AnoCLIP: Text-Guided Zero-shot Anomaly Localization via Self-Supervised Adaptation

This study aims to explore the intrinsic potential of CLIP (Contrastive Language-Image Pretraining) for zero-shot anomaly localization, a challenging task that involves localizing anomalous regions without normal samples. To this end, we propose AnoCLIP, an efficient and training-free framework for zero-shot anomaly localization. By employing visual attention modifications and comprehensive text prompts, we effectively harness CLIP’s natural capability for zero-shot anomaly localization. To further refine the results of AnoCLIP in visual anomaly localization, we propose a novel test-time adaptation method, Text-Guided Anomaly Self-Supervised Adaptation (TGASA). Our approach constructs a pair of balanced self-supervised tasks using text-prompted pseudo-labels and noise-corrupted visual tokens, respectively, to optimize a lightweight adapter during inference. The proposed approach is time-efficient and significantly enhances the anomaly localization performance of AnoCLIP. Through comprehensive experiments, we demonstrate the efficiency and effectiveness of AnoCLIP and TGASA in zero-shot anomaly localization on various datasets.