CM -YOLO: A Highly Accurate YOLOv5 S-Based Model for Steel Surface Defect Detection

Steel surface defect detection is essential for quality control in manufacturing, but traditional methods often struggle with fine-grained defects and complex textures. Although deep learning models like YOLOv5 have shown promise, they face challenges in defect localization and feature representation. In this paper, we propose an enhanced YOLOv5 model, named CM- YOLO, with the following ideas: 1) enhancing feature extraction by adaptively recombining features via content-aware interpolation, which helps preserve spatial details and enhances the representation of small defects, and 2) improving bounding box regression by considering the maximum possible distance between the predicted box and the ground truth box, which improves localization accuracy, especially for overlapping or irregularly shaped defects. Experimental results show that the detection accuracy of the proposed CM- YOLO model is better than other models, providing a more effective solution for steel surface defect detection in industrial applications.