Guided by visual-textual feature and wavelet similarity for zero-shot anomaly detection

Zero-shot anomaly detection (ZSAD) is of utmost importance in industrial domains. It aims to detect anomalies in various fields without training samples, yet faces challenges due to significant differences in anomalies. Although large pre-trained vision-language models like CLIP are powerful in zero-shot recognition, they falter in ZSAD as they emphasize foreground object semantics over anomaly/normality. Traditional CLIP-based ZSAD methods have flaws like poor text prompt generalization and simplistic feature fusion. This paper proposes an innovative framework. It features vision-guided text prompt design, enhanced feature integration through text space refinement and local visual attention optimization, and a loss function incorporating frequency information. Experiments on MVTec AD and VisA datasets, evaluated with image-AUROC, AP, pixel-AUPRO and PRO, demonstrate VWCLIP's outstanding performance and fine-grained anomaly recognition.Ablation study and hyperparameter analysis validate the framework's effectiveness, offering novel ideas and avenues for future ZSAD research.