Abstract:
The primary objective of this project is to construct a deep learning-based Generative Adversarial Network. Super-resolution, the process of enhancing the resolution of lowresolution images, has garnered significant attention in the field of computer vision. In recent years, Generative Adversarial Networks (GANs) have emerged as a powerful tool for achieving remarkable improvements in super-resolution tasks. This project focuses on leveraging GANs and deep learning techniques to address the challenge of super-resolving low-resolution images.The proposed approach involves training a GAN architecture consisting of a generator and a discriminator. The generator network aims to transform lowresolution images into high-resolution counterparts, while the discriminator network learns to distinguish between generated high-resolution images and real high-resolution images. This adversarial training setup fosters a competitive learning process, driving the generator to produce increasingly realistic and visually appealing results. By leveraging the outcomes from the first phase, I progressively refine the generator’s performance, ultimately leading to superior results compared to its previous state. Despite significant advancements in single image resolution through the use of fast and deep neural convolution networks, a central challenge remains unresolved: effectively preserving fine texture details while accurately addressing substantial objects. In light of this, I propose SRGAN— an innovative Super-Resolution (SR) Generative Adversarial Network (GAN).The results obtained demonstrate significant improvements over traditional interpolation-based methods and state-of-the-art super-resolution techniques. Overall, the proposed approach holds promise for various applications, including medical imaging, surveillance systems, and digital entertainment, where high-resolution imagery plays a vital role.
