Gray-scale Image Colorization using Generative Adversarial Networks

Abstract

Gray-scale image colorization is a classical problem in computer vision. The purpose of colorization is to assign a color value to each pixel of a gray-scale image which as similar as possible to the real one. In this paper, we propose an automatic colorization model by using conditional generative adversarial networks. The model is trained on two different dataset including a cartoon dataset and a real-world image dataset. Each dataset consists of 1200 image data. We divide them into 960 training images and 240 testing images. Experimental results demonstrate that the proposed model can achieve a reasonable result even though the training set is small.