Surface Defect Detection of Steel Products Based on Improved YOLOv5

The defect detection of steel surface is the key to ensure the quality of steel products, while the real-time detection and detection accuracy have not kept pace with the automation of the processes. The above issues are caused by two aspects: (1) the high production speed improves the real-time requirements of detection; (2) the surface defects of steel with irregular shapes and numerous small defects increase the difficulty of accurate detection. This paper provides an improved algorithm based on YOLOv5 to resolve the above issues. First, YOLOv5s is used as the basic network structure to ensure the high-speed detection. Second, a deformable convolution method and a dual attention module are adopted so that the defects of different shapes and sizes can be accurately detected. Experimental results show that the improved YOLOv5s can maintain a high speed of 60.51FPS while improving the detection accuracy up to 0.974mAP. The smallest irregular defect size that can be detected is Sx7 (mm
²
), which is 3.4 times smaller than that of YOLOv5. And the improved YOLOv5 algorithm can fully suitable for the surface defect detection of the steel production line.