Segmentation and analysis of cement particles in cement paste with deep learning

Abstract

Scanning electron microscopy (SEM) is a widely used method for the analysis of concrete micro structure. To quantitatively analyze the SEM images with high efficiency and accuracy, an automatic segmentation framework is proposed in this paper. The deep segmentation algorithm is purposely optimized from PointRend based on the characteristic of SEM images to improve prediction accuracy, especially the performance around boundaries. Moreover, the SEM images can be segmented without additional treatment. Cement paste samples with 0.2 and 0.4 water-to-cement ratios are prepared and cured for 1, 3, 7, 14, and 28 days. Totally SEM images with 2267 labeled cement particles are included to build the dataset. From the results of intersection over union and pixel accuracy, the proposed algorithm outperforms the trainable waikato environment for knowledge analysis (WEKA) segmentation, Fully Convolutional Networks (FCN), and the original PointRend method. The segmentation results are used to calculate the hydration degree of two cement paste samples. Good agreement is obtained with the hydration degree calculated by using nonevaporable water in the samples for the 5 curing durations. At last, the shape of the cement particles is analyzed. Irregularity and roundness of the cement particles do not change significantly with an increase in curing duration.

More information

Identifiers

ORCID for Ye Li: orcid.org/0000-0002-2379-5497

Catalogue record

Export record

Altmetrics