Adapting Domain-Aware Knowledge to Vision-Language Model for Zero-Shot Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot anomaly detection (ZSAD) faces challenges due to anomaly rarity, diversity, and domain-specific manifestations.
• Vision-language models (VLMs) show promise for ZSAD but struggle with domain adaptation due to limited domain-specific knowledge.

Purpose of the Study:

• To propose Domain Adaptation CLIP (DA-CLIP), a novel approach for enhancing ZSAD by adapting domain-aware knowledge to VLMs.
• To improve the generalization capabilities of VLMs for detecting anomalies in unseen domains.

Main Methods:

• DA-CLIP employs a Domain-Aware Knowledge Adaptation (DAKA) strategy with specialized experts for target domains.
• Learnable domain-aware prompts are injected into both CLIP encoders and DAKA modules for dual-pathway learning.
• This enables dynamic selection and combination of experts tailored to anomaly characteristics and domain-specific features.

Main Results:

• DA-CLIP consistently outperforms state-of-the-art methods on benchmark datasets across industrial and medical domains.
• Significant improvements were observed in both image-level and pixel-level anomaly detection tasks.
• The dual-pathway learning effectively captures domain-specific features for better adaptation.

Conclusions:

• DA-CLIP offers a robust solution for domain adaptation in zero-shot anomaly detection.
• The proposed DAKA strategy and domain-aware prompts enhance VLM performance on diverse anomaly detection tasks.
• This approach advances the field of ZSAD by improving generalization across different domains.