PatchCLIP: Enhencing CLIP with Pair-Images for Zero-Shot Anomaly Detection

The vision-language model CLIP has profoundly transformed the filed of zero-shot anomaly detection. Recent studies acquire anomaly maps by aligning images with normal and abonormal prompts. However, CLIP does not effectively interpret text prompts in the context of anomaly detection. As a result, relying solely on text prompts is insufficient, making visual cues are particularly crucial. In this paper, we propose PatchClip for zero-shot anomaly detection, integrating both text-image feature matching and image-image matching. We concurrently process two images, treating each as the visual prompt for the other, and combine their respective text-image matching scores. We conduct extensive experiments on anomaly detection benchmarks, demonstrating superior performance compared to previous methods. The code will be open soon.