Ensemble Image Colorization using Convolutional Neural Network

Abstract

Colorization of grayscale images brings more life to images and presents a more realistic view of the objects present in the real world. It can be applied to the restoration of historic monochrome images and films as well as in assisting cartoon artists. Traditionally, image colorization involved scribbling color hints or choosing a reference color image to be able to transfer the color traits into the input grayscale image. In this study, the burden of requiring user intervention in the colorization process is removed by using a convolutional neural network(CNN) that is trained on a large dataset of color images. Previous colorization studies have produced useful colorizations for elements of an image's global semantics, such as the sky, grass, and trees, but they also revealed color bleeding. We propose an approach for CNN-based colorization using ensemble colorization where separate colorizations by two different CNNs are further processed by a refinement network. Mode cost is evaluated using L2 regression loss and by performing a "Colorization Turing Test". Survey results reveal that people are fooled by most of the hand-picked colorized images considering that they are our network's best colorization results.