Training and Validation of Automatic Target Recognition Systems using Generative Adversarial Networks

Abstract

This research provides advances aiming to improve the adaptability and usability of Automatic Target Recognition (ATR) algorithms in new environments. We propose to use a Generative Adversarial Networks (GAN) based approach to add simulated contacts into real sidescan sonar images. Our results show that the GAN approach is able to create realistic contacts. We carried out a visual experiment to validate that a trained operator was unable to distinguish real objects from simulated objects. In addition, we demonstrate that an ATR tuned using simulated objects, generated by the GAN, achieves a comparable performance to an ATR tuned using real data.