GAN Using Capsule Network for Discriminator and Generator

In this paper, we propose Capsule GAN, which incorporates the capsule network into the structure of both discriminator and generator of Generative Adversarial Networks (GAN). Many CNN-based GANs have been studied. Among them, Deep Convolutional GAN (DCGAN) has been attracting particular attention. Other examples include convolutional GAN, auxiliary classifier GAN, Wasserstein GAN (WGAN) which uses Wasserstein distance to prevent mode collapse during the learning process, and Wasserstein GAN-gp (WGAN-gp). However, image generation by GAN is not stable and prone to mode collapse. As a result, the quality of the generated images is not satisfactory. It is expected to generate better quality images by incorporating a capsule network, which compensates for the shortcomings of CNN, into the structure of GAN. Therefore, in this paper, we propose two approaches to generate images with better quality by incorporating the capsule network into GAN. The experimental results show that the proposed method is superior to the conventional method.