An Improved YOLOv5 Model for Steel Surface Defect Detection by Integrating Multi-Dimensional Atrous Convolution

In deep network-based steel surface defect detection models, using only serial convolution operations can hinder the detection of small defects due to insufficient multi-scale detail information. To address this issue, a detection model based on joint multi-dimensional atrous convolution (JMDC) is proposed. By introducing atrous convolution with various kernel sizes into the feature extraction layer, a multi-channel parallel convolution chain is constructed to capture multi-scale defect features, enhancing the detail perception. Subsequently, a JMDC operator, generated from the residual network, fuses these features and establishes a channel constraint to improve the training stability and global feature extraction. Then by integrating the JMDC into the multi-level feature extraction layers, the improved YOLOv5 realizes the detection enhancement for the multi-scale defects. Correspondingly, the k-means algorithm is adopted to generate adaptive multi-scale anchor boxes with the reduction of the positional deviation. Experimental results show that the proposed approach has provided a detection accuracy of 81.7% on the NEU-DET dataset, yielding a significant improvement on the defect detection, especially for small targets.