A Normalizing Flow-based Unsupervised Anomaly Detection Approach

Abstract

Currently, unsupervised industrial anomaly detection and anomaly localization based on deep learning have achieved great success. The most commonly used dataset for industrial anomaly detection and anomaly localization is MV-TAD, and the most commonly used evaluation metric is AUROC. Most research methods are based on the above datasets and experiment evaluation metrics. Although the most advanced method has nearly 100% AUROC index values in the above dataset, the results are still not ideal, as seen by observing the anomaly segmentation maps after the experiments. Therefore, in this paper, we use a new metric, F1-measure, to evaluate the experimental performance of industrial anomaly detection and anomaly localization models. Compared with the AUROC metric, the F1-measure ensures complete and accurate detection by reconciling Precision and Recall. We use ResNet34 and WideRes-Net101 pre-trained encoders based on the current state-of-the-art normalized flow-based generative model anomaly detection method to train and test on the dataset, and achieve good experimental performance. In addition to MVTAD, we extended the dataset to DAGM 2007, BTAD, trained and tested them. Furthermore, we use the new evaluation metric, F1-measure, to evaluate the experimental results.