TPCR-AD: Test-Patch Clustering Refinement with Normal-Feature PCA Reconstruction for Few-Shot Industrial Anomaly Detection

Abstract

Few-shot industrial anomaly detection typically relies on a small number of normal samples to establish a normalfeature representation. Although recent methods built upon strong pretrained features such as DINOv2 have improved detection performance, they still tend to produce scattered noise and discontinuous anomaly responses under complex textures, local specular reflections, and boundary perturbations. To address this issue, we propose a new paradigm for few-shot industrial anomaly detection based on test-patch clustering refinement and normal-feature PCA reconstruction. Specifically, initial anomaly cues are first derived from reconstruction residuals in a PCA normal subspace, and the anomaly scores are then refined through clustering consistency among test patches. This design characterizes anomaly deviation while enhancing the continuity of defect regions and suppressing isolated noise and local false positives. Finally, the reconstruction residuals and clusteringrefined responses are fused to generate anomaly maps and image-level anomaly scores. Experiments on MVTec AD and VisA show that, under the 4-shot setting, the proposed method achieves P-AUROC scores of 97.42% and 97.84%, respectively, demonstrating its effectiveness for few-shot anomaly localization.