EDB-YOLO: An enhanced multi-scale feature fusion model for steel surface defect detection

Abstract To address the challenges in steel surface defect detection, such as diverse morphologies, complex hierarchies, a high proportion of small-sized defects, insufficient accuracy, and high computational complexity, we propose an improved real-time detection model, EDB-YOLO, based on YOLO11. First, we introduce an efficient multi-scale residual attention (EMRA) module to fuse multi-scale features, thereby enhancing detection accuracy. Second, we design a dynamic feature extraction (DFE) module to enhance the model’s ability to capture contex-tual information, effectively suppressing background noise and improving the feature extrac-tion capability for small defects. Additionally, a novel bidirectional enhanced feature pyramid network (BE-FPN) is designed. This network achieves efficient fusion and enhancement of multi-resolution features through hierarchical cross-scale interaction mechanisms. Experiments on the NEU-DET and GC10-DET datasets show that EDB-YOLO model achieves 81.2% and 70.3% mAP50, surpassing YOLO11 by 5.1% and 2.5%, respectively, while maintaining competitive parameter counts and computational complexity. Additionally, EDB-YOLO achieves 84 s and 85 frames per second on the two datasets, respectively. The model achieves an optimal trade-off between detection accuracy, computational cost, and efficiency, making it highly suitable for steel surface defect detection.