DDR-YOLO: An Efficient Detection Model Based on YOLO11s for Steel Surface Defect

In the process of steel production, surface defects on the product are inevitable. Due to challenges such as high surface reflectivity and significant variation in defect scales on steel strips, this paper proposes an improved YOLO11-based algorithm to enhance the accuracy of steel surface defect detection. First, we propose a Dual-Branch Dynamic Structure Convolution module (DualSPDG), featuring a dual-channel architecture comprising a backbone and an auxiliary branch. This module preserves texture details and structural information during downsampling and enhances the backbone network’s feature extraction capability. Second, to improve the detection of small-scale defects, we propose a Dynamic Attention Multi-Scale Fusion (DAMF) module, which imposes dynamic weights on the fusion features after upsampling and summing fusion, and realizes more discriminative cross-scale feature fusion. Finally, steel surface defects often exhibit complex shapes, such as elongated strips and arcs. To extract such large-structure information, we propose the restructurable parameter fusion large-kernel convolution attention aggregation block (RepLK-AA). This module combines large-kernel depthwise separable convolution with attention mechanisms to parallelize multiple receptive fields, thereby enhancing feature representation capability. Ablation and comparative experiments demonstrate that DDR-YOLO outperforms existing detection algorithms. Compared with YOLO11s, our method improves mAP50 by 3.6% on NEU-DET dataset and 3.9% on GC10-DET dataset. The proposed method significantly enhances detection accuracy without compromising speed, meeting the industrial requirements for real-time steel surface defect detection.