A Masked Reverse Knowledge Distillation Method Incorporating Global and Local Information for Image Anomaly Detection

Knowledge distillation is an effective image anomaly detection and localization scheme. However, a major drawback of this scheme is its tendency to overly generalize, primarily due to the similarities between input and supervisory signals. In order to address this issue, this paper introduces a novel technique called masked reverse knowledge distillation (MRKD). By employing image-level masking (ILM) and feature-level masking (FLM), MRKD transforms the task of image reconstruction into image restoration. Specifically, ILM helps to capture global information by differentiating input signals from supervisory signals. On the other hand, FLM incorporates synthetic feature-level anomalies to ensure that the learned representations contain sufficient local information. With these two strategies, MRKD is endowed with stronger image context capture capacity and is less likely to be overgeneralized. Experiments on the widely-used MVTec anomaly detection dataset demonstrate that MRKD achieves impressive performance: image-level 98.9% AU-ROC, pixel-level 98.4% AU-ROC, and 95.3% AU-PRO. In addition, extensive ablation experiments have validated the superiority of MRKD in mitigating the overgeneralization problem.

💡 Research Summary

The paper addresses a critical limitation of knowledge‑distillation‑based anomaly detection (AD): the tendency of student networks to over‑generalize because the input and supervisory signals are identical (both derived from normal features). To break this equivalence, the authors propose Masked Reverse Knowledge Distillation (MRKD), which converts the conventional reconstruction task into a restoration task that explicitly forces the student to convert abnormal representations back to normal ones. MRKD consists of two complementary masking strategies: Image‑Level Masking (ILM) and Feature‑Level Masking (FLM).

In ILM, normal training images are randomly masked using the Normal Sample Augmentation (NSA) technique: patches are cut from the same image and pasted elsewhere, creating synthetic abnormal images that retain realistic texture but disrupt global context. Both the original normal image and the synthetic abnormal image are fed into a frozen WideResNet‑50 teacher network (pre‑trained on ImageNet). The teacher extracts feature maps for the normal image (used as the supervisory signal) and for the abnormal image (used as the input signal). The student network is then trained to regress the normal feature map from the abnormal one, i.e., to “restore” the missing global information. This forces the student to learn high‑level semantic relationships rather than merely copying pixel values.

ILM alone, however, may neglect fine‑grained local cues such as edges and small texture variations. FLM addresses this by randomly masking individual pixels in the student’s output feature map and passing the masked map through a lightweight generation module (a 1×1 convolution followed by ReLU). Because neighboring pixels contain sufficient contextual information, the generator learns to infer the missing values, thereby strengthening the student’s ability to capture local correlations.

The loss is based on cosine similarity between the restored student features and the teacher’s normal features, encouraging alignment in the high‑dimensional feature space. A bottleneck module with random weights compresses the abnormal teacher features before they are fed to the student, reducing redundancy and sharpening the restoration target.

Experiments on the widely used MVTec AD benchmark (15 categories, both image‑level and pixel‑level evaluation) demonstrate that MRKD achieves 98.9 % image‑AUROC, 98.4 % pixel‑AUROC, and 95.3 % AU‑PRO, surpassing state‑of‑the‑art methods such as RD4AD, DRAEM, PatchCore, and memory‑augmented autoencoders. Ablation studies show that removing either ILM or FLM degrades performance, confirming that global and local information are both essential. Compared to reconstruction‑based approaches (AE, VAE, GAN) and memory‑based methods (MemAE, MGNAD), MRKD requires only a teacher network, a student network, and a simple generation block, resulting in lower computational cost and faster inference.

In summary, MRKD introduces (1) a novel reverse‑distillation paradigm that eliminates input‑supervision equivalence, thereby mitigating over‑generalization, and (2) a dual‑masking scheme that jointly captures global semantics and local texture cues. These contributions provide a more robust and efficient framework for unsupervised image anomaly detection, with potential impact across industrial inspection, medical imaging, and video surveillance applications.