A study on high efficiency phase organization segmentation model for SEM images based on deep learning

Abstract

Quantitative characterization and analysis of material microstructures play a crucial role in understanding material properties. Traditional U-Net architectures, however, are often vulnerable to noise and prone to overfitting when applied to small datasets in material scanning electron microscopy (SEM) image segmentation. In this work, we propose a novel deep learning model, DNU-Net, specifically designed for SEM image analysis. By integrating a dedicated denoising layer and Dropout regularization into the U-Net framework, the model achieves efficient, robust, and generalizable microstructural phase segmentation even with limited training data. The model is applied to Ti-6Al-4V (TC4) titanium alloy, a material with widely used engineering applications and a complex phase structure. DNU-Net successfully identifies β-phase features before and after annealing with high precision. The mean intersection over union (IoU) and pixel accuracy (PA) of the model reach 90.56% and 98.46%, respectively, demonstrating excellent segmentation accuracy and pixel-level consistency. Compared with conventional binary digital image processing methods, DNU-Net provides more efficient and precise phase segmentation, exhibiting lower sensitivity to noise and greater robustness. Even under challenging illumination conditions and noise perturbations, the quantitative analysis error of the target phase remains below 5%, demonstrating the robust performance of the model. The proposed DNU-Net framework enables precise quantitative analysis of complex microstructures and offers a novel deep learning-based approach for material characterization. Its methodology is expected to be transferable across diverse alloy and material systems, supporting the advancement of artificial intelligence-driven materials design.