Enhancing microstructural fidelity in near-field high-energy diffraction microscopy reconstruction with Noise2Void denoising.

The resolution and fidelity of grain reconstruction from near-field high-energy diffraction microscopy (NF-HEDM) are often limited by poor signal-to-noise ratios, especially for fine microstructural features. Traditional denoising methods require clean ground-truth data, which is physically impossible in these experiments. To overcome this, we introduce a self-supervised deep learning network, Noise2Void (N2V), to denoise NF-HEDM images at the detector level prior to segmentation. By avoiding the requirement for clean ground-truth images or paired noisy data sets, which are typically unavailable in experimental settings, the N2V technique presented here is highly suitable for NF-HEDM and potentially transferable to other diffraction-based microstructure imaging techniques compromised by uncorrelated noise, such as micro Laue diffraction and far-field HEDM. The N2V model is trained solely with noisy images and then applied to those same images to obtain denoised data. The resulting cleaner images allow for the application of a lower segmentation threshold, enabling the capture of very weak signals that were previously lost. This yields a quantifiable improvement in reconstruction confidence by ∼9% on average and, validated grain by grain against an electron backscatter diffraction (EBSD) ground truth, recovers more of the true grain population - concentrated in the small grains and fine microstructural features that conventional thresholding discards - without altering the boundaries of grains already detected.