Self-supervised dual-layer 2D normalizing flow method for industrial anomaly detection

Abstract

To address the shortcomings in feature representation and abstraction ability of NFs in the field of unsupervised industrial anomaly detection, as well as the ambiguity in determining the decision boundary between normal and anomaly features, this study proposes the SDNF. This method is dedicated to accurately identifying anomalies by mastering the differences in the distribution of normal and anomaly features. The adopted AFS enables the model to more clearly define the decision boundaries, while the application of the DNF enhances the feature representation and abstraction capabilities. By utilizing the mapping principle of NF, we successfully distinguished between normal and anomaly features, achieving efficient anomaly localization. Furthermore, the ESAM is introduced into the DNF’s sub-networks, allowing the model to focus more on anomaly features, thereby significantly improving the performance of anomaly detection. Experimental validation shows that this method significantly improves the image-level and pixel-level AUC metrics on the MVTec AD dataset, reaching 99.33 % and 98.48 % respectively, surpassing previous research results in terms of the accuracy and quality of anomaly detection. While ensuring high-precision anomaly detection, this research also balances the complexity and inference efficiency of the model, contributing to the advancement of industrial anomaly detection. The related code has been made open source, for more details please visit https://github.com/FutureIAI.