NFCF: Industrial Surface Anomaly Detection with Normalizing Flow Cross-Fitting Network

The effective fully-supervised defect detection methods in industry are done by training many defect labels. These methods face application challenges because of insufficient defect samples. Moreover, the subtlety of defects, the similarity between defect-free and defect regions, and the interference factors carried by defect-free images also pose problems. To solve these difficulties above, we propose a Normalizing Flow Cross-Fitting network (NFCF) to use only defect-free images as priori knowledge for training. Specifically, the incremental broadening module in the network focuses on information expansion for minor defects, and the interactive filtering module completes the weight filtering through same-scale mutual reasoning. The processed information is fitted to a defect-free distribution by the normalizing flow module and the distribution is used as the basis for the determination. The NFCF does experiments on four types of datasets, and it achieves an average of 96.3% and 95.88% on two metrics, AUC-Image and AUC-Pixel. Experimental results show that the network can distinguish defect images using only defect-free training. Moreover, visualization experiments show that it can also complete localization segmentation for minor defects. Based on semi-supervision with achieved accuracy, NFCF alleviates the labeling dependence problem of defects and significantly increases the application value.

Automated summary

The study proposes a Normalizing Flow Cross-Fitting network (NFCF) that uses only defect-free images as prior knowledge for training to address the issue of insufficient defect samples in industry. The network includes an incremental broadening module for information expansion of minor defects and an interactive filtering module for weight filtering through same-scale mutual reasoning. The processed information is fitted to a defect-free distribution by the normalizing flow module, serving as the basis for determination. Experimental results show that NFCF achieves an average of 96.3% and 95.88% on two metrics, AUC-Image and AUC-Pixel, on four types of datasets. The network can distinguish defect images using only defect-free training and perform localization segmentation for minor defects.