GAN-Based Image Restoration and Colorization

The importance of images in today’s society has made it essential for them to be of the highest quality and visually indicative of their essential traits and attributes. Significant research has been done individually on colorizing and restoring degraded images. Separate studies of Generative Adversarial Networks (GANs) have also been conducted in each of these fields. However, it’s rare to find GAN architectures that can focus on both the tasks at once. With an emphasis on nature photographs, this study proposes a unique GAN architecture that was trained on a customized image dataset including images of landscapes, flowers, and mountains combined with the GoPro Light dataset. The proposed methodology makes use of a combination of different loss functions that enable the model to focus on both tasks simultaneously. Alongside the L1 loss and adversarial loss traditionally used in GANs, the proposed model includes the perceptual loss that performs feature-wise comparisons between images to restore its inherent features. To prove that the GAN can perform both restoration and colorization, its performance has been compared with other models that perform each of the two tasks separately. The model is tested on the curated dataset and evaluated on image-specific metrics like peak signal-to-noise ratio (PSNR) and structural similarity Index (SSIM). The model gives results that compare well with existing models, and it can colorize and restore images that have been degraded with motion blur or camera misfocus—successfully striking a good balance between the two tasks. The paper concludes by providing insight into the future work that can be carried out.