Abstract: Unsupervised anomaly detection is widely used in industrial quality inspection and other domains due to its requirement of only normal samples for training.Existing single-class anomaly detection methods exhibit a significant performance decrease when directly applied to multi-class anomaly detection.Among them,knowledge distillation-based anomaly detection methods distill the feature knowledge of pre-trained teacher models on normal samples into student models.However,they have the problem that they can't guarantee thestudent models learn only normal sample knowledge in multi-class anomaly detection.This paper proposes an unsupervised multi-class anomaly detection method,PRD(Prototype based reverse distillation),based on a reverse knowledge distillation framework.It utilizes the Multi-class Normal Prototype module and Sparse Prototype Recall training stra-tegy to learn prototypes of multiple-class normal sample features from the teacher model.These prototypes are then used to filter the input features of the student model,ensuring that the student model only learns the feature knowledge of normal samples from the teacher model.PRD surpasses existing state-of-the-art methods on various industrial anomaly detection datasets.Qualitative,quantitative,and ablation experiments validate the effectiveness of the PRD framework and its internal modules.

Key words: Anomaly detection, Unsupervised learning, Prototype learning, Knowledge distillation, Pre-trained features