CONAN: CONcept-Oriented ANomaly Detector for Interpretable Zero-Shot Detection

초록

Large vision-language models (LVLMs) have advanced zero-shot anomaly detection (ZSAD) by enabling generalization to unseen classes, yet most operate as semantic “black boxes” without explicit rationale. Incorporating concept bottleneck models (CBMs) is challenging because anomalies are condition-centric and transient, unlike stable object attributes. We propose CONAN (CONcept-oriented ANomaly Detector), an interpretable framework that performs active concept curation through an object-agnostic concept bank organized into five macro-level anomaly dimensions. By aligning CLIP image features with category prototypes using LVLM-generated pseudo-labels, we inject fine-grained anomaly semantics into the embedding space. Experiments show strong zero-shot performance on industrial benchmarks, robust cross-domain generalization to medical datasets, and support for human-in-the-loop explanation refinement.

키워드