Abstract
Image colorization is the process of artificially coloring a black and white image such that this fabrication appears realistic and authentic to the viewer. There are many nontrivial applications of this process, such as the colorization and augmentation of historical photos as well as the removal of color tone filters from images. This research employs a unique architecture of a convolutional neural network, a type of artificial neural network, to train a machine to identify patterns within images and color them accordingly. However, in place of a traditional regression-based application of this model, a novel classification-based approach is implemented through the use of a cross-entropy loss function and a discretized color space. After training this algorithm using either approach, the model is able to take grayscale images as input and output images representing attempts at their colorized counterparts. Images produced using the latter approach far outperform those produced using the former in terms of the plausibility of the coloring, as determined by user interaction. The presentation highlights implementation of this architecture, as well as the key differences between the baseline approach and the novel method. Additionally, sample colorized images are examined from both methods to understand the cases in which the different implementations succeed and fail.