Visual Sensitivity Aware Rate Adaptation for Video Streaming via Deep Reinforcement Learning
Dan Meng, Ye Jin, Wenchao Jiang, Yuanchao Shan · 2021
Adaptive bitrate (ABR) technology is widely adopted in video streaming to improve quality of experience (QoE). However, existing methods ignore the inherent characteristic of Human Visual System (HVS), i.e., HVS has different sensitivity to the quality distortion of different video content. Therefore, video contents with high visual sensitivity have higher visual importance and need to be allocated more bitrate resources. In this paper, we propose a novel rate adaptive algorithm by incorporating newly-defined visual sensitivity model. More specifically, a set of relevant features are extracted from video content to establish a deep neural network model that accurately reflects the visual sensitivity. The network state, buffer occupancy and visual sensitivity are comprehensively considered under deep reinforcement learning framework to select appropriate bitrate for maximizing QoE. Experimental results show that the proposed algorithm can generate an ABR strategy consistent with visual sensitivity, and improves the perceptual video quality and user QoE by 14.2% and 16.3%, respectively.