Improved Grain Boundary Reconstruction Method Based on Channel Attention Mechanism

Area of Science:

• Materials Science
• Image Analysis
• Computational Materials Science

Background:

• Metal material properties are highly dependent on grain size.
• Metallographic images often present challenges like noise, poor contrast, and incomplete grain boundaries, hindering accurate analysis.
• Precise grain size measurement is critical for predicting material performance.

Purpose of the Study:

• To develop an advanced method for reconstructing incomplete grain boundaries in metallographic images.
• To improve the accuracy of grain size measurement and material performance prediction.
• To enhance the semantic understanding and reconstruction capabilities of generative networks for image analysis.

Main Methods:

• Utilized a generative adversarial network (GAN) as the core architecture.
• Integrated a custom-designed channel attention module within the GAN's generator.
• Incorporated a global context attention mechanism to capture comprehensive image information.
• Optimized the loss function with Focal Loss to mitigate Mode Collapse and improve robustness.

Main Results:

• The proposed method achieved high performance metrics: MIoU of 86.25%, Accuracy of 95.06%, and Precision of 86.54%.
• The improved channel attention module demonstrated superior performance compared to other attention mechanisms.
• The method effectively reconstructed missing grain boundaries and leveraged long-range feature correlations.

Conclusions:

• The developed grain boundary reconstruction method significantly enhances the accuracy and generalization ability of networks.
• This approach provides robust technical support for microstructure characterization and material performance prediction.
• The improved channel attention mechanism offers a promising solution for analyzing challenging metallographic images.