From fragmented to flawless: A large-scale synthetic micrograph library for benchmarking microstructural image...

The accurate extraction of grain boundaries from metallographic images is a prerequisite for quantitative microstructural analysis, but it is often hindered by experimental artifacts such as noise, etching pits, scratches, and incomplete boundaries. Manual correction of these defects is a significant bottleneck, being both time-consuming and subjective. This article presents a large-scale, synthetic dataset of 512×512 pixel images designed to train and benchmark deep learning models for two critical automated image processing tasks: noise cleaning and grain boundary reconstruction. The dataset was generated by first creating clean, single-phase polycrystalline micrographs using the polySim library. These base images were then programmatically degraded by adding a combination of real and artificial noise and by introducing discontinuities into the grain boundary network. The final dataset comprises 14,999 paired PNG images, consisting of 5999 pairs for noise removal (degraded vs. clean ground truth) and 9000 pairs for grain boundary reconstruction (fragmented vs. connected networks), enabling the development of robust models for automated, objective, and reproducible microstructural image analysis.