Surface Defect Detection of Steel Based on Improved YOLOv5s

Abstract

To address the issues of low accuracy and frequent occurrences of missed and false detections in steel surface defect detection, this paper proposes an algorithm named GSM-YOLO, which is based on YOLOv5s. Firstly, the Ghost Bottleneck module is introduced to improve the C3 module of the backbone network. The Ghost module is used to generate diverse feature maps to obtain more detailed information and increase the detection accuracy of the model. Secondly, a Spatial and Channel Reconstruction Convolution (SCConv) is incorporated to refine the feature information generated by the C3 module, making the output feature information more precise and consequently improving the detection accuracy of the algorithm. Finally, a Multi-dimensional Collaborative Attention (MCA) mechanism is added at the bottom of the SPPF module. Spatial and channel feature information from the three-branch structure of MCA is fused to enhance the model's attention for multi-scale target features, and then missed and false detections are reduced. The experimental results show that the proposed algorithm achieves a recall rate (R), precision (P), and mean average precision (mAP) which are 2.8%, 5.3% and 5.2% higher than those of the original network, respectively, on the NEU-DET dataset. This effectively enhances the ability to detect surface defects in steel materials.