Research on Surface Defect Detection of Steel Strip Based on Deep Learning

Surface defect detection of steel strip has become an indispensable part of quality inspection in steel strip production. An enhanced algorithm is suggested by this paper, GRW-YOLO, based on the YOLOv5 model, which solves the problem of insufficient precision in surface defect detection of steel strips and meets the requirements of real-time detection in industry. First, the backbone network incorporates the GAM (Global Attention Mechanism) to enhance the model's capacity in representing targets. Secondly, the Res2Net is used to reconstruct the C3 module to reduce the number of parameters and improve the model's generalization ability and precision. Then, the WIoU bounding box loss function is used to solve the problem of category imbalance and improve the model's robustness. Finally, to enable the network to detect objects of different scales, a detection layer is incorporated. Experimental results show that GRW-YOLO achieves a mean average precision (mAP) of 81.9% on the NEU-DET dataset, which is an improvement of 5.8% compared to YOLOv5s. Additionally, the detection speed reaches 67 FPS, surpassing the industrial requirement of 30 FPS. This demonstrates a favorable balance between accuracy and speed for GRW-YOLO.