Video Super Resolution with Generative Adversarial Network

This paper exhibits a novel implementation of video super resolution through Generative adversarial network. Generative adversarial networks (GANs) are a class of artificial intelligence algorithm implemented by a system of two neural networks provide a unique way to learn deep representations. The results of GAN shows slightly lower PSNR compared to traditional methods, since the frames produced by GAN is more appealing when viewed by human. GAN can learn from large datasets and automatically add high-frequency details and features to frames while traditional methods can't. The generator discriminator architecture in GAN pushes it to generate more realistic and appealing frames.

I. Introduction

Video Super-resolution is a classic problem in video processing that addresses the question of how to reconstruct a high resolution (HR) frame from its downscaled low-resolution (LR) version. The temporal relationships in the input can be used to improve reconstruction for video super-resolution and combine the information from as many LR frames as possible to reach the best video super resolution results. This paper mainly focuses on video super resolution through generative adversarial network. In general, the existing SR approaches can be divided into two categories: Frequency domain based SR methods and Spatial domain based SR methods.