Related Experiment Videos
UAD: A Unified Model for Zero-shot Anomaly Detection
Summary
This study introduces UAD, a unified framework for zero-shot anomaly detection (ZSAD). UAD effectively identifies unseen anomalies by integrating semantic context and structural information, improving cross-domain generalization.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Zero-shot anomaly detection (ZSAD) is vital for privacy-sensitive domains but faces challenges in cross-domain generalization.
- Existing methods often fail due to domain-specific coupling or fragmented designs neglecting semantic and structural aspects.
Purpose of the Study:
- To propose UAD, a unified framework for holistic zero-shot anomaly detection.
- To enhance generalization by jointly modeling semantic regularity and anomaly-aware representations.
Main Methods:
- UAD aligns multi-level semantic understanding with fine-grained structural cues.
- Image representations are organized into semantic contexts to detect deviations from local and high-level patterns.
- Cross-domain robustness is improved via prompt concatenation and intensity-guided anomaly synthesis.
Main Results:
- UAD demonstrates superior zero-shot performance across 17 diverse real-world anomaly detection datasets.
- The framework effectively detects and segments anomalies in defect inspection and medical imaging domains.
- Significant improvements in generalization to unseen anomaly types and domains were observed.
Conclusions:
- UAD offers a holistic approach to ZSAD, outperforming existing methods.
- The framework's ability to integrate semantic and structural information is key to its success.
- UAD advances the state-of-the-art in cross-domain anomaly detection.
