HR-Pro: Hierarchical Residual Prototype Memory for Efficient Industrial Anomaly Detection

Abstract

Industrial anomaly detection is a critical task in intelligent manufacturing systems. Existing memory-based methods generally require storing a large number of patch-level normal features, resulting in high storage demands and reduced inference efficiency in large-scale settings. To address this issue, this article proposes Hierarchical Residual Prototype Memory (HR-Pro), a compact industrial anomaly detection framework based on hierarchical prototype memory, residual normal patch augmentation, and score-level fusion. HR-Pro constructs level-specific prototype memories from multi-level backbone features and augments them with a small set of residual normal patches that are insufficiently represented by prototypes, thereby improving the coverage of under-represented normal patterns while maintaining compact memory storage. Experimental results on MVTec AD and VisA show that, compared with same-backbone PatchCore 10% baselines under ResNet50 and WideResNet50 settings, HR-Pro reduces memory-bank storage by 73.4% on MVTec AD and 76.1% on VisA. It also improves inference speed by more than $150\times $ on MVTec AD and $370\times $ on VisA. Meanwhile, HR-Pro achieves strong pixel-level localization performance and provides a favorable accuracy–efficiency trade-off.