Improving the Steel Surface Defect Detection Algorithm of YOLOv5 Network

Abstract

Automated surface defect detection in industrial steel production is critical for ensuring product quality, but remains a challenging task due to the prevalence of small, low-contrast defects and the multi-scale nature of these anomalies. Current detection systems often struggle with insufficient accuracy, particularly for small targets, and a lack of robustness across varying defect sizes. To address these limitations, this paper proposes a series of targeted improvements to the YOLOv5s algorithm. First, at the input stage, the K-Means++ algorithm is employed to recluster the dataset and generate optimized initial anchor boxes, which provides a better prior for the model to learn from and improves localization, especially for small defects. Second, an attention mechanism is integrated into the backbone network to enhance feature representation. This module enables the model to focus computational resources on more informative spatial regions and channel features associated with defects, effectively suppressing irrelevant background noise and amplifying subtle defect signatures. Comprehensive experiments were conducted on a dedicated industrial steel defect dataset. The results demonstrate that the improved algorithm achieves a mean Average Precision (mAP@0.5) of 83.3%, representing a significant 6.2% increase over the baseline YOLOv5s model. Crucially, this performance gain is achieved without sacrificing inference speed; the enhanced model maintains a real-time detection rate of 96.5 frames per second (fps) on a standard GPU. These findings confirm that the proposed enhanced YOLOv5s algorithm successfully balances high precision with real-time processing capabilities, making it a viable and effective solution for automated visual inspection in demanding industrial environments such as steel manufacturing.