Single-Class Detection Method for Bar Material Surface Defects Based on Self-Supervised Learning

This paper addresses the issue of surface defect detection in steel bars, proposing a one-class method based on self-supervised learning and thoroughly explores its effectiveness in online detection scenarios. Due to the limited number and incomplete variety of defect samples in the actual production environment, the traditional deep learning algorithms that rely on supervised learning face great challenges. On the contrary, defect-free samples are abundant and easy to obtain, so the single-classification task trained with single-class samples is more suitable for this scenario. In this study, normal samples were enhanced by DotPaste, PatchPaste and LinePaste to generate pseudo-samples with similar defect samples, which supports the self-supervised learning framework to train the model without real defect samples. At the same time, the class activation mapping (CAM) technology is introduced to realize the accurate positioning of defect positions. After experimental verification, the method proposed in this paper performs excellent in the surface defect detection of steel bars. The AUC value reaches 0.9934, which is significantly better than the traditional method, effectively solving the problem of scarce defect samples, and providing an efficient and reliable solution for the online detection of the surface defects of steel bars.