SRGAN-Based Joint Super-Resolution and Denoising for Mitigating Geometric and Topological Biases in Fine-Grained Electron Backscatter Diffraction Images

Accurate microstructural characterization for materials with grain sizes ranging from tens to hundreds of nanometers is often constrained by insufficient resolution and noise, leading to distorted statistics. In this study, we constructed a simulated dataset incorporating noise induced by effective resolution based on experimental electron backscatter diffraction (EBSD) characteristics and proposed an SRGAN-based method for simultaneous super-resolution and denoising. Compared with conventional interpolation-based denoising methods widely adopted in EBSD experiments, the proposed framework effectively alleviates the underestimation of grain size and grain boundary number distributions, and more faithfully recovers microstructural geometry and topology. Quantitatively, conventional methods cause an average of approximately 16.03% of the grains to shift toward smaller size bins, whereas the SRGAN-based approach reduces this proportion to about 3.35%. In addition, the SRGAN-based method shows a 2.58% improvement in the accuracy of grain boundary counts compared with conventional method, effectively mitigating statistical distribution distortion. Moreover, although trained exclusively on simulated images, the network performs effectively on experimental images, demonstrating practical applicability for microstructural analysis.