Generative adversarial networks with data augmentation and multiple penalty areas for image synthesis

Abstract

The quality of generated images is one of the significant criteria for Generative Adversarial Networks (GANs) evaluation in image synthesis research. Previous researches proposed a great many tricks to modify the model structure or loss functions. However, seldom of them consider the effect of combination of data augmentation and multiple penalty areas on image quality improvement. This research introduces a GAN architecture based on data augmentation, in order to make the model fulfill 1-Lipschitz constraints, it proposes to consider these additional data included into the penalty areas which can improve ability of discriminator and generator. With the help of these techniques, compared with previous model Deep Convolutional GAN (DCGAN) and Wasserstein GAN with gradient penalty (WGAN-GP), the model proposed in this research can get lower Frechet Inception Distance score (FID) score 2.973 and 2.941 on celebA and LSUN towers at 64×64 resolution respectively which proves that this model can produce high visual quality results.