Deep learning methods for phase segmentation in backscattered electron images of cement paste and SCM-blended systems

Quantitative microstructure analysis of supplementary cementitious material (SCM) - blended paste through backscattered electron (BSE) imaging has long been intractable owing to the sophisticated micro phase distribution and overlapped greyscale histogram. This study explores the use of Convolutional Neural Network (CNN) based supervised semantic segmentation methods for quantifying phase assemblage in BSE images of blank cement paste and SCM-blended cement pastes. Four types of SCMs, namely limestone, slag, quartz and metakaolin, were separately blended with WPC and OPC paste for analysis. U-Net architecture with and without ResNet backbones were trained to perform pixel-level image segmentation of anhydrous cement and SCM particles. The results indicate that deep learning models can robustly segment anhydrous cement particles from BSE images and achieve same level of precision as QXRD. For limestone, quartz and slag, deep learning models show strong potential for semi-quantitative segmentation. While metakaolin cannot be reliably segmented based solely on graphic information.