Reproduce “Prototypical Residual Networks for Anomaly Detection and Localization” on OpenResearch

About this paper

Hui Zhang, Zuxuan Wu, Zheng Wang, Zhineng Chen, Yu-Gang Jiang

A Prototypical Residual Network, coupled with online anomaly generation strategies, improves anomaly detection and localization by learning explicit feature residuals and capturing multi-sized inconsistencies. The framework achieves state-of-the-art performance on MVTec AD, DAGM, BTAD, and KolektorSDD2 datasets, showing an average Pixel-AUROC of 99.0% and AP of 78.6% on MVTec AD, surpassing previous supervised methods by significant margins.

Key insights

• Explicitly learning feature residuals against multi-scale normal prototypes allows the model to precisely identify deviations from expected patterns.
• Incorporating multi-size self-attention enables the network to effectively detect local inconsistencies and anomalies of diverse sizes.
• Combining in-distribution (Extended Anomalies) and out-of-distribution (Simulated Anomalies) generation significantly enhances model robustness and generalization to both seen variations and unseen anomaly types.

Summary by alphaXiv.