Recognizing of the A-Scan Image of a Buried Object Using a Deep Network

There are various techniques for detecting buried objects underground, and ground penetrating radar (GPR) is a common non-destructive practical method used in the detection process. GPR uses the difference in dielectric properties between the buried object and its surroundings to locate the object. The basic operation of a GPR device is to transmit electromagnetic waves and receive the returned signal. Processing and interpretation of the received signal using GPR equipment is still a challenging task in terms of time consumption, cost, and effective accuracy. This research aims to study the performance of various deep neural network architectures for classifying different buried objects. This work consists of two parts, the first part is to use a simulation system (gprMax) to obtain a reflection signal. The second part is to process the input images using a deep learning network for analysis in four categories. The advantage of this procedure using a deep learning network is to improve efficiency and reduce the time required to locate and identify a buried object. The training progress results for all networks show that four networks (EfficientNet-b0, ResNet-18, Inception-ResNet-v2, and ResNet-101) have the highest validation accuracy of 87.5%. In comparison to these networks, ResNet-18 is more efficient as less time elapsed compared to other networks. The performance and validation accuracy can be improved by increasing the number of the input data and adjusted hyperparameters values of the training options.