A GAN-based model for single image super-resolution on consumer-grade GPUs: comprehensive analysis and development of MSRGAN

Abstract

In a world where visual content plays a crucial role in anything imaginable, the need for sharper, more detailed images has never been more important. This research paper explores innovative approaches to improve the quality of single images through the application of Deep Learning techniques, specifically GAN architectures. The research focuses on developing an efficient and feasible model that can be trained in a consumer-grade GPU. Among various architectures, the research focused on the RRDB model for further development. With the modification of the RRDB model and the implementation of activation functions combination and proper loss functions, this research seeks to achieve enhanced performance and effectiveness. Finally, the proposed model MSRGAN was developed which was capable of training on a consumer-grade GPU with an average amount of video RAM. The model possesses the capability for 4x upscaling. For testing the performance, the research used PSNR and SSIM evaluation metrics where the MSRGAN has outperformed the basic SRGAN and ESRGAN.