Object Identification form GPR Images by Deep Learning

We have developed an identification method of an underground object form a ground penetrating radar (GPR) image by the deep neural network. In this study, in order to automatically detect an underground object from the GPR image by the DNN, we have generated several hundred thousand GPR images for training the DNN using a fast finite-difference-time-domain (FDTD) simulation with graphics processing units (GPUs). Furthermore, Characteristics of underground objects are extracted and learned from generated GPR images by a 9-layers convolutional neural network (CNN). It is shown that the CNN can identify six materials with roughly 80% accuracy in inhomogeneous underground media.

I. Introduction

Recently, the deterioration of social infrastructures such as tunnels, bridges, and roads becomes a severe social problem. It is required to rapidly and adequately detect for abnormal parts of social infrastructures [1], [2]. The ground penetrating radar (GPR) is efficient for inspection of various social infrastructures. However, it is difficult to detect an underground object from a GPR image. Automatic detection systems of the underground object based on a machine learning and an artificial intelligence technique are required. Although detection methods by several neural networks have been investigated [3], [4], these previous methods cannot identify more information such as the material and the size of the underground object.