Steel Defect Detection Method Based on Improved YOLOv8

Steel, a crucial material in construction and machinery, demands high surface quality. This paper presents an enhanced YOLOv8 for steel surface defect detection. To detect small defects better, the original C2f module is replaced with the new DBBNCSPELAN4 module, improving small - defect detection accuracy. For better feature fusion, the traditional Concat module in the Neck network is swapped with the ContextGuideFusionModule, enhancing fusion and reducing computational complexity. To address overlapping defect issues, the WaveletPool module is added to certain convolutional layers in the Backbone and Neck, strengthening the model’s robustness and multi - scale feature extraction. Experiments show that, though with increased inference time and complexity, the model improves Precision by $5.3 \%$, Recall by $2.0 \%$, and mAP50 by $3.0 \%$, significantly enhancing steel defect detection accuracy, especially for tiny defects.