An improved steel defect detection model using multi-scale feature fusion based on YOLO-MFD

To address the challenges of limited detection accuracy for small targets in complex backgrounds and the difficulty in model deployment due to a large number of parameters in steel surface defect detection, this paper proposes a lightweight detection model named YOLO-MFD, based on YOLO11n. First, to tackle the issue of missed detection caused by blurry defect images and indistinct edge features, a Multi-scale Edge Feature Enhancement (MEFE) module is designed, which enhances the discrimination capability for subtle defects by integrating multi-scale edge feature differences. Subsequently, in the feature fusion stage, a Global-to-Local Spatial Attention Bidirectional Feature Pyramid Network (GLSA-BiFPN) is constructed, which simplifies the feature propagation path while improving the fusion efficacy of cross-scale semantic information. Finally, a Group detection head (Ghead) is adopted to replace the original detection structure, significantly reducing the number of parameters through the introduction of grouped convolution while maintaining accuracy. Experimental results on the NEU-DET dataset demonstrate that YOLO-MFD achieves an mAP of 83.8%, representing a 5.9% improvement over YOLO11n, while reducing the number of parameters by 27% and computational cost by 6.3%. The results indicate that the proposed model effectively balances detection accuracy and lightweight design, making it suitable for real-time steel surface defect detection in complex industrial scenarios.