Estimating Compressive Strength of Concrete Using Deep Convolutional Neural Networks with Digital Microscope Images

Abstract

Compressive strength is a critical indicator of concrete quality for ensuring the safety of existing concrete structures. As an alternative to existing nondestructive testing methods, image-based concrete compressive strength estimation models using three deep convolutional neural networks (DCNNs), namely AlexNet, GoogLeNet, and ResNet, were developed for this study. Images of the surfaces of specially produced specimens were obtained using a portable digital microscope, after which the samples were subjected to destructive tests to evaluate their compressive strength. The results were used to create a dataset linking the experimentally determined compressive strength with the image data recorded for each. The results of training, validation, and testing showed that DCNN models largely outperformed the recently proposed image processing–based ANN model. Overall, the ResNet-based model exhibited greater compressive strength estimation accuracy than either the AlexNet- or GoogLeNet-based models. These finding indicate that image data obtained using a portable digital microscope contain patterns that can be correlated with the concrete’s compressive strength, enabling the proposed DCNN models to use these patterns to estimate compressive strength. The results of this study demonstrate the applicability of DCNN models using microstructure images as an auxiliary method for the nondestructive evaluation of concrete compressive strength.