Pencil images translation into realistic images using GAN

Abstract

This thesis presents an innovative approach to image-to-image translation, focusing on the conversion of colour pencil art, black and white pencil images, and grayscale images into RGB images using the Pix2Pix Generative Adversarial Network (GAN). Building upon the established Pix2Pix framework, which utilises a UNet-based architecture with skip connections for the generator and a PatchGAN discriminator, this research introduces a novel modification by incorporating an intermediate layer into the Unet architecture. This modification aims to refine the translation process, addressing the intricate details and colour dynamics of the input images. Empirical results demonstrate that this enhanced generator architecture significantly outperforms the traditional UNet architecture in producing more vivid, detailed, and colour accurate RGB images. By leveraging this advanced method, the study not only contributes to the body of knowledge in neural network applications for artistic image translation but also opens new avenues for further research in high-fidelity image conversion technologies. This work underscores the potential of targeted architectural modifications within GAN frameworks to achieve superior image translation outcomes, thereby setting a precedent for future explorations in the field.