Distribuição de vídeo na internet aprimorada por super-resolução baseada em redes neurais adversárias generativas

Abstract

Over the years, video content distribution over the internet has increased dramatically, with video content predicted to represent 82% of all internet traffic. Major video content providers, such as Netflix, Prime Video, and YouTube, use content delivery networks (CDNs) to replicate their content in locations closer to their users, improving latency and avoiding rebuffering. Content providers use adaptive video streaming to enable viewers to watch videos with adequate image quality based on their end-to-end connection with the provider. However, this technology requires multiple representations of the same video with varying resolutions and bitrates, increasing the volume of redundant data flowing through distribution infrastructures and overburdening CDN infrastructure. Recent literature has highlighted deep neural networks, particularly generative adversarial networks, for image and video super-resolution methods. These methods can restore low-resolution images and videos to high resolution with unnoticeable quality loss to human vision. In this study, the application of video super-resolution using a generative adversarial network was evaluated in two ways: i) To reduce video traffic in cloud infrastructures, videos were replicated in low-resolution versions between data centers and geographically distributed surrogate servers. These videos were then restored to high resolution on these servers using a super-resolution model; ii) To optimize the quality of experience for viewers of live video streaming applications, enhanced perceptual quality was achieved through super-resolution. The research was conducted in an experimental setting designed to simulate real-world scenarios. The findings demonstrate two frameworks that use a super-resolution model, one for a cloud video replication service and the other for a live video distribution service supported by edge computing. The experiment results revealed a reduction in video-related traffic in infrastructures of up to 88.37%. Additionally, the quality of the session experience during live video streaming was improved, as measured by perceptual quality. Overall, the study suggests that using super-resolution techniques for video content delivery can reduce network traffic and improve the quality of experience for viewers. These findings could have implications for the future of video content delivery, especially as video content continues to grow in popularity and demand.