Transfer learning approach with synthetic data for high-accuracy segmentation of equiaxed grain boundaries

Abstract

Accurate segmentation of microstructural images plays a critical role in quantitative materials analysis. However, it remains a challenging task due to the scarcity of annotated data and the inherent diversity of microstructures. In this study, we propose a novel transfer learning approach utilizing Voronoi diagrams to enhance the extraction accuracy of equiaxed grain boundaries in α-phase pure titanium. We evaluate the effects of systematically controlled synthetic image generation conditions on segmentation performance. These conditions include seed point placement, grayscale value assignment, and the addition of Gaussian noise. The results demonstrate that introducing moderate Gaussian noise clearly improves extraction accuracy, achieving a higher Intersection over Union (IoU) score than models pre-trained on ImageNet. Furthermore, the model pre-trained on Voronoi diagrams is shown to capture microstructure-specific features, such as triple junctions, thereby enhancing grain boundary extraction even in low-contrast regions. This study provides a foundation for pre-trained models tailored to microstructural analysis and contributes to the advancement of deep learning applications in materials science. The pretrained models and code are publicly available on GitHub.

Dataset

dataset

We used nine SEM-BSE images (2,560 × 1,536 pixels) of α-phase equiaxed grain microstructure in commercially pure titanium (99.995% purity), with manually annotated ground truth labels. Samples were mechanically and electropolished.

Generation of synthetic data (Voronoi diagrams)

Synthetic Data

Results

results

While conventional models struggled to extract low-contrast grain boundaries, which are difficult to distinguish visually, our model successfully identified them. This success is likely due to the inherent triple junction features of Voronoi diagrams. The model learned these features during training, which enabled accurate grain boundary extraction even in challenging low-contrast regions.

BibTeX

@unpublished{ozaki2025transfer,
    title={Transfer learning approach with synthetic data for high-accuracy segmentation of equiaxed grain boundaries},
    author={Ozaki, Koichi and Nohira, Naoki and Tahara, Masaki and Kumazawa, Itsuo and Hosoda, Hideki},
    year={2025},
    note={under review}
}