Automatic Image Colorization Based on Convolutional Neural Networks

Abstract

Analysis of methods and tools for image colorization was performed. It was explained why artificial neural network model was chosen for graphics information processing. The task of automatic colorization of arbitrary images was formulated. Initial data, conditions and constraints necessary for colorization model are listed. As a result of text classification, set of neural network hypercolumns was retrieved for each image processed. Colorization model was created which allows to determine color of each pixel based on hypercolumns set. In fact, this model consists of two related parts: classifier and colorizer. Classifier is based on using convolutional neural network, and colorizer is based on hash table which stores mapping of hypercolumns and colors. Algorythm of using this model for image colorization is proposed. Comparison of colorization results for developed and existing models was performed. Software tool was created which allows to perform learning of different neural networks and colorization of graphical information. Experiments shown that developed model determines image color quite correctly. Proposed algorithm allows to use convolutional neural network for colorizing black-and-white images, for color correction of pictures, etc.