Fine Tuning GANs for Image Colorization: Effects of Training Data and Epochs

Ever since the advent of Deep Learning, there has been a lot of experimentation and research being carried out in image processing. This gave rise to the conversion of black and white i.e. grayscale images into colorized images. Image colorization is a popular image-to-image translation problem. It finds its application in various domains like image restoration, movie colorization, colorization of old aerial surveillance shots, medical image colorization, etc. Earlier this was done manually by artists and photographers but recent advancements in deep learning have enabled automatic image colorization using techniques like Convolutional Neural Networks and Generative Adversarial Networks. CNN is a neural network that has layers that are essential for extracting features from images and can learn the complex mapping between grayscale images and their corresponding ground truth images. On the other hand, GANs is a relatively new technique that consists of two components: generator and discriminator. This research paper aims to present the results obtained by using both the above-mentioned methodologies for image colorization. Many examples of generated images and their respective accuracies and structure according to the technique used are discussed in the later part.