Improved YOLOv5 Network for Steel Surface Defect Detection

Steel surface defect detection is crucial for ensuring steel quality. The traditional algorithm has low probability. This paper proposes an improved based on the YOLOv5 model to enhance Firstly, deformable convolution introduced in backbone network, and a module replaced by convolution; secondly, CBAM attention mechanism added network; then, Focal EIOU used instead of CIOU loss function YOLOv5; lastly, K-means cluster Anchor box, box parameters that are more suitable this obtained. experimental results show using can get feature information, which conducive learning network. uses mechanism, heat map shows beneficial extraction. optimized high wide compared with function, accelerates convergence model. favorable achieved probability 78.8% NEU-DET dataset, 4.3% better than original inference time each image only increased 1 ms; therefore, proposed effective.