Yolo-HF: A Novel High-Frequency Attention-Based Approach for Enhanced Fine Defect Detection on Cold Rolled Strip Steel

To address the problem of low detection accuracy for fine defects on the surface of cold rolled strip steel, this paper proposes a new fine defect detection algorithm, Yolo-HF. Based on frequency domain analysis, Yolo-HF introduces a high-frequency attention extraction module, alongside Residual Feature Distillation Self-Attention (RFDSA) and Multi-Scale Efficient Spatial (MES) modules, to effectively capture and fuse semantic information and multi-scale features. In the detection head design, four different-scale detection heads are employed to better handle the varying sizes of target bounding boxes and categories. Experimental results indicate that the Yolo-HF model achieves a Mean Average Precision (mAP) of 94.3% and a Frames Per Second (FPS) of 89.29 on the self-constructed Cold Rolled Strip Steel (CRSS) dataset, demonstrating a significant improvement in fine defect detection accuracy and verifying its effectiveness in detecting surface defects on cold rolled strip steel.