Data augmentation in feature-space with generative adversarial networks, applied to GPR-based buried threat detection
In this work we consider the problem of developing algorithms for automatic buried threat detection (BTD) in ground penetrating radar (GPR) data. Many such algorithms are supervised, and perform best when they can be trained on large quantities of labeled threat and non-threat GPR data, respectively. Unfortunately, such data is costly to collect, and therefore relatively scarce. One approach to mitigate this problem is data augmentation, in which novel training data is created by applying transformations to existing data. Prior work has shown that augmentation can indeed improve the training of GPR-based BTD algorithms. In this work, we explore the use of Generative Adversarial Networks (GANs) for data augmentation. GANs can be trained to generate novel, but highly realistic, data after training on a real-world dataset. GANs have yielded impressive results on many types of data, but they are notoriously difficult to train. In this work, we propose an approach, entitled featureGAN, that mitigates some of the challenges training GANs. We show that augmentation using featureGAN yields improved detection performance, and yields better performance than some naïve alternative augmentation strategies. We also propose a metric for quantifying the success of GAN training, called the q-metric, which was crucial to achieving good results.
