Abstract
Ground Penetrating Radar (GPR) has been widely used to detect highway pavement structures. In recent years, deep learning techniques have achieved significant success in image recognition, which is potentially relevant for interpreting ground-penetrating radar data. This is because the various types of damage develop at different levels and in different quantities. So the number of datasets of various types of road injuries is not balanced. This leads to poor accuracy of deep learning for injury classification. And the cost of collecting a large amount of data in the field is higher. The aim of this paper is to improve classification accuracy at a lower cost relative to field collection, we propose a damage data expansion method based on generative adversarial network, which consists of encoder and a generative adversarial network. We have made a number of improvements to the generator and discriminator, as well as to the newly added encoder. All of these improvements have improved the generation results in terms of metrics. So that the network can stably generate damage samples with a small number of samples to improve the classification network's accuracy. The effect on accuracy by varying the proportions of different kinds of samples and traditional expansion methods is also explored. The improvement of the classification network accuracy and FlD metrics illustrates the better performance of the proposed method.