<p>Industrial anomaly detection is crucial in smart manufacturing, but faces challenges such as diverse types of anomalies, fine granularity of defects, and scarcity of labeled data in practical applications. In recent years, Zero-shot Anomaly Detection (ZSAD), which detects new types of defects without the need for anomaly samples, has gained much attention. However, most of the existing mainstream methods rely on visual language models (VLMs) such as CLIP to reason through graphical semantic alignment, which makes it difficult to capture the fine-grained differences between abnormal and normal. To this end, this paper proposes a purely visual zero-sample detection method based on visual models, which utilizes multilayer visual features from a pre-trained Vision Transformer (ViT) to construct a multiscale feature fusion and patch mutual refinement mechanism (MAD) framework. We design a multi-scale feature fusion module (MSFF) to enhance anomaly recognition at different scales, and introduce a patch mutual refinement module (Patch-MRM) to recognize anomalous regions by local feature similarity without semantic cues. On two industrial datasets, MVTec AD and VisA, MAD achieves 97.7% / 97.1% and 92.4% / 98.7% AUROC at the image level and pixel level, respectively, which significantly outperforms existing zero-shot detection methods. This method demonstrates that pre-trained visual models still have strong migration generalization capabilities for fine-grained industrial anomaly detection even without relying on linguistic cues or semantic supervision.</p>