A Comprehensive Review of Super-Resolution Techniques: Progress and Prospects from SRCNN to ESRGAN

Abstract

In computer vision, Super-Resolution (SR) is considered to an essential technology that tackles the problem of reconstructing high-resolution (HR) images from it’s low-resolution (LR) inputs. With the backing of developments in deep learning and computational architectures, numerous SR models have been created over time. From early CNN-based techniques like that of SRCNN and VDSR to more sophisticated designs like EDSR, RCAN and the attention-enhanced models like SAN and HAN, a variety of SR models have been created to tackle various challenges. Numerous Real-time and resource-constrained applications might be benefit from lightweight models like CARN and DRRN that balance efficiency and quality, whereas the Generative Adversarial Networks (GANs) such as SRGAN and ESRGAN may prioritize perceptual quality. These models are widely applicable in variety of domains such as gaming, entertainment, satellite imaging, and healthcare. This research tries to outlines the evolution, strengths and applications of SR models highlighting how they  can contribute to improving image quality and outlining potential paths for accessibility and optimization in the future.