Image Colorization Using a Deep Transfer Learning

Abstract

Over the past decade, the automatic image coloring has been of particular interest in applications such as repairing damaged or old images. One of the problems with the auto-coloring is the ability to predict multiple color results for gray image pixels. In the presence of noise, the problem becomes more complicated. Recently, some researchers employ conventional neural networks (CNN) to the problem of image colorization. Usually, the output of the last layer of CNN is used as a feature representation. However, the information contained in this layer may be too coarse spatially to allow exact localization. Conversely, earlier layers may be precise in localization but will not capture semantics. In this article, we use a concept called hypercolumns to achieve the best in both cases and develop a fully automatic image coloring system. Our approach exploits recent advances in deep neural networks and uses the semantic representation to provide an accurate color prediction. Since many elements of the scene naturally appear by the color distribution, we train our model in such a way to predict the color texture of each pixel. The DIV2K dataset has been used for training, and the obtained results are compared with other methods based on PSNR, which are promising.