Abstract:
Automated inspection of raw materials for steel and wood surfaces in manufacturing settings is experiencing unprecedented development thanks to the integration of vision-...Show MoreMetadata
Abstract:
Automated inspection of raw materials for steel and wood surfaces in manufacturing settings is experiencing unprecedented development thanks to the integration of vision-based systems using object detection models. Considering time efficiency, most current surface defect detection approaches are based on YOLOv5 (You Only Look Once, fifth version). Meanwhile, the state-of-the-art in real-time object detection has largely evolved, and the more recent systems have yet to be evaluated for surface defect detection. Thus, this paper presents the results of an extensive effectiveness assessment of YOLOv5, YOLOv6 v3.0, YOLOv7, YOLOv8, YOLOv9, and Real-Time Detection Transformer (RT-DETR) by evaluating the trade-offs between inference speed and accuracy reached by these models for defect detection. Experimental results show that YOLOv6 v3.0 maintains the highest level of combined performance when considering both mean Average Precision and Frames Per Second metrics and using industrial datasets of wood and steel defects.
Date of Conference: 06-09 August 2024
Date Added to IEEE Xplore: 12 September 2024
ISBN Information: