Transfer and multitask learning method for buried wire detection via GPR

Abstract

Ground Penetrating Radar (GPR) is a widely used tool for buried target detection, classification and identification applications. Classification of targets in same shape and different materials is a particularly difficult problem, especially for targets with similar electrical conductivity. In such a classification problem, deep learning is used because it is very difficult to extract features. Deep learning is a method that has demonstrated state-of-art performance over the last five years thanks to automatic learning of the features and classifier. The purpose of this research is to correctly classify buried targets with high performance. For this purpose, firstly, data, composed of targets buried in 3 types of soil, is produced. Then, deep learning, transfer learning and multitask learning methods were proposed for detection and identification of copper and plastic wires in this dataset. Transfer learning is the process of transferring information acquired from a deep learning model previously trained with a large number of data to the current model. For transfer learning, the first six layer of the VGGNet architecture, which yields very successful results in the deep learning literature, is used for GPR classification. Multitask learning is an approach that improves learning of a task, by using training information from other related tasks. For multitask learning, a deep learning architecture that detects the type of soil (dry, wet, damp) was developed. It has been shown that the proposed methods improve target detection.