Steel surface defect detection algorithm based on improved YOLOv12

To tackle the shortcomings of existing algorithms in detecting small defects amidst complex backgrounds, we present an enhanced YOLOv12-based steel surface defect detection algorithm, designated as P-YOLO. The introduction of pinwheel-shaped convolution (PConv) in the backbone of the original model, enhances the capture of spatial features, thereby improving the model’s capability to recognize defects on intricate surfaces. Experiments conducted using the publicly available NEU-DET steel defect dataset demonstrate that the improved model outperforms YOLOv12, achieving increases of 3.8% in mAP@50 and 0.7% in mAP@50–95, along with improvements of 0.9% in recall and 1% in precision. Furthermore, generalizability experiment on GC10-DET dataset indicates that P-YOLO algorithm exhibits a significant degree of generalization and effectively fulfills the task of detecting surface defects on steel.