S2SWCLIP: semantic-optimized prompts with spatial-wavelet synergy for zero-shot anomaly detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot anomaly detection is vital for privacy-sensitive applications with scarce data.
• Existing visual-language models struggle with semantic overlap and inadequate visual detail representation.
• Alignment deviation between text and visual embeddings hinders performance in current methods.

Purpose of the Study:

• To propose S2SWCLIP, a novel approach for zero-shot anomaly detection.
• To refine prompt learning, enrich visual representations, and optimize cross-modal alignment.
• To address limitations in semantic understanding and visual detail capture in prior models.

Main Methods:

• Integration of semantic-optimized prompts (object-agnostic, contrastive, anomaly-referenced) for sharper semantic boundaries.
• A cross-informative adaptive fusion mechanism for comprehensive semantic information optimization.
• A spatial-to-wavelet transformation module and hierarchical feature fusion for detailed visual representation.
• Entropy-gain similarity for adaptive quantification of information richness and optimized image-text alignment.

Main Results:

• S2SWCLIP demonstrates superior performance compared to existing methods on anomaly detection tasks.
• Experiments conducted on 14 real-world datasets validate the effectiveness of the proposed approach.
• The method successfully delineates semantic boundaries and retains fine-grained image details.

Conclusions:

• S2SWCLIP significantly advances zero-shot anomaly detection capabilities.
• The proposed semantic-optimized prompts and wavelet-spatial synergy effectively improve cross-modal alignment.
• The approach offers a robust solution for privacy-sensitive scenarios requiring accurate anomaly detection.