DGS-YOLO: A Lightweight YOLOv8-Based Method for Steel Surface Defect Detection

To overcome the limitations of high parameter count and insufficient detection accuracy inherent in applying the YOLOv8n model to steel surface defect detection, we propose DGS-YOLO, an improved lightweight detection model. This model incorporates two principal enhancements. Firstly, the YOLOv8n backbone is redesigned by integrating a novel D-Starblock module, significantly bolstering multi-scale feature extraction capabilities. Secondly, the original neck structure is replaced with a GS-FPN module, achieving a more efficient feature fusion process and contributing to the model’s lightweight design. Collectively, these modifications yield superior detection performance and enhanced model efficiency. Experimental validation on the NEU-DET dataset demonstrates that DGS-YOLO attains a mean Average Precision (mAP) of 77.4%, surpassing the baseline YOLOv8n by 1.9% while concurrently reducing the parameter count by 45.2%. Experimental results confirm that DGS-YOLO effectively addresses the practical requirements for high-precision, lightweight industrial defect detection systems.