arxivcs.CV2026-07-06
ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection
Joongwon Chae, Lihui Luo, Yang Liu, Dongmei Yu, Peiwu Qin, Runming Wang, et al.
Memory-based anomaly detection is attractive because it localizes defects from normal images without training a decoder or synthesizing pseudo anomalies. However, most memory methods still use the memory bank as a nearest-neighbor lookup table: a test patch is treated as normal i…