[2O5-GS-13] AI application: Civil Engineering

[2O5-GS-13-04] Improvement of identification performance by effective use of unlabeled radar images for buried objects identification using ground penetrating radar

〇Tomoyuki Kimoto¹, Jun Sonoda² (1. National Institute of Technology, Oita College, 2. National Institute of Technology, Sendai College)

Keywords:Ground penetrating radar, Social infrastructure maintenance, Unsupervised learning, Variational auto encoder: VAE

The our purpose is to develop a system for identify whether it is a risk factor or not from ground penetrating radar images of underground objects. In order to train the CNN whether a radar image is a risk factor, a large number of radar images with risk factor labels required, but in reality, a large number of unlabeled radar images and only a few labeled radar images can be obtained. In this study, we report that the recognition rate can be improved by perform the supervised learning a small number of labeled radar images with MLP after the unsupervised learning of a large number of unlabeled radar images with VAE.