Surpassing Traditional Image-Colorization Problems with Conditional Generative Adversarial Networks

Color helps to understand the semantic information of the image more accurately and reveals a lot more details which grayscale images cannot. By looking at an image, humans can automatically segment different objects present in an image making it easier for us to color an image. We propose a completely automated system to colorize grayscale images which learns to segment and color images in a realistic manner. We leverage the recent advancements in deep learning, Generative Adversarial Networks and improved cost functions, to overcome the problems of traditional Convolutional Neural Networks with image colorization. Given the unconstrained nature of the problem, we propose this algorithm to make a colorization model that achieves realistic colorizations. We have experimented different deep network architectures with various training algorithms and cost functions to come up with this network where we can clearly see realistic colors for given gray scale image and differentiate the characteristics of generative adversarial network from a traditional convolutional neural network.