Deep feature clustering for multi-class industrial image anomaly detection

Abstract

Existing unsupervised multi-class anomaly detection algorithms usually train unified reconstruction networks to capture the distribution of all classes simultaneously. However, under such a challenging setting, popular reconstruction networks need to be elaborately designed to avoid the “identical shortcut”. In addition, the distribution of each category is different, which may mean different requests for expression ability. To solve these problems, built on the intuitive “classification-then-detection” idea, we utilize clustering algorithm to expose the category information hidden in the pre-trained deep features, then propose a simple and application-friendly approach for multi-class anomaly detection. The proposed approach consists of Category Anchor Construction (CAC), Category Information Mining (CIM) and Local Feature Routing (LFR). Firstly, CAC is proposed to extract the corresponding pre-trained features from a small subset of training images to construct category anchors, preserving the valuable category information provided by the training set. Then, CIM is introduced to mine category information embedded in pre-trained features by category anchors voting and acquires the category labels. Finally, to achieve multi-class anomaly detection, we propose LFR, splitting multi-class distribution into multiple single-class distributions according to category labels so that separate single-class anomaly detection heads can be trained to express them. In spite of simplicity, the proposed method outperforms state-of-the-art algorithms in terms of accuracy and stability on the widely used MVTec-AD, VisA, MVTec-LOCO, MPDD and BTAD datasets.