A Fast Detection Method of Steel Plate Surface Defects Based on Weight‐Sharing Dilated Convolution and Self‐Attention Mechanism

ABSTRACT Surface defects on steel plates have a direct influence on product integrity and operational reliability. The wide variety of defect types and significant size variations in steel materials pose a major challenge for defect detection. This study proposes a novel steel surface defect detection method, designated as FMD‐YOLO, which incorporates weight‐sharing dilated convolution and a self‐attention mechanism. The C2f module is optimized by employing PConv and PWConv as the primary operators to construct the FasterCSP module, thereby reducing memory usage and computational redundancy while markedly lowering both the model's complexity and its processing demands. A Multi‐Scale Shared Convolution Feature Pyramid Network (MSCFPN) is developed to extract multi‐scale features through shared‐weight convolutional layers with varying dilation rates, effectively decreasing computational cost. Furthermore, a Down Sampling with Self‐Attention (DSADown) module is introduced, utilizing depthwise separable convolution to capture local image features. The integration of a channel self‐attention mechanism enables flexible adaptation to feature maps of varying scales and emphasizes critical features, thereby enhancing the detection of small targets. Experimental results show that the proposed method achieves a mean average precision (mAP) of 77.7% on the NEU‐DET dataset, outperforming YOLOv10n by 2.1%, while reducing parameters by 18.5% and computational cost by 32.3%. Compared to other state‐of‐the‐art detection algorithms, the proposed approach demonstrates superior performance. Validation on the industrial‐grade GC10‐DET dataset further confirms its robust generalization capability.