankar (Technological University of Shannon: Midlands and Midwest & The NPD Group, Ireland); Enda Fallon (Athlone Institute of Technology, Ireland); Kieran Flanagan (The NPD Group, L. P., Ireland); Paul Connolly (The NPD Group, Inc, Ireland)
Document information extraction, which combines item classification with object localization within a scene, is a major difficulty in computer vision. With the advent of modern advances in deep learning, significant advancements in object detection have been made. Majority of research is focused on designing increasingly more complex object detection networks for improved accuracy, such as YOLOv5, SSD, R-CNN, Faster R-CNN, and other extended variants of these networks. This paper proposes to use the YOLOv5 algorithm to identify data in an invoice and also observe the effects of degraded image quality on the performance of the algorithm. The BRISQUE score is utilized to parameterize the quality of the image. The difference in performance under the same training conditions between three different variants of the YOLOv5 algorithm are also detailed.
Journal: International Journal of Simulation- Systems, Science and Technology- IJSSST V23
Published: no date/time given