AWS Compute Video Super-Resolution powered by the Intel® Library for Video Super Resolution
Carlos Salazar, Surbhi Madan, Anand Bodas, Arturo Velasco, Christopher A. Bird, Onur Barut, Tahani Trigui, Xiaoxia Liang · 2024
The rise of free ad-supported streaming TV (FAST) services has boosted personalized content, including classic movies and shows. Most of this content is available in lower-resolution formats (SD) and needs an enhanced viewing experience. Traditionally, low-complexity upscaling methods like Lanczos and bicubic have been used. However, they often introduce image artifacts such as blurring and pixelation. While deep learning (DL) techniques such as SRCNN [1], and EDSR [2] have shown remarkable results in terms of quality metrics (VMAF, SSIM, PSNR). However, they are computationally expensive, and can require GPUs to implement, making them unsuitable for low-cost channels,the most typical case in FAST services. Therefore, we (Intel®1 and AWS) propose a cost-effective architecture for video super-resolution that leverages the benefits of Amazon Spot instance to process video assets using the Intel® Library for Video Super Resolution (Intel® Library for VSR) on Intel® Processors balancing quality and performance for real world use cases.