On the Problem of Consistent Anomalies in Zero-Shot Industrial Anomaly Detection

TMLR Paper5414 Authors

Abstract: Zero-shot image anomaly classification (AC) and anomaly segmentation (AS) play a crucial role in industrial quality control, where defects must be detected without prior training data. Current representation-based approaches rely on comparing patch features with nearest neighbors in unlabeled test images. However, these methods fail when faced with consistent anomalies—similar defects that consistently appear across multiple images—leading to poor AC/AS performance. We present Consistent-Anomaly Detection Graph (CoDeGraph), a novel algorithm that addresses this challenge by identifying and filtering consistent anomalies from similarity computations. Our key insight is that for industrial images, normal patches exhibit stable, gradually increasing similarity to other test images, whereas consistent-anomaly patches show abrupt spikes after exhausting a limited set of images with similar matches. We term this phenomenon ``neighbor-burnout'' and engineer a robust system to exploit it. CoDeGraph constructs an image-level graph, with images as nodes and edges linking those with shared consistent-anomaly patterns, using community detection to identify and filter out consistent-anomaly patches. To provide a theoretical explanation for this phenomenon, we develop a model grounded in Extreme Value Theory that explains why our approach is effective. Experimental results on MVTec AD using the ViT-L-14-336 backbone show 98.3\% AUROC for AC and AS performance of 66.8\% (+4.2\%) F1 and 68.1\% (+5.4\%) AP over state-of-the-art zero-shot methods. Additional experiments with the DINOv2 backbone further enhance segmentation, achieving a 69.1\% (+6.5\%) F1 and a 71.9\% (+9.2\%) AP, demonstrating the robustness of our approach across different architectures.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Satoshi_Hara1

Submission Number: 5414