Maximizing the performance of reverse distillation in anomaly detection

Knowledge distillation has been widely used in unsupervised anomaly detection, but it often suffers from over-generalization. Reverse distillation was introduced to address this issue, but its effectiveness is limited by the capacity of the teacher encoder and the loss of fine-grained reconstruction details. To address these limitations, the Unlocking the Potential of Reverse Distillation (URD) framework integrates an expert network and a guided information injection (GII) module to enhance anomaly localization. However, URD still suffers from two major limitations. First, its over-reliance on the expert network may limit detection performance. Second, the GII module brings only limited improvements to anomaly detection. In this work, we propose a reverse distillation method enhanced with a sensitivity factor derived from expert network features. The sensitivity factor fuses expert and teacher features, serving as a more robust optimization target for the student network. To reduce manual category-specific tuning, a calibration-based semi-automatic strategy is introduced to estimate the sensitivity factor without using the test set. In addition, the GII module is redesigned as a new information fusion module (IFM), which employs conditional control to balance feature transfer from the teacher to the student, thereby jointly improving anomaly detection and localization. Finally, an inference-stage optimization strategy is developed to further enhance performance. Extensive experiments on industrial and medical datasets demonstrate that the proposed method achieves competitive and balanced performance among existing unsupervised anomaly detection approaches based on reverse distillation. In particular, using the semi-automatically estimated sensitivity factor, our method achieves 99.7% detection AUC and 98.8% localization AUC on the MVTec AD dataset, demonstrating strong and balanced performance under the reverse distillation framework. The code is publicly available at https://github.com/GE-123-cpu/MRD.