Optimizing PatchCore for One-Shot Industrial Anomaly Detection

PatchCore is a widely used algorithm for industrial anomaly detection, due to its high performance compared to alternative approaches. This paper focuses on optimizing its architecture for the challenging case of one-shot anomaly detection. More precisely, we modified the feature extraction step by substituting the original ResNet50 with the Anti-aliased ResNet50, and the dimensionality reduction phase by replacing the original Average Pooling with the Gaussian Random Projection. Comparative experiments conducted on the MVTec Anomaly Detection (MVTec-AD) dataset showed that the proposed modifications can significantly increase performance with respect to the original PatchCore on several common categories of objects.