FreePatch: a training-free multiscale patch matching method for few-shot industrial anomaly detection

Abstract

Industrial anomaly detection requires both image-level anomaly recognition and pixel-level defect localization. In real production lines, anomalous samples are scarce, defect patterns are hard to enumerate, and product categories may change frequently. Category-specific training methods therefore incur high deployment costs, while simple patch matching is often limited by single-layer representations and insufficient coverage of normal variations under few-shot conditions. This paper proposes FreePatch, a training-free multi-scale patch matching method for few-shot industrial anomaly detection. FreePatch extracts layer2 and layer3 features from a frozen ImageNet-pretrained ResNet-50 and builds a patch memory bank from only a few normal reference images. Its core design consists of multi-scale normalized PatchScore fusion and density-calibrated nearest-neighbor scoring. The former uses leave-one-out statistics from normal references to robustly normalize anomaly scores across feature layers, improving the comparability of multi-scale scores. The latter calibrates each anomaly score according to the reference stability of the matched normal patch, reducing false alarms caused by valid but rare normal patterns. On all 15 categories of MVTec AD under the 4-shot setting, FreePatch improves Image AUROC from 0.9015 to 0.9235 and Pixel AP from 0.4194 to 0.4903. These results show that FreePatch improves few-shot industrial anomaly detection and localization without anomalous samples or backbone updates.