CLIP-Based Adaptive Weighted Anomaly Detection with Cross-Layer Semantic Decoupling

Image anomaly detection is an important task in industrial quality inspection and has also been investigated on medical imaging benchmarks. However, many existing methods require normal samples from the target domain, which may be difficult to obtain in practical applications. Recent zero-shot anomaly detection methods based on vision–language models have shown promising performance. However, cross-layer semantic redundancy in hierarchical visual–text alignment and the aggregation of sparse anomaly information for image-level prediction remain challenges. To address these challenges, we propose CLIP-based adaptive weighted anomaly detection (AWAD), a cross-domain zero-shot anomaly detection framework. AWAD introduces a cross-layer semantic decoupling (CLSD) module to construct stage-specific textual representations and reduce excessive cross-layer semantic similarity. An adaptive anomaly guidance (AAG) module is further developed to aggregate spatial features according to cross-modal anomaly responses for image-level discrimination. Experiments on 12 public industrial and medical imaging benchmarks and a supplementary Toothbrush Bristle Defect Dataset (TBDD) evaluate the proposed framework under cross-domain zero-shot settings. AWAD achieves a pixel-level AUROC of 95.8% and a PRO of 89.6% on VisA, together with a pixel-level AUROC of 91.6% on MVTec AD. The results show promising pixel-level anomaly localization performance on several benchmarks, while image-level detection performance varies across datasets. The medical imaging experiments serve as cross-domain transfer benchmarks and do not constitute clinical validation.