PromptAD: Zero-Shot Anomaly Detection Using Text Prompts

Yiting Li, Adam Goodge, Fayao Liu, Chuan-Sheng Foo; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024, pp. 1093-1102

We target the problem of zero-shot anomaly detection, in which a model is pre-trained on a set of seen classes and expected to detect anomalies in other unseen classes at test time. Although providing exceptional results for many anomaly detection (AD) tasks, state-of-the-art AD algorithms catastrophically struggle in zero-shot scenarios. However, if knowledge of additional modalities exist (e.g. text), we can compensate for the lack of visual information and improve the AD performance. In this work, we propose a knowledge-guided learning framework, namely PromptAD, which achieves the compatibility of a abnormality view and a normality view through a dual-branch vision-language decoding network. Concretely, the normality branch establishes a normality profile to exclude anomalies. Meanwhile, the abnormality branch directly models anomaly behaviors provided by natural language. As the two views capture complementary information, we naturally think of the compatibility of them for achieving better performance. Therefore, a cross-view contrastive learning (CCL) s proposed to regularize the intra-view training with additional reference information from the other complementary view, and a cross-view mutual interaction (CMI) strategy further promotes the mutual exploration of useful knowledge from each branch.