Support of Data Augmentation with GAN on Faster R-CNN Based Buried Target Detection

Ground penetrating radar (GPR) is a common technique used for target through-the-obstacle imaging. However, it is quite difficult to detect the presence of a buried object in the obtained images in GPR, and the reason for this is the presence of clutter. Many methods have been proposed for clutter removal. In the classical approach, it is aimed to suppress the clutter and increase the target signal strength and then target detection algorithms used. Nowadays, region based convolutional neural network (RCNN) algorithm is adapted for the buried target detection which does not need any clutter removal algorithms. The accuracy performance of these algorithms is related to the size of the training data set. In this study, a realistic simulation data set was created with the gprMax and this data set is augmented by using the generative adversarial network (GAN) algorithm. The Faster R-CNN structure has been trained with an enhanced simulated data set and tested both in simulated and real data sets. In the results obtained, it has been observed that increasing the data with GAN achieves a performance increase up to %5-54 and it is shown that target detection can be performed on real data with a realistic simulated data set for GPR.