UTRAD: Anomaly detection and localization with U-Transformer

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection is crucial for industrial defect and medical disease identification.
• Existing methods face challenges with training stability and universal evaluation criteria for feature distributions.

Purpose of the Study:

• To introduce UTRAD, a novel U-Transformer based Anomaly Detection framework.
• To improve the stability, precision, and multi-scale detection capabilities of anomaly detection systems.

Main Methods:

• Representing deep pre-trained features as 'word tokens' processed by transformer-based autoencoders.
• Utilizing reconstruction on informative feature distributions rather than raw images for enhanced detection.
• Implementing a multi-scale pyramidal hierarchy with skip connections for comprehensive anomaly detection.

Main Results:

• Achieved a more stable training process and precise anomaly detection and localization.
• Successfully detected both multi-scale structural and non-structural anomalies.
• Demonstrated superior performance compared to state-of-the-art methods on MVtec AD, Retinal-OCT, Brain-MRI, and Head-CT datasets.

Conclusions:

• UTRAD offers a robust and efficient framework for anomaly detection across various domains.
• The method's multi-scale architecture and feature distribution reconstruction contribute to its high performance.
• UTRAD represents a significant advancement in anomaly detection technology, validated across industrial and medical applications.