Research On Visual Sensitivity Aware Adaptive Bitrate Algorithm Based On Reinforcement Learning

In recent years,video streaming traffic continues to grow and has occupied the main part of the total Internet traffic.In order to cope with the fluctuation of network bandwidth and provide smooth video playback service,adaptive video streaming technology is proposed.At present,adaptive bitrate(ABR)algorithm is widely used in dynamic adaptive streaming over HTTP(DASH)to improve the user’s quality of experience(Qo E).The traditional ABR algorithms usually select the bitrate of video chunk only according to the player and network environment,and ignore the influence of video content characteristics on user’s Qo E.Therefore,content-aware ABR algorithms are proposed and become the mainstream.However,the bitrate decision of existing algorithms still ignores the inherent visual sensitivity characteristic of human visual system(HVS).Since HVS has different sensitivity to the quality distortion of different video contents,video content with high visual sensitivity needs to be allocated more bitrate resources.Therefore,the existing ABR algorithms still have limitations in reasonably allocating bitrate and maximizing the user’s Qo E.To solve this problem,this paper designs an adaptive bitrate strategy from the perspective of user’s vision,studies the method of modeling visual sensitivity,and proposes a visual sensitivity aware ABR algorithm combined with reinforcement learning(RL)technology.The main contributions of this paper are as follows:1.Due to the complexity of HVS interactive mechanism,the existing visual sensitivity modeling methods are still insufficient in simulating HVS characteristics.Based on the analysis of the impact of different visual masking effects on the perception of HVS to video quality distortion,this paper establishes a total masking effect model for different video contents.The model adopts a variety of video features as input,and combines multi-stream feedforward convolutional neural network(CNN)with the feedback module to simulate HVS perception mechanism.Compared with the latest existing models,this model can reflect the visual sensitivity more accurately.2.This paper analyzes the corresponding relationship between the prediction results of the total masking effect model and visual sensitivity to quantify the visual sensitivity.By combining the visual sensitivity information with the input state and reward function of RL algorithm,a visual sensitivity aware ABR algorithm is proposed.The algorithm can generate a bitrate strategy consistent with visual sensitivity,and allocate bitrate based on more accurate visual sensitivity information to further optimize the resource utilization and the perceptual quality of users.3.Through extensive experimental evaluation and analysis,compared with the latest visual sensitivity prediction methods,the total masking effect model proposed in this paper has higher prediction accuracy and is robust to the video resolution.Compared with ABR algorithms without considering visual sensitivity,the proposed algorithm achieves good performance in many metrics,and improves the user’s Qo E by about 22.8%.Compared with ABR algorithms based on other visual sensitivity prediction methods,the proposed ABR algorithm also achieves a certain improvement of the perceptual quality of users.