BHF-YOLO: Efficient Detection Network for Steel Surface Defects Based on Lightweight Multi-Scale Feature Fusion

Steel surface defect detection is a key link to ensure the quality of industrial products, but the traditional methods have problems such as low detection efficiency and large number of model parameters in complex backgrounds. In this paper, we propose an improved YOLOv8n lightweight steel surface defect detection algorithm. First, a bidirectional dynamically weighted feature pyramid network (BDWFPN) is designed to enhance multi-scale feature characterization through cross-layer interaction and adaptive weighting; second, a heterogeneous scale-shared convolutional pyramid (HSCP) is constructed to improve the detection accuracy of small targets by utilizing parameter sharing and multi-scale cavity convolution; lastly, the depth-separable convolution and spatial-channel-attention (SCSA) that fuses depth-separable convolution and space-channel-attention (FEL-Block) is embedded into the c2f FEL-Block module to optimize the feature extraction efficiency. Experiments show that the improved model achieves 79.6% mAP@50 with 2.07M parameters and 6.4GFLOPs on the NEU-DET dataset, which improves the accuracy by 4.1% over the baseline accuracy, and the detection speed reaches 204.6 frames/s, providing an efficient solution for industrial real-time detection.