Innovative AI-Powered Video Upscaling: Globo's Efficient Cloud System

Abstract

Globo, holder of a vast archive of video content with over 300 soap operas, mainly in standard definition (SD), frequently retransmits this material in high definition (HD), resulting in high processing demand. To tackle this challenge, we developed an innovative cloud system that adapts to our workflow, integrating seamlessly with the company's operational tools. This system not only offers scalability but also reduces processing costs by activating machines only when necessary and remaining in stand-by during idle periods.

Our system utilizes Artificial Intelligence models trained with videos from our archive, enabling them to handle images with different focus adjustments and intense action scenes, common in soap operas and series produced by Globo. The upscaling models are based on the ESPCN convolutional network architecture, implemented in Ffmpeg, a cross-platform solution used in the backend of many applications for tasks such as media converters, streaming solutions, and audio and video players.

All the tools, libraries, and frameworks used are open-source, eliminating costs associated with proprietary software licenses. The system is designed to be agnostic, capable of being implemented on most cloud platforms (Google, AWS, Azure, etc.) without the need for complex changes. The only requirements are buckets, VMs, messaging services (PUB/SUB), and serverless services, allowing it to run in any cloud environment.

Our solution ensures adaptability and cost efficiency, with a service level agreement (SLA) of less than 10 minutes. This efficiency is maintained regardless of the volume or duration of the video, thanks to the strategic use of preemptive machines that distribute the workload efficiently and allow for rapid provisioning in case of interruptions.

The process is fully automated and integrates seamlessly with the commercial platforms used in Globo's workflow. Additionally, the tool is available to Globo's CGI team, enabling the super-resolution of backlog content in SD or HD formats. This dual application highlights the platform's versatility and its significant role in improving broadcast content quality.

In our next phase, we aim to make this system accessible to other companies and individual users online, exploring innovative business models that provide monetization opportunities while meeting our operational needs. This initiative represents a crucial step in expanding the reach and benefits of our cutting-edge technology, offering a valuable tool for high-quality, low-cost video processing to a broader audience.