PatchSR: Achieving HD Video Streaming with Pre-downloaded Universal SR Models

Jiahong Xu, Zhenxiao Luo, Zelong Wang, Gangqiang Zhou, Miao Hu

Abstract: The high-definition (HD) video streaming has gained tremendous popularity with the proliferation of smartphones and mobile networks. However, it is quite challenging to deliver HD online video streams directly to devices with very low bandwidth in current systems. In this paper, we propose a neural-enhanced HD video streaming system named PatchSR to provide HD video streaming for bandwidth-constrained devices. PatchSR delivers universal super-resolution (SR) models with high performance to devices in advance. Only low-resolution video streams are sent to bandwidth-constrained devices, and the video quality at the device side can be enhanced with SR techniques. The main challenge is training multiple universal SR models with high performance and selecting the dedicated SR model for each video content. To overcome this new challenge, we propose an image classification algorithm of texture features according to the Discrete Fourier Transform (DFT) feature map of the training patch. We also design a dynamic selection algorithm of SR models for clients to improve video quality. Finally, we achieve and evaluate our proposed PatchSR system with real network traces and the experimental results show that PatchSR achieves higher video quality and up to 28.65% QoE improvement compared to baselines.