HDC-YOLOv8s: An Improved YOLOv8s Algorithm for Steel Surface Defect Detection

To address the issues of low contrast and strong background interference in steel surface defect detection, which lead to high false positive and false negative rates, an improved YOLOv8s steel surface defect detection algorithm, HDCYOLOv8s, is proposed. Firstly, a lightweight backbone network based on HGNetV2 is used to replace the backbone network of YOLOv8s, enhancing feature extraction capabilities while reducing model parameters and computational load. Secondly, a hybrid dilated convolutional residual (HDCR) module is designed to replace the original C2f module in YOLOv8s’s neck, effectively integrating multi-scale information and enhancing feature representation. Furthermore, a dynamic weighted downsampling (DWD) module is introduced to adaptively adjust feature weights and reduce information loss during the feature extraction process. Lastly, a content-guided attention (CGA) feature fusion module is incorporated to adaptively aggregate low-level and high-level information through learnable spatial weights, improving detection performance. Experimental results demonstrate that the improved YOLOv8s algorithm achieves mAP50 of $81.1 \%$ and mAP50:95 of 48.1% on the NEU-DEF dataset, with improvements of $4.0 \%$ and $2.9 \%$ over the baseline algorithm, while reducing parameter count and computational load by $26.8 \%$ and $29.3 \%$ respectively.