Detection of Surface Defects in Steel Based on Dual-Backbone Networks-MBDNet-Attention-YOLO

Steel surface defect detection plays a crucial role in industrial quality inspection. However, automated recognition remains challenging due to the complexity and diversity of defect types, such as cracks, pores, scratches, and dents. Traditional detection methods often struggle to balance accuracy and efficiency, especially in scenarios with complex backgrounds or small defect sizes. Although mainstream algorithms have made progress in improving recognition performance, they commonly exhibit limitations in sensitivity to small defects, robustness against background interference, and practicality for industrial applications. In order to confront the aforementioned difficulties, this paper proposes an enhanced surface defect detection method named MBY (MBDNet-Attention-YOLO), which integrates the MBDNet backbone with the YOLO framework. First, the backbone network incorporates modules such as HGStem，Dynamic Align Fusion，and C2f-DWR to achieve efficient extraction and fusion of multi-scale features while optimizing computational resource utilization. Second, in the neck structure, we design a MultiSEAM feature fusion module to enhance synergy across multi-scale features, thereby strengthening the model’s capability in detecting subtle defects and handling complex backgrounds. Furthermore, the Inner-SIoU cost function is incorporated to improve bounding box regression accuracy and accelerate training convergence by refining the alignment between predicted and ground-truth boxes. Experiments conducted on two public steel defect datasets, NEU-DET and PVEL-AD， demonstrate the effectiveness of our method. The MBY model achieves 85.8% mAP@0.5 on NEU-DET and 75.9% mAP@0.5 on PVEL-AD, outperforming state-of-the-art defect detection algorithms. These results validate that the proposed approach not only enhances detection accuracy but also exhibits strong generalization capability and practical potential.