TSENet Video Super Resolution for Broadcast Television 

The emergence of Video Super Resolution (VSR) has advanced the media industry. It holds great potential for enhancing broadcast and streaming, improving visual quality, and boosting user engagement. While there are benefits, applying VSR technology to broadcast presents challenges, as it requires a high-quality solution that maintains consistent quality across diverse broadcast standards, handles live content with minimal latency, and ensures compatibility with existing broadcast infrastructure. This demands sophisticated AI models, often needing high-performance GPUs, which can increase costs. In this paper, we introduce a method to upscale low-resolution, compressed video frames based on an enhanced version of Equivalent Transformation and Dual Stream Network (ETDS)[1], called a Temporally Stabilized ETDS Network (TSENet) that performs equivalent or better compared to more compute extensive solutions. This paper examines various VSR models, emphasizing the key performance indicators identified by the AIM2024 Challenge for Efficient VSR[2]. The focus is on deploying these models in workflows, considering visual quality through both subjective and objective analyses, compression artifact reduction, and real-time processing capabilities.

Surbhi Madan | Intel Corp | San Diego, Calif., United States
Onur Barut | Intel Corp | Berlin, Mass., United States
Anand V Bodas | Intel Corp | Bangalore, India
Christopher A. Bird | Intel Corp | Chandler, Az., United States
Jerry Dong, Lin Xie | Intel Corp | Beijing, China