Generative adversarial networks with augmentation and penalty

Abstract

In the original GANs model, there exist two drawbacks for image generation task. On the one hand, a generator cannot deal with a noise that causes an abrupt loss change and inputs of these noises also affect a discriminator, resulting in instability in training process. On the other hand, the discriminator’s discriminating ability is limited in the later stage of training. Eventually, generated images are blurred and targets are incomplete. In order to solve the above problems, a GANs model with hybrid augmented discriminator and fake sample penalty is firstly proposed. In this model, we design a hybrid augmented discriminator. We add real and fake samples into this discriminator. These hybrid samples are conducive to improve the discriminating ability of discriminator. Then, to stabilize the training process and achieve local convergence, we add a penalty to the generator on the basis of designed discriminator, which constrains fake samples with ill condition number. Secondly, we validate the generalization of proposed method on three different loss functions including Hinge, GANs and LSGANs loss. Finally, experimental results show that the proposed method is more effective than baseline models.