Colorizing Black and White Images Using Deep ConvNets and GANs

Abstract

In this paper, a technique to colorize black-and-white images combining both localized and global features has been presented. The technique is based on Deep Convolutional Neural Networks and Generative Adversarial Networks (GANs), which merge localized scene information derived from small area with the global scene information of sample image. The whole framework, including the colorization model, is trained in an E2E manner, i.e., end-to-end, because of its ability to perform well irrespective of the knowledge of the problem. Unlike other convolution-based approaches, the model is able to process images of varied resolutions. To train the model, an existing large-scale scene classification dataset has been used. The class labels about the types and scenes represented in various images have been utilized for efficient training. In this work, the adversarial networks on top of deep convolutional nets have been used as a generalized approach toward I2I translation. What makes these networks generic is the way their algorithm learns loss function mapping along with mapping from input to output layers, which otherwise would have required separate approaches. At last, we have compared the proposed model with some existing models used for grayscale image colorization. The proposed model demonstrated promising result on many different types of black-and-white images and delivered realistic colorization, even on random images taken from the Internet.