Generative adversarial networks with denoising penalty and sample augmentation

For the original generative adversarial networks (GANs) model, there are three problems that (1) the generator is not robust to the input random noise; (2) the discriminating ability of discriminator gradually reduces in the later stage of training; and (3) it is difficult to reach at the theoretical Nash equilibrium point in the process of training. To solve the above problems, in this paper, a GANs model with denoising penalty and sample augmentation is proposed. In this model, a denoising constraint is firstly designed as the penalty term of the generator, which minimizes the F-norm between the input noise and the encoding of the image generated by the corresponding perturbed noise, respectively. The generator is forced to learn more robust invariant characteristics. Secondly, we put forward a sample augmentation discriminator to improve the ability of discriminator, which is trained by mixing the generated and real images as training samples. Thirdly, in order to achieve the theoretical optimization as far as possible, our model combines denoising penalty and sample augmentation discriminator. Then, denoising penalty and sample augmentation discriminator are applied to five different GANs models whose loss functions include the original GANs, Hinge and least squares loss. Finally, experimental results on the LSUN and CelebA datasets show that our proposed method can help the baseline models improve the quality of generated images.