SEM Image Segmentation Method for Copper Microstructures Based on Enhanced U-Net Modeling

Grain boundary segmentation in scanning electron microscope (SEM) images of pure copper presents substantial challenges for traditional image processing methods, including constrained segmentation precision and difficulties in identifying elongated grain boundaries and intricate topological structures. To overcome these constraints, this research introduces a comprehensive framework that integrates dataset development, advanced data augmentation, and model optimization to achieve precise grain boundary segmentation. This work proposes three principal innovations. First, a meticulously curated small-scale dataset, combined with a sophisticated adaptive data augmentation strategy, addresses data scarcity and ensures high-quality, robust training data. Second, the U-Net model was refined by incorporating a self-attention mechanism, markedly enhancing its capability to capture global contextual information and accurately detect complex grain boundary features. Third, an optimized stratified K-fold cross-validation method was implemented to ensure equitable data partitioning and reduce overfitting, thereby strengthening the model’s generalization capability. Experimental results demonstrate that the proposed framework delivers exceptional performance on the validation dataset, achieving a global accuracy of 0.96, a Dice coefficient of 0.91, and a mean Intersection over Union (mIoU) of 0.85. These metrics underscore significant advancements in grain boundary segmentation precision for polycrystalline metal systems. The framework validates the power of deep learning in microstructural characterization and establishes a reliable computational tool for quantitative metallographic analysis. It is well-positioned to extend to the microstructural analysis of a broad range of heterogeneous materials, enabling deeper insights into microstructure–property relationships in materials engineering.