Unsupervised Transformation Network Based on GANs for Target-Domain Oriented Image Translation

Image-to-image translation usually refers to the task of translating an input image from the source domain to the target domain while preserving the structure in the source domain. Recently, generative adversarial networks (GANs) using paired images for this task have made great progress. However, paired training data will not be available for many tasks. In this paper, a GAN-based unsupervised transformation network (UTN-GAN) is proposed for image-to-image translation. Importantly, UTN-GAN employs hierarchical representations and weight-sharing mechanism to translate images from the source domain to the target domain without paired images. We employ two groups of unsupervised GANs to generate the images in different domains first, and then discriminate them. In UTN-GAN, an auto-encoder reconstruction network is designed to extract the hierarchical representations of the images in the source domain by minimizing the reconstruction loss. In particular, the high-level representations (semantics) are shared with a translation network to guarantee that the input image and the output image are paired up in the different domains. All network structures are trained together by using a joint loss function. The experimental studies in qualitative and quantitative aspects on several image translation tasks show that the proposed algorithm is effective and competitive compared with some state-of-the-art algorithms.