Colourization of Old Black-and-White Media Using GAN (Generative Adversarial Network)

Colourization of black and white media has been a long-standing challenge in image processing. This research explores the utilization of Generative Adversarial Networks (GANs) to address this task. The study entails a deep learning framework designed to automatically add colour to greyscale images. The methodology involves the creation of a GAN ar- chitecture comprising a Generator and Discriminator network. The Generator learns to transform input greyscale images into plausible colourized versions, while the Discriminator aims to distinguish between generated and real colour images. The dataset consists of pairs of black and white images alongside their corresponding colour counterparts. The GAN models are trained iteratively, alternating between updating the Generator and Discriminator weights, facilitating the learning process. Pre- liminary experimentation involves dataset preparation, resizing images, model architecture design, and training procedures. The training loop involves optimizing the networks’ weights based on binary cross-entropy and mean absolute error loss functions. The outcomes of this study are represented in the form of loss convergence plots for the Generator and Discriminator over epochs, illustrating the progression of the models’ learning process.