Research On 360 Degree Dynamic Video Quality Optimization Transmission Based On Deep Learning Method

Nowadays,the rapid development of Virtual Reality technology and the video sharing market of VOD(video-on-demand)stream make online360-degree video a new form of internet media.In the past few years,many360-degree video delivery schemes are proposed,focusing mainly on compression methods of 360-degree images and videos.Yet,the characteristics of high resolution in 360-degree video require a high network bandwidth,otherwise will suffer from significant video quality drop.Thus to propose an effective system to adjust 360-degree video quality and to reduce network storage consumption based on image tiles while keeping a good quality of video is a challenging problem.In this paper,our main contribution is the design of a transmission optimization system of 360-degree images or videos.This paper focuses on using deep learning Salient Detection Method with Motion State Detection to reset 360-degree image or video quality.The system first predicts the salient portion of video and then separate one frame into foreground frame and background frame.Then by using a newly proposed D-MS SSIM(Differential Multi-Scale Structural Similarity Index Method)algorithm,the system detects important frames and estimates the motion state of foreground,background and the entire view.The following procedure includes methods which lower the quality of dynamic background view meanwhile maintaining the high quality of foreground.This system separates a 360-degree video into two videos with a much smaller file size or bit rate of the background during network transmission.Finally,a newly proposed Qo E(Quality of Experience)method combining both subjective and objective standards evaluates the quality of video or image before and after processing to validate the system design.The main advantage of the scheme is that it is able to propose a new optimization system to detect salient foreground and less noticed dynamic background of VR graphics,and is able to effectively judge the motion state of each view.As a result,in the transmission of the VR graphics,the network can execute a parallel transmission of a foreground and a size-reduced background,costing less network resources.Due to the processing cost of local calculation resources,this system achieved a balance between network transmission cost and calculation cost.Four experiments about Salient Model Comparison,Motion State Detection,File Size and Bit Rate Optimization and Qo E Evaluation are implemented.The results prove that comparing with traditional 360-degree image or video optimization for the entire view,this newly proposed scheme can effectively reduce global transmission bandwidth of the network by segmenting the views into stable high quality salient and dynamic low quality non-salient areas,with adequate Qo E while watching.