Surface Defect Detection of Strip Steel Based on YOLOv5

Abstract

 Given the various kinds of surface defects and the lack of obvious features, which lead to detection error and leakage detection, a method of strip surface defects detection with a decoupling head, YOLOv5s, is proposed. Firstly, the K-means + + method was utilized to relocate the anchor frames to produce an optimised coupling between the transcendental and the real frame. SimAM's attention free mechanism was incorporated into the Neck network to improve the performance of the model. Finally, the head of the network is replaced by the decoupling head, which separates classified tasks from regressive tasks, thus enhancing convergence and recognition accuracy. The results indicate that the average precision of the proposed algorithm is 80.7% on the NEU-DET dataset, an increase of 3.2% compared to YOLOv5s, and a transfer frame number of FPS of 50 per second, which balanced detection accuracy and operational efficiency. The increased accuracy of detection as compared to other techniques meets the requirements of precision and timeliness.