Steel surface defect detection algorithm based on improved YOLO11

With the development of industrial automation, steel surface defect detection is very important to improve production quality. Traditional methods are limited by efficiency and accuracy, but deep learning technology provides new ideas for this. In this paper, a steel surface defect detection algorithm (YOLO-GH) based on improved YOLO11 is proposed. By replacing C3k2 module in backbone network with C3k2_HFERB module, high-frequency feature capture capability is enhanced. A global-local self-attention module (GLSA) is introduced into the neck network to optimize multi-scale feature fusion. The experiment was based on NEU-DET data set, and the results showed that the mAP50 of the improved model reached 81.4%, which was 5.5% higher than that of the basic model, the parameter number was reduced to 2.52M, and the computation amount was only 6.1 GFLOPs, which was significantly better than the mainstream models such as YOLOv7 and YOLOv8. The algorithm achieves a balance between precision and efficiency, and provides an efficient solution for industrial real-time inspection.