Attention-guided YOLOv5s-SDF model for accurate detection of strip steel surface defects

Surface defects on strip steel are often inevitable due to limitations in raw materials and manufacturing processes. To improve defect detection accuracy, we propose the YOLOv5s-SDF algorithm, which integrates ShuffleAttention in the neck and DyHead in the head of the YOLOv5s framework. The novelty lies in the use of ShuffleAttention to model spatial-channel dependencies and the incorporation of Focal-CIoU loss to mitigate the influence of low-quality samples. Experiments on the NEU-DET dataset show that YOLOv5s-SDF achieves 69.37% Precision, 74.21% Recall, 76.32% [Formula: see text], and 38.60% [Formula: see text], outperforming the baseline YOLOv5s by 0.2%, 1.72%, 3.57%, and 0.84%, respectively.