CHSA-CLIP: zero-shot anomaly detection via cross-modal hierarchical semantic alignment

Zero-shot anomaly detection (ZSAD) aims to identify anomalous regions from previously unseen categories and plays an important role in industrial inspection and medical image analysis. However, existing ZSAD methods commonly suffer from cross-modal hierarchical semantic misalignment, which is mainly manifested as insufficient spatial correspondence between fine-grained textual descriptions and multi-level image patches, as well as inherent semantic discrepancies between hierarchical textual prompts and global CLS tokens. This dual misalignment adversely affects the accuracy of anomaly detection. To address these issues, we propose a cross-modal hierarchical semantic alignment-based ZSAD method, termed CHSA-CLIP. Specifically, CHSA-CLIP integrates a task-aware prompt matrix with filtered fine-grained textual prompts to dynamically generate hierarchical textual prompts, enabling collaborative perception of global semantic context and local anomalous features, thereby improving the precision of cross-modal hierarchical semantic alignment. In addition, a multi-level anomaly aggregation mechanism based on K-Means clustering is introduced to mine key anomalous regions from visual features, and a global-local CLS fusion strategy is employed to effectively enhance the model’s sensitivity to subtle anomaly signals. Experimental results on 15 industrial and medical benchmark datasets demonstrate that CHSA-CLIP achieves superior performance and strong cross-domain generalization capability on both image-level anomaly classification (AC) and pixel-level anomaly segmentation (AS) tasks.