SRGAN-Based Joint Super-Resolution and Denoising for Mitigating Geometric and Topological Biases in Fine-Grained...

Area of Science:

• Materials Science
• Computational Materials Science
• Image Analysis

Background:

• Microstructural characterization of nanomaterials is limited by low resolution and noise in techniques like electron backscatter diffraction (EBSD).
• Existing denoising methods can distort statistical distributions of grain size and grain boundary counts.
• Accurate characterization is crucial for understanding material properties.

Purpose of the Study:

• To develop a method for simultaneous super-resolution and denoising of EBSD data.
• To improve the accuracy of microstructural statistics, including grain size and grain boundary number distributions.
• To provide a practical tool for analyzing experimental microstructural images.

Main Methods:

• Construction of a simulated EBSD dataset with noise mimicking experimental conditions.
• Implementation of a Super-Resolution Generative Adversarial Network (SRGAN) for image enhancement.
• Comparison of the SRGAN-based method with conventional interpolation-based denoising techniques.

Main Results:

• The SRGAN-based method significantly reduces the underestimation of grain size distributions compared to conventional methods (3.35% vs. 16.03% shift to smaller bins).
• It improves the accuracy of grain boundary counts by 2.58%.
• The network demonstrates effective performance on experimental images despite being trained on simulated data.

Conclusions:

• The proposed SRGAN-based framework offers superior microstructural characterization for nanomaterials.
• It accurately recovers microstructural geometry and topology, mitigating statistical distortions.
• The method shows practical applicability for real-world materials analysis.