Defect identification method for steel surfaces based on improved YOLOv5

Abstract

Traditional machine vision detection methods suffer from low accuracy in identifying small-scale defects. To address this, a nondestructive identification method for steel surface defects is proposed based on an enhanced version of the fifth version of the You Only Look Once(YOLOv5)algorithm. In this improved approach, the Res2Block module is incorporated into the backbone of the YOLOv5 algorithm to expand the receptive field and improve computational efficiency. Additionally, the recursive gated convolution structure is fused into the neck of the YOLOv5 algorithm to further enhance the computational performance of the surface defect identification method. To validate the effectiveness of the proposed method, a series of ablation experiments were conducted using different module combinations. These results were then compared with those obtained through other object detection methods. This comparison reveals that the proposed method achieves a mean average precision of 67.8% and an F₁-score of 86.0% in steel surface defect identification. When compared with the original YOLOv5 algorithm, the proposed method exhibits superior performance, particularly in the identification of small-scale steel surface defects. Furthermore, it also surpasses other object detection methods, such as SSD, YOLOv3, YOLOv5-Lite, and YOLOv8, demonstrating significant improvements in computational accuracy.

Keywords

steel / defect detection / convolutional neural network / You Only Look Once(YOLO)