Zero-shot industrial anomaly detection and localization via reweighted density estimation

To achieve high quality industrial Anomaly Detection (AD) is difficult as there are no anomalous samples in the training set and one need to treat the defective region in the test images as outliers. Recently, some researchers even propose to perform the AD in a zero-shot manner. The algorithm is required to predict the defective state of each test image without the training phase. Surprisingly, the pioneering work has achieved relatively high accuracy in this challenging scenario thanks to the smart low density assumption of the anomalous patches in the feature space. In this paper, we propose to improve the successful zero-shot method by introducing the dynamic density estimation of the patch features. The experiment results illustrate a significant performance gain in all the involved 4 metrics for anomaly detection or localization.