Estimation of Detailed Information on Internal Defects in Concrete Using Ground Penetrating Radar and Generative Adversarial Networks

Abstract

In this study, we validated the proposed method to estimate the position, angle, and size information of internal defects of concrete from GPR images using pix2pix, a kind of conditional generative adversarial networks. First, in order to obtain GPR images for training, concrete specimens with embedded thin plate-shaped defects at various positions, angles, and sizes were created and GPR tests were performed. Then the training dataset was learned by pix2pix. As a result of validation of the model, the proposed method was able to roughly estimate the defect information. However, as the position of the defect becomes deeper, the response of the reflected wave from the defect decreases, and as a result, it becomes indistinguishable from the scattered wave from the aggregate distributed inside the concrete, and the accuracy decreases.