Automated segmentation and grain detection in resolution-limited images of nanocrystalline alumina
Abstract
Accurate segmentation of scanning electron microscopy (SEM) images is essential for quantitative grain characterization in nanocrystalline alumina, used to obtain derived macro properties of the material. However, the fine grain structure often produces resolution-limited SEM images in which only a subset of grains can be confidently annotated. These target images are also scarce, as their sample-specific preparation and imaging conditions are not readily reproducible. This work proposes a synthetic SEM noise data augmentation applied on clean SEM images using a physics-inspired SEM noise model with evidence-filtering. The proposed method suggests that synthetic SEM noise data augmentation can leverage clearer, fully annotatable SEM images to improve grain extraction quality and measurement reliability, given a low volume of resolution-limited SEM data that is partially annotatable, supporting more informative quantitative microstructural characterization.