(A) study on SAR image generation using GAN-based multi-task learning

Abstract

Generative Adversarial Networks (GANs) have become most commonly used deep generative models that desire to learn a target data distribution. Although they are successfully applied to diverse fields, training GANs on synthetic aperture radar (SAR) data is a challenging task mostly due to speckle noise. On the other hands, in a learning perspective of human’s perception, it is natural to learn a task by using diverse information from multiple sources. However, in previous GAN works on SAR image generation, the pose angle information of the targets has not been taken into consideration for GANs. In this thesis, we propose a novel GAN-based multi-task learning (MTL) for SAR image generation with several GAN’s training techniques, called PeaceGAN that uses both pose angle and target class information, as concatenated to a noise vector of its generator’s input part and that has two additional structures with a pose estimator and an auxiliary classifier at the side of its discriminator to combine the pose and class information more efficiently. The PeaceGAN can produce SAR images with intended target classes at desired pose angles so it can be utilized for SAR data augmentation as well. In addition, the PeaceGAN is jointly learned in an end-to-end manner as MTL with pose angle and target class to make it enhanced both the diversity and the quality of generated SAR images. Lastly, we also propose two indirect evaluation methods for the quality of generated SAR images. By those methods, we can verify both the generator’s ability and the quality of generated SAR images indirectly.