REVIVING BLACK AND WHITE IMAGES: ENHANCING COLORIZATION WITH GENERATIVE ADVERSARIAL NETWORKS (GANS)

Abstract

One of the more intriguing deep learning applications is colourizing black and white photographs. The technology of automatic image colorization has sparked a lot of attention in the recent decade for a range of applications, including the restoration of aged or degraded photos. This task used to need a lot of human input and hardcoding, but now, thanks to AI and deep learning, the entire process can be automated from start to finish. Using GANs to restore and recolorize historical photos is one answer to this problem. The purpose of this project is to demonstrate GAN functionality and superiority by using a Generative Adversarial Network (GANs) that accepts fixed size black and white images as input and produces corresponding coloured images of the same size as output. Concerning the dataset, I have approximately 3000 rgb photographs from diverse domains such as mountains, forests, and cities, which we will convert to grayscale and use as labels for our model. I used binary cross entropy as the discriminator's loss function and mean squared error as the generator's loss function, & then I used Adam to optimise the generator and discriminator. Moving on to the results, the colourized output from the generator was significantly closer to the original rgb image. The project's future work will include colourizing photographs of our grandparents in order to make the image more remembered, and I also intend to continue this work by colourizing videos of historical heroes such as Charlie Chaplin. Finally, I want this model to be implemented on a web platform so that users all over the world can convert their old black and white images to colourized versions.