Multi-Level Normalizing Flow for Comprehensive Anomaly Detection and Localization

Unsupervised anomaly detection identifies deviations from normal patterns as anomalies. Recently, unsupervised methods have made significant strides in anomaly detection. However, single-scale feature extraction struggles to capture subtle anomalies and existing methods frequently emphasize exclusively on local information while disregarding global semantic information. In this paper, we propose a new normalizing flow called Multi-Level Normalizing Flow (MLFlow) for anomaly detection and localization. First, we input normal images and extract multi-scale features using a pre-trained feature extractor. Second, MLFlow receives the multi-scale feature maps and density estimate. StepFlow and ConvergeFlow are the two main modules of MLFlow. Specifically, the StepFlow independently transforms each layer of feature maps, allowing the lower layer to capture detailed features, the middle layer to extract local features and the top layer to extract semantic information. Additionally, the ConvergeFlow combines transformed multi-scale feature maps, enhancing the comprehensive analysis capability for anomalies. Experimental results on MVTec AD, BeanTech AD and VisA datasets reveal that the proposed method performs exceptionally well in anomaly detection and localization tasks, surpassing existing methods and achieving the state-of-the-art performance.