What methods are used for anomaly detection in images (PaDiM, PatchCore)?

PaDiM and PatchCore are modern anomaly detection methods for images based on patch analysis (small fragments of an image).

• PaDiM (Patch Distribution Modeling) builds a statistical model of patch feature distributions of normal images using a multivariate normal distribution. When testing new images, the method assesses how much the patches deviate from the normal distribution, identifying anomalies.

• PatchCore uses caching of normal image patch representations and applies a nearest neighbor approach to assess anomalies. For each patch of the test image, the distance to the nearest patches from the normal set is calculated, which helps in detecting deviations.

Both methods are effective for tasks where anomalies are localized and require detailed analysis of small image regions.