Generative Adversarial Network Architectures For Image Synthesis Using Capsule Networks
share
In this paper, we propose Generative Adversarial Network (GAN) architectures using dynamically-routed Capsule Networks for conditional and random image-synthesis. The architectures benefit from the fact that Capsule Networks encode the properties and spatial relationships of the features in the images. Capsule Networks being a more effective critic provide multiple benefits when replacing Convolutional Neural Network discriminators being used in the current work-horses for image synthesis - DCGANs. Our architectures use a loss analogous to Wasserstein loss and demonstrate that they can encode a dataset's representation faster than currently existing GANs resulting in lesser number of training samples required. Our experiments show that the generator is pushed to give better and more diverse results in significantly lesser number of epochs without over-fitting. We have used MNIST, Fashion-MNIST and their variants for demonstrating the results achieved from this architecture.
READ FULL TEXT