Answers are grounded in this paper and cite their sources.
Purpose Objective and rapid evaluation of microscopic images is critical in materials science. This is because the mechanical, thermal and electronic properties are directly related to microstructure. Especially in recent years, with the increase in deep learning techniques, precise and rapid microscopic characterization has become possible. However, deep learning-based segmentation methods typically require large annotated datasets, which are difficult to obtain in metallography. This article proposes a scenario for a modified Res-UNet architecture optimized under highly constrained data conditions for semantic segmentation of B4C-reinforced aluminium metal matrix composites using the Taguchi. Design/methodology/approach A total of 90 image-mask pairs were collected, with 45 reserved exclusively for testing. To systematically investigate the effects of training data size, batch size and learning rate on segmentation performance. Using Taguchi experimental design and analysis of variance (ANOVA), the impacts of dataset size, batch size and learning rate were methodically examined. Findings The findings indicated that dataset size and batch size considerably affect segmentation performance, although learning rate has a comparatively negligible impact. The optimized model attained an F1-score of 0.91 and sustained satisfactory performance even when trained with merely 8–10 training images. These findings also indicate that dependable and quantitative microstructural analysis can be accomplished despite significant data constraints. Originality/value This study's novelty resides in tackling the limited-data barrier in metallographic image segmentation and offering a systematic approach to enhance semantic segmentation performance under restricted annotation situations.
No takes yet. Share an insight, caveat, or question.
Kalkan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: