In the production of shaft metal parts, accurate detection of endface defects is critical to ensure proper assembly and operational safety. However, in the early stages of production, the extremely low occurrence of defects, the scarcity of available defect samples, and the high cost and time consumption of manual data labeling pose significant challenges for traditional supervised learning methods. To address these issues, this paper proposes a two-stage weakly supervised defect detection framework focused on the endface of shaft parts. In the first stage, an unsupervised anomaly detection model (URS-AD) is employed to identify the majority of anomalous regions on the metal surface. This method achieves a detection rate of 99.2% with a false positive rate of 1.6%, and AUROC scores of 0.986 at the pixel level and 0.937 at the image level. In the second stage, regions with similar anomalous features are grouped into defect categories by using Mean Shift clustering algorithms to generate pseudo-labels. These pseudo-labels are then used to train an improved UNet model for fine-grained defect segmentation under a supervised learning framework. The experimental results show that the improved UNet model outperforms the comparison methods in segmentation performance on imprecisely labeled data, with increases of 4.7% in Mean IoU and 6.4% in Dice coefficient. The proposed two-stage framework offers a practical and effective solution for addressing the challenges of limited samples and costly labeling in early-stage endface defect detection for shaft metal parts.
Liao et al. (Sat,) studied this question.