Deep colorization by variation

Abstract

We propose an adversarial learning based model for image colorization in which we elaborately adapt image translation mechanism that are optimized according to the task. After developing approaches on improving the global and local quality of the image colorization by analyzing this processing made by network architecture and objective functions, we formulate a diverse mapping from the gray scale images to colorful images by latent space variation within the model. At last, discussion on the theoretical framework for studying color information distribution and video colorization is given.