Pixel-Level Feature Clustering Learning for Image Anomaly Detection and Localization

Abstract

Image anomaly detection and localization not only need to provide image-level anomaly judgment but also need to locate pixel-level anomaly areas. This paper proposes a model named pixelAD, which builds an end-to-end network through pixel- level feature clustering learning to solve this problem. The normal prototype is obtained during training by clustering the normal pixel-level features. We generate pixel-level cluster labels of normal samples according to the prototypes, which guide the model to update parameters by calculating the assignment loss. For inference, pixelAD directly outputs the pixel-level anomaly score end-to-end. The experimental results of the real industrial dataset MVTecAD show that PixelAD has an excellent performance in anomaly detection and anomaly localization.