DeOldify: A Review and Implementation of an Automatic Colorization Method

Abstract

DeOldify is a recent automatic colorization method based on Convolutional Neural Networks which yields impressive results. The method was initially created by Jason Antic with the support of the Californian start-up Fast.ai and thus does not come from the academic research world. The goal of this paper is twofold. First we propose a rigorous mathematical presentation of the method along with a critical analysis of its different steps. Second, we provide an open-source implementation of a simplified but effective version of the approach, based on Pytorch and without dependence on the Fast.ai framework.