Recognition of anomalies of an a priori unknown type

In the present article we propose a modification of the PaDiM anomaly detection method which maps images to vectors and then calculates the Mahalanobis distance between such vectors and the distribution of the vectors of the training set. Of the coordinate axes of the vectors we choose a subset of such that the distribution along them is close to normal according to the chosen statistical criterion. The uniformization procedure is then applied to those coordinates and the Mahalanobis distance is calculated. This approach is shown to increase the ROCAUC value in comparison with the PaDiM method.