A dual reverse distillation scheme for image anomaly detection

Image anomaly detection and localization are challenging in industrial applications due to highly unbalanced distributed real-world data. Reverse Distillation (RD), utilizing a teacher–student model, has been employed for this task, yet it often suffers from over generalization issue, as the student network struggles to effectively recover abnormal features. To address this issue, this paper proposes a dual reverse distillation (DRD) scheme that combines pseudo abnormal and normal images and restores their features simultaneously with the reverse distillation method. Specifically, DRD incorporates a novel multi-feature cascade fusion (MFCF) block, a combined knowledge distillation loss function, and a detection refinement method. The MFCF block enhances feature integration across layers and captures more comprehensive information, while the knowledge distillation-loss function balances features from batch and individual images to better align output features with normal features. Furthermore, a refinement method at the inference stage further improves detection accuracy. Such strategies help improve the recovery of abnormal features while reducing over generalization. Experiments with many public datasets, including MVTec, BTAD and VisA, were conducted to verify effectiveness of the proposed scheme. Results show that, for MVTec, it achieved 99.46% AUROC at the image level and 98.39% AUROC at the pixel level, outperforming RD by 1.02% and 0.59%, respectively. Ablation experiments also confirmed that DRD can mitigate the over generalization issue effectively.