Steel Surface Defect Detection Algorithm Based on DM-YOLOv11

To address the issues of multi-scale steel surface defects and the poor multi-scale feature processing of existing algorithms, this study proposes an improved detection modeling algorithm, called DM-YOLOv11. Firstly, the C3k2 module of the backbone network is integrated with the Dynamic Weight Reparameterization (DWR) module, expanding the receptive field at higher network levels to enhance multi-scale target recognition. Secondly, the Multi-path Coordinate Attention (MPCA) is integrated into the backbone network, enhancing the localization and discrimination of multi-scale defective features. Lastly, the Wise-IoUv2 loss function is employed to dynamically adjust the gradient modulation factor and incorporate non-monotonic gradient focusing, enhancing regression robustness. The experimental results demonstrate that the mean average precision (mAP) of DM-YOLOv11 on the NEU-DET and GC10-DET datasets improves by 4.1% and 2.6%, reaching 80.6% and 71.4%, respectively. The proposed method is highly effective for steel defect detection in industrial settings, showcasing its strong potential for robust generalization performance.