Abstract
Ultra-high-definition (UHD) videos are enjoying increased popularity in people’s daily usage because of the good visual experience. However, the data size of UHD videos is 416 times larger of HD videos. This will bring many challenges to existing video delivery systems, such as the shortage of network bandwidth resources and longer network transmission latency. Super resolution (SR) algorithms are widely used in video delivery applications to tackle these challenges. However, applying the super resolution model on UHD videos requires much more GPU memory, as compared with HD videos, which brings a significant challenge to existing systems.In this paper, we propose a deep compression framework named Pearl, which utilizes the power of deep learning to compress UHD videos. New channel-based super resolution models are developed to overcome the GPU memory shortage problem. In pearl, instead of applying the traditional RGB-based super resolution model, three separate super resolution models are trained based on the Y, U, and V channels of UHD videos. These super resolution models are used to reconstruct a UHD video from a low-resolution video. With Pearl, super resolution algorithms can be successfully applied to UHD videos. As a result, the data size of UHD videos can be significantly reduced during network transmission. At the same time, the efficiency of video encoding and decoding can also be improved with Pearl. To the best of our knowledge, Pearl is the first deep learning driven compression framework on UHD videos. We evaluate the performance of Pearl with extensive experiments. In all considered scenarios, Pearl can compress up to 95% of video data size during the video transmission and achieve 2.4 times faster, as compared with existing systems.
| Original language | English |
|---|---|
| Title of host publication | IEEE International Conference on Communications |
| Publication status | Published - 2021 |