SAR image synthesis with GAN and continuous aspect angle and class constraints

Abstract:
Target classification generally requires large databases, especially for deep learning methods. However, it is not always possible to have access to a database of sufficient size for certain imaging modalities. For example, in synthetic aperture radar (SAR) imaging only limited incidence angles and aspect angles can be available. Unfortunately, to overcome this problem, most of the classical data augmentation methods are inappropriate for SAR data. Thus, in a previous work, we evaluated conditional Generative Adversarial Networks to generate synthetic SAR images at given aspect angles and for specific target classes. Among the various models evaluated the so-called StyleGAN2-ada, slightly modified to take into account the specificity of SAR images, appear to be the most efficient model. However, we observed that some of the generated images had wrong aspect angles. In this contribution we propose to correct this problem by adding a regularization term recently proposed in a model called Generator Regularized-cGAN. Our experiments show that this modification strongly reduce the problem.