Improving Learning time in Unsupervised Image-to-Image Translation

Abstract

Unsupervised image-to-image translation can map local textures between two domains, but typically fails when the domain requires big shape changes. It is difficult to learn how to make such big change using the basic convolution layer, and furthermore it takes much time to learn. For faster learning and high-quality image generation, we propose to use Cycle GAN that is combined with Resnet in a network that is connected with the residual block for upsampling to make big shape change and construct faster image-to-image translation.