COMPREHENSIVE STUDY OF DEEP LEARNING-BASED SUPER-RESOLUTION WITH EMPHASIS ON GANS

Abstract

Image super-resolution using Generative Adversarial Networks (GANs) has been ex tensively researched in recent years due to its ability to recover high-perceptual-quality high-resolution images from low-resolution inputs. Various GAN-based methods have been proposed over the years, which employ di!erent architectures and loss functions to increase the fidelity and realism of output images. This work integrates these develop ments and investigates their impact on various categories of images in various application domains. By extensive experimentation, we compare three highly acclaimed GAN-based super-resolution models SRGAN, ESRGAN, and Real-ESRGAN on twelve disparate im age classes. The results confirm that the performance of the model varies significantly depending on the image features and domain, which calls for the need of domain-specific methods that are capable of learning to generalize across varying image content. To address these findings, we add a new component to loss functions with orthogonal reg ularization for, Wide Activation SRGAN (WDSR-GAN), which employs wide activation residual blocks to increase feature representa-tion and training stability. Furthermore, in this work we explore how various loss functions impact super-resolution quality and illustrate how various combinations impact image sharpness and perceptual detail. To quantitatively compare model performance, we use a collection of metrics consisting of PSNR and SSIM, which collectively capture pixel-level accuracy and structural integrity. The findings of this thesis provide valuable insights into the problems and opportuni- con nections of GAN-based image super-resolution. By extensive analysis of di!erent models and loss functions in di!erent domains and metrics, this work lays a strong foundation for the design of more e”cient and flexible super-resolution algorithms. Such e!orts seek to steer future research towards more fidelity, improved perceptual quality, and increased adaptability to real-world imaging applications.