Image-based microstructure classification of mortar and paste using convolutional neural networks and transfer learning

Abstract

The scanning electron microscopy (SEM) is widely applied to analyze the microstructure of concrete. SEM results are generally analyzed by human experts with different levels of expertise, and some tasks are extremely time consuming. In this study, a dataset consisting of 3600 SEM images was first built. Then, a deep-learning framework based on a convolutional neural network (CNN) was implemented for classifying cement paste mixtures with different water-to-cement ratios and different amounts of added silica fume. The accuracy of the classification reaches a high level of 94%. To improve the generality and efficiency of the proposed method, transfer learning technology with three transfer configurations was implemented and tested on a dataset of mortar samples. The result indicated that transfer learning enabled the new model to achieve higher accuracy and generality than training a network with randomly initialized parameters. The model accuracy increases with an increasing number of free convolutional layers, although the training time becomes longer. Finally, the critical features that greatly influence the classification were identified via visualization of the CNN model. Relatively small unhydrated cement particles have higher influence on mixtures with lower water-to-binder ratios, whereas hydration products are more influential in the case of mixtures with higher amounts of water or without silica fume.

More information

Identifiers

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

Catalogue record

Export record

Altmetrics