XMatchAD: A Cross-Modal Matching Perspective on Reconstruction-Based Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Reconstruction-based Unsupervised Anomaly Detection (UAD) methods model image discrepancies but struggle with subtle anomalies and blurred boundaries.
• Existing methods are limited in complex multi-class scenarios due to challenges in capturing fine details and precise localization.

Purpose of the Study:

• To introduce XMatchAD, a novel UAD framework that reinterprets the task using a pseudo cross-modal matching perspective.
• To enhance sensitivity to subtle anomalies and improve precision in anomaly detection and localization.

Main Methods:

• Utilized a pre-trained feature extractor for discriminative representation encoding.
• Introduced an attention-guided cross-modal matching mechanism to align inter-modal anomaly patterns and refine features.
• Developed an adaptive frequency-aware fusion module to sharpen anomaly boundaries by integrating high-frequency components.

Main Results:

• XMatchAD demonstrates superior performance in multi-class anomaly detection and localization across MVTec-AD, VisA, and MPDD benchmarks.
• The proposed method consistently outperforms state-of-the-art techniques.
• Achieved enhanced sensitivity to diverse anomaly shapes and subtle deviations.

Conclusions:

• XMatchAD effectively addresses limitations of traditional reconstruction-based UAD methods.
• The pseudo cross-modal matching framework offers a promising direction for advanced anomaly detection.
• The method provides significant improvements in both detection accuracy and localization precision.