Hierarchical Vector Quantized Transformer for Multi-class Unsupervised Anomaly Detection

• DOI: 10.48550/arxiv.2310.14228
• Publication Date: 2023-01-01
• Authors: Ruiying Lu; Yujie Wu; Long Tian; Dongsheng Wang; Bo Chen; Xiyang Liu; Ruimin Hu
• Keywords: Interpretability; Discriminative model

Unsupervised image Anomaly Detection (UAD) aims to learn robust and discriminative representations of normal samples. While separate solutions per class endow expensive computation limited generalizability, this paper focuses on building a unified framework for multiple classes. Under such challenging setting, popular reconstruction-based networks with continuous latent representation assumption always suffer from the "identical shortcut" issue, where both abnormal samples can be well recovered difficult distinguish. To address pivotal we propose hierarchical vector quantized prototype-oriented Transformer under probabilistic framework. First, instead learning representations, preserve typical patterns as discrete iconic prototypes, confirm importance Vector Quantization in preventing model falling into shortcut. The prototype is integrated reconstruction, that data point flipped point.Second, investigate an exquisite relieve codebook collapse issue replenish frail patterns. Third, optimal transport method proposed better regulate prototypes hierarchically evaluate score. By evaluating MVTec-AD VisA datasets, our surpasses state-of-the-art alternatives possesses good interpretability. code available at https://github.com/RuiyingLu/HVQ-Trans.