The goal of this project was to study how a grayscale image can be colorized using a convolutional neural network (CNN). The method presented in the paper Colorful Image Colorization, written by R. Zhang et al., was used as reference [1]. The data used to train the model presented in this paper is confined to portrait images of people’s faces. The dataset used in this study is thus more narrow than what is used in the original paper, which is a wide variety of image categories. Two different loss functions are implemented with results showing that a multinomial cross-entropy loss function produces much better results than a euclidian loss function. Although it has been trained for far fewer iterations, the model presented in this paper colorizes portrait images more realistically in 58% of the cases when compared to the model presented by R. Zhang et al. in [1] in a qualitative study.
See below for some examples of results of images colored by our model compared against the ground truth:
