Soft-Margin Ellipsoid Generative Adversarial Networks Based on Gradient Regularization

This paper presents the results on the designing and applying of a kind of generative adversarial networks(GANs) with soft-margin Ellipsoid GAN which are used in a variety of applications including computer vision and image processing. However, GANs are often subjected to problems linked to instability and non-convergence in training process. Defined an integral probability metric (IPM) on hypersphere, Sphere Generative Adversarial Network(Sphere GAN) makes discriminator have the characteristics of Lipschitz continuity to ensure the convergence in the training process. Thus, developed from Sphere GAN, we propose a new GAN model called Ellipsoid Generative Adversarial Networks(Ellipsoid GAN), in which IPM defined on hypersphere is generalized to hyperellipsoid. The hyperellipsoid is realized to relax the upper bound of IPM by extending measurable functions space, and induce a more sensitive Wasserstein distance to improve the quality of generated samples. In addition, in order to improve the stability of discriminator, the idea of soft-margin is introduced into the designed Ellipsoid GAN to prevent the discriminator from gradient vanishing and exploding on the classification boundary. Comparison experimental results on CIFAR10 and LSUN datasets show that the designed soft-margin Ellipsoid GAN is superior to Sphere GAN with better quality of generated samples.