FocusPatch AD: Few-Shot Multi-Class Anomaly Detection With Unified Keywords Patch Prompts.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Industrial few-shot anomaly detection (FSAD) faces challenges with limited normal samples and the need for separate models per category.
• Existing methods incur high computational and storage costs due to single-category model training.

Purpose of the Study:

• To develop a unified anomaly detection framework for multi-class, few-shot settings.
• To address the limitations of current FSAD methods by reducing computational overhead and improving generalization.

Main Methods:

• Introduced FocusPatch AD, a novel framework leveraging vision-language models for unified FSAD.
• Developed a method to link anomaly keywords to specific image regions, enhancing focus on anomalies and reducing background interference.
• Mitigated false detection issues common in global semantic alignment approaches.

Main Results:

• Achieved significant gains in both image-level and pixel-level anomaly detection on MVTec, VisA, and Real-IAD datasets.
• Demonstrated superior classification and localization performance compared to prevailing anomaly detection methods.
• Validated the framework's excellent generalization and adaptability across diverse categories and domains.

Conclusions:

• FocusPatch AD offers an effective solution for unified few-shot, multi-class anomaly detection.
• The proposed region-focused approach enhances accuracy and efficiency in industrial anomaly detection.
• The framework shows strong potential for real-world applications requiring adaptable and robust anomaly identification.