Application of generative adversarial networks for image super Resolution

Abstract

In image super-resolution applications, generative adversarial networks (GANs) are often employed to convert low-resolution images into high-resolution ones, thus improving image quality. These networks are made up of a discriminator network that assesses the validity of the generated samples and a generator network that generates new samples based on the data. Image quality has been significantly improved using GANs, especially in situations where data samples are scarce. In this thesis, we are implemented and tested a several models of image super resolution based on GANs such as: SRGAN, ESRGAN and Real ESRGAN. We compared the obtained results of the deep learning methods with traditional methods like Bicubic interpolation using the PSNR and the SSIM metrics.