SCAD: A self-constrained solution to automate context-guided zero-shot image anomaly detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional image anomaly detection (IAD) requires labeled training data, which is costly and time-consuming.
• Zero-shot IAD (ZS-IAD) methods bypass training sets but often rely on external prompts and lack scene adaptation.
• Context-guided ZS-IAD methods leverage scene context but struggle with blind hyperparameter tuning and mixed anomalies.

Purpose of the Study:

• To automate context-guided zero-shot image anomaly detection (ZS-IAD) by addressing limitations of existing methods.
• To introduce a novel Self-Constrained Anomaly Detector (SCAD) for reliable and efficient IAD.
• To eliminate the need for external prompts and manual hyperparameter tuning in ZS-IAD.

Main Methods:

• Developed a self-constrained mechanism for automatic hyperparameter value determination.
• Designed an online self-constrained sampler with an efficient stopping point to reduce computational cost.
• Implemented self-constrained normality refinement strategies to manage anomalies and adjust thresholds.

Main Results:

• SCAD automates hyperparameter selection, a novel contribution to IAD.
• The method achieves performance comparable to classic IAD and ZS-IAD methods using hindsight knowledge.
• Demonstrated significant reduction in computational cost through an optimized sampling process.

Conclusions:

• SCAD offers a robust and automated solution for context-guided ZS-IAD.
• The approach enhances reliability by addressing hyperparameter sensitivity and mixed anomalies.
• SCAD represents a significant advancement in efficient and adaptable image anomaly detection.