Zero-shot visual anomaly detection (ZSAD) aims to identify and localize anomalies in unseen categories without target-domain adaptation, yet heterogeneous anomaly patterns and cross-domain variations challenge reliable detection. Existing approaches primarily rely on local visual features, patch-text similarities, or memory-based retrieval, leaving inter-patch relational information underexplored in anomaly evidence modeling. In this paper, we propose RACE-AD, a relation-aware evidence routing and confidence-guided adaptation framework that integrates local anomaly responses with attention-derived relational information for generalizable ZSAD. To capture inter-patch relationships, we characterize the dispersion of visual attention distributions using Shannon entropy, providing complementary relational statistics beyond local anomaly evidence. To organize heterogeneous evidence, we introduce relation-aware routing that combines memory-derived normal/anomaly responses with attention entropy to construct image-conditioned representations through entropy-regularized soft aggregation. To mitigate unreliable retrieval responses under domain shift, we design confidence-guided adaptation that regulates anomaly-conditioned representations according to the strength of the memory anomaly evidence and the relational dispersion before visual-semantic alignment. Extensive experiments on ten industrial and medical benchmarks demonstrate the effectiveness of RACE-AD, achieving average image-level AUROC and AP scores of 91.9% and 93.7%, and pixel-level AUROC and AUPRO scores of 93.6% and 80.0%, respectively. Ablation and diagnostic analyses show that entropy-guided routing mainly improves region-level localization and that the confidence gate is essential under domain shift.
翻译:暂无翻译