MemFlow-AD: An anomaly detection and localization model based on memory module and normalizing flow

Abstract

Presently, in most anomaly detection methods, the training dataset contains a low frequency of anomalous data with diverse categories. However, these methods exhibit limited learning capacity for anomalous information, leading to weak model generalization and low detection accuracy. This paper proposes MemFlow-AD, an anomaly detection and localization model that integrates a memory module and a normalizing flow. MemFlow-AD supplements anomalous data using anomaly simulation, retains general patterns from normal samples via the memory module to discern potential differences between normal and anomalous samples, employs a 2D normalizing flow to extract distributional feature information from the data, and through multiscale feature fusion and attention mechanism to further enhance the feature expression ability of the model. Experimental results demonstrate outstanding performance in detecting and localizing anomalies on the MVTec dataset, achieving accuracies of 98.61% and 94.02%, respectively. Moreover, on the BTAD dataset, the model exhibits a 2.15% improvement in detection accuracy compared to current mainstream methods.