A Multi-Class Anomaly Detection Method Based on Reverse Distillation and Mixed-Attention

Reverse distillation is an efficient unsupervised anomaly detection framework that localizes anomalies by exploiting representation discrepancies between the teacher and student networks. However, despite being trained only on normal images, the student network can often reconstruct anomalous regions accurately - a phenomenon termed over-generalization, which degrades anomaly detection performance, especially in the multi-class setting. To enhance the reverse distillation framework's ability in the multi-class setting, we propose a Mixed-Attention Feature Selection (MAFS) module. By fusing spatial attention, channel attention, and self-attention, MAFS guides the model to select the most discriminative features for reconstruction while blocking potential anomalous information propagation to the student network, thereby alleviating over-generalization. Additionally, we design a more effective Multi-Scale Feature Fusion (MSFF) module, which leverages multi-scale features from the teacher network to enhance the reconstruction of normal regions. Following the unsupervised multi-class anomaly detection setting, our method achieves competitive performance on MVTecAD and VisA datasets.