Alternate Approach to GAN Model for Colorization of Grayscale Images: Deeper U-Net + GAN

Abstract

Image colorization refers to applying appropriate colors in a given grayscale image, such that the viewer can accept the results as close to reality. By analyzing existing colorization algorithms based on AutoEncoder and VGG-16, this paper showed that they are not able to provide an accurate result in most cases, and are inefficient in terms of computation time. In contrast to these models, we suggested a new model developed from an established GAN model. By reforming the generator part and adding 1x1 convolutional layers based on VGG-11, we were able to create a deeper model where we could also apply nonlinear functions such as ReLU, and Leaky-ReLU. Comparing the results printed by our new model and the conventional model, we proved that our model produced better results in terms of the accuracy and clarity of colors and computation time. However, there is still room for further research in which one can investigate the optimal number of convolution layers and depth that maximizes accuracy and minimizes computation time. Still, this research holds value in that it successfully provides an alternate algorithm with better performance, and opens a path toward further development for colorization algorithms.