Resolution enhancement of cementitious microstructure images and phases quantification using deep learning

This paper introduces a deep learning method for reconstructing and segmenting cement paste microstructural images in Backscattered Electron (BSE) mode. Using a dataset of 2400 full-scale BSE images (258,818 after cropping), the Local Implicit Image Function (LIIF) super-resolution network enhanced image resolution up to 30 × , reducing noise and improving microstructural detail. This improved the accuracy of segmentation, which was carried out using the SegFormer network. SegFormer outperformed traditional methods like U-Net and DeepLabv3 + in terms of mean Intersection over Union (MIoU) and detail preservation. The segmentation results quantified porosity, unhydrated cement, and hydration products, aligning closely with Mercury Intrusion Porosimetry (MIP) and Quantitative X-Ray Diffraction (QXRD) measurements. The hydration degree from BSE segmentation also matched well with Thermogravimetric Analysis (TGA) calculations. The method showed strong generalization capabilities, effectively handling diverse BSE images. This study confirms the method's reliability and accuracy for concrete microstructure analysis. • Introduced a deep learning method for enhancing image resolution and segmenting concrete BSE images. • Utilized a large dataset of 2880 full-scale BSE images, expanding to 258,818 images after cropping for detailed analysis. • Segmentation accurately quantified porosity, unhydrated cement, and hydration products, aligning with experimental measurements. • Strong generalization in handling diverse BSE images, confirming the method's reliability.