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.

Purpose of the Study:

• To introduce XMatchAD, a novel UAD framework.
• To address limitations in detecting subtle anomalies and localizing them with sharp boundaries.

Main Methods:

• Reinterpreting UAD as pseudo cross-modal matching, treating input and reconstructed images as complementary modalities.
• Utilizing a pre-trained feature extractor for discriminative representations.
• Employing an attention-guided cross-modal matching mechanism for inter-modal pattern matching and feature refinement.
• Integrating an adaptive frequency-aware fusion module for sharp boundary delineation.

Main Results:

• Enhanced sensitivity to anomalies of diverse shapes and subtle deviations.
• Significant improvements in the precision of anomaly detection and localization.
• Superior performance on MVTec-AD, VisA, and MPDD benchmarks, outperforming state-of-the-art methods.

Conclusions:

• XMatchAD offers a robust solution for multi-class anomaly detection and localization.
• The cross-modal matching perspective effectively enhances anomaly detection capabilities.
• The method demonstrates state-of-the-art performance across multiple benchmarks.