A GAN to Fight Video-related Traffic Flooding: Super-resolution

Abstract

Image and Video Super-Resolution problems become relevant in applied Deep learning due to recent results on using Convolutional Neural Networks and Adversarial Training method to solve such problems. This body of work has focused on conceiving or improving super-resolution methods, and validating them. Little attention has been devoted to their application. Video streaming has the highest popularity among Internet users being responsible for the most significant portion of today's Internet traffic. In this paper, a single image super-resolution model is applied to conceive a video super-resolution model. The designed model was tested against a video base made up of 220 clips, and each clip was encoded in four resolutions. The numerical results showed that the conceived model output is virtually indistinguishable from its ground truth and its use in the context of video distribution decrease by almost 84.5% the related traffic.