Cross-Attention Regression Flow for Defect Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Defect detection in industrial settings is vital but challenged by rare anomalies and poor performance on small defects.
• Existing methods struggle to capture subtle anomalies, especially small-sized defects, due to feature representation limitations.

Purpose of the Study:

• To develop an advanced defect detection framework capable of accurately identifying small-sized defects.
• To improve the modeling of normal visual patterns for robust outlier separation in image data.

Main Methods:

• Proposed a Cross-Attention Regression Flow (CARF) framework.
• Introduced an interactive cross-attention pattern flow module to preserve multi-layer scale information.
• Implemented a layer-conditional autoregression module to handle complex multi-layer feature distributions.

Main Results:

• CARF effectively models normal visual patterns in a latent space for outlier detection.
• The framework demonstrates superior performance in detecting small-size defects compared to state-of-the-art methods.
• Experiments on multiple datasets validate CARF's effectiveness, especially for challenging small defect scenarios.

Conclusions:

• The CARF framework offers a significant advancement in industrial defect detection, particularly for small and hard-to-detect anomalies.
• The proposed cross-attention and autoregression modules enhance the ability to characterize normal patterns and detect subtle defects.