Feature Reconstruction via Reverse Distillation for Multi-class Anomaly Detection

Jiaxiang Wang, Haodi Xu, Zheyuan Cai, Xiaotong Tu

Abstract: Unsupervised anomaly detection has attracted considerable attention in industry due to its broad applicability and practical significance. However, most existing methods are based on a one-class-one-model paradigm, which poses significant challenges in model management and incurs high computational and storage costs. In this paper, we target the more challenging yet practical task of multi-class unsupervised anomaly detection. Following the multi-class-one-model paradigm, we propose a simple and effective anomaly detection framework based on knowledge distillation, termed FRRD. To improve the accuracy and precision of feature reconstruction, we introduce a teacher knowledge injection module, where a student model is guided by knowledge transferred from a teacher model, enabling more precise reconstruction. In addition, to enable multi-level feature integration while preventing trivial identity mapping, we design a feature compression module that compresses and fuses multi-level features into low-dimensional representations. This not only increases the reconstruction difficulty during training but also mitigates shortcut learning. Finally, we employ a dynamic hard-mining loss to further encourage the model to focus on regions that are more difficult to reconstruct, thereby enhancing anomaly localization performance. Extensive experiments demonstrate the effectiveness of our approach, achieving superior performance in both anomaly detection and localization tasks across the MVTec AD, VisA, and Uni-Medical datasets.

External IDs:dblp:conf/icic/WangXCT25