Reconstruction-based distillation for anomaly detection

Anomaly detection plays an important role in industrial production. Recent advances have established knowledge distillation as a prominent anomaly detection method, leveraging the paradigm where a student network learns feature representations from a pre-trained teacher network. In practice, this traditional feature imitation strategy leads to overgeneralization, which degrades detection performance. To mitigate this limitation, this paper proposes a reconstruction-based distillation network that replaces direct feature imitation with feature reconstruction. This method improves the student network’s understanding of the semantic information of features. In addition, to improve the accuracy of the student network in predicting anomalous regions, we introduce a prediction consistency loss to ensure that the predictions of the student network are consistent in the training phase with the inference phase. Extensive experiments on the MVTec AD and VisA datasets validate the effectiveness and generalization capability of our method. On the MVTec AD benchmark, our method achieves 99.61% image-level AUROC for anomaly detection and 98.23% pixel-level AUROC for anomaly localization.