Multiple Domain Image to Image Translation for Facial Attribute Transfer
6 Pages Posted: 30 Apr 2019
Date Written: April 27, 2019
Abstract
In the field of Deep learning, image to image translation is an emerging concept of Computer Vision. Its objective is to learn the mapping between an input image and an output image. Over the years, remarkable results have been obtained in image translation for two domains. However, the existing studies require different generators for modelling multiple domain mappings. To overcome this drawback, we propose a scalable approach which can perform multi-domain image-to-image translations using only a single Generative network and a Support Vector Machine. Our architecture allows to control the different generative tasks of training the multiple datasets of varying domains using a single network. Our architecture also includes a SVM classifier for classifying the generated images into the target domains, thus providing a mechanism to check its own generated images for correctness. The results of our experiment exhibit superior qualitative performance in comparison to the existing models in translation between two domains. Our approach is extensible and we represent its success for the facial attribute transfer task.
Suggested Citation: Suggested Citation