Research Of Video Coding And Transport System Based On Cross Random Access Point Referencing

As the consuming demand of visual contents increasing,video data on the Internet grows explosively.The storage and transport of massive video data consume mounts of memory and bandwidth resources.New generations of technologies for higher video compression ratio are pro-posed by thousands of experts in the video coding framework in order to decrease the size of video data and the resources it consumes.However,there are many Random Access Points(RAPs)in-serted in video bitstream,which are used to support flexible random access.Those RAPs divide the bitstream into multiple Random Access Segments(RAS s)that are independent from each other Each RAS can only be compressed or coded by use its own information rather than the reference informaton in other RASs,even though other RASs have similar contents with the current RASThe independent restriction results in that the redundant information between multiple RASs cannot be eliminated,which restricts the coding efficiency for future increasing.In fact,there are many videos in practical services where similar contents repeat alternately between RASs that are far from each other in temporal domain.In order to make full use of the correlated information existing in large temporal duration,previous work has proposed a scene library based video coding scheme which introduces a scene library containing External Reference Pictures(ERPs).ERPs are composed of diverse scene contents and provide external reference information shared by multiple RASs with similar scene contents.However,the ERPs bring two issues for video coding.One issue is in the coding layer where the content and quality of generated or selected ERPs decide the coding efficiency.Obtaining a set of optimal ERPs requires high coding complexity,which is not efficient for the compression of massive video data.Thus the key problem in the coding layer is how to obtain a set of optimal ERPs with the least coding complexity and also the highest coding efficiency.Another issue is in the system layer where sharing ERPs in a large temporal duration challenges the random access functionality and transport efficiency of a video.Hence another key problem in the system layer is how to realize the compact storage and efficient transport of ERPs and the RASs depending on themThis thesis provides a comprehensive discussion and design of the video coding scheme using external reference information,which is also referred to as video coding based on Cross Randsom-access-point Referencing(CRR).We focus on the research of the efficient ERP selection in the coding layer and the efficient storage and transport of CRR coded video in the system layer.The main contributions are as the followings·In the coding layer,we aim at the selection of optimal ERPs for improving the coding per-formance.On the one hand,we try to decrease the complexity of CRR based video coding,which is more time-efficient for videos with long duration than the previous work.On the other hand,we try to improve the coding efficiency of the CRR based video coding,which futher increase the performance compared to the previous work1.We propose a crowdsourcing based optimal ERPs selection which formulates the joint rate distortion of the ERPs and the main video referencing them and transform the joint Rate Distortion Optimization(RDO)problem to a user selection problem in crowd-sourcing theory.According to the submodular property of the joint rate distortion func-tion,the RDO problem is solved by a local search based optimization algorithm.The proposed method reduces 20%encoding complexity on long drama sequences com-pared to the previous work and still acheives 28%coding gains compared to HEVC2.We propose a local distortion propagation based quality enhancement for ERPs.The proposed method builds the distortion propagation chain of blocks in ERPs based on the dependency between main video frames and ERPs.The weight of the distortion propagation chain is used to decide the quantization parameters of blocks in ERPs Experiment results show that the proposed method further improves coding efficiency of CRR based video coding by 4%compared to the above crowdsourcing based CRR method.Combining both the proposed methods in coding layer,the CRR based video coding finally gains more than 19%on sequences in HEVC common test condition and more than 38%on long drama sequences,compared to HEVC·In system layer,we focus on storage in local play and transport in streaming service for CRR coded video,which solves the joint storage and synchronized streaming of ERPs and the main video referencing them and realizes the compact storage and efficient transport with convenient operations in the system layer1.In file format for storage,we propose a sample group based file format for j oint storage of ERPs and the main video.The proposed method clusters main video frames into multiple groups according to the ERPs they rely on.Each group records the indices of the referenced ERPs which will be fetched from file for decoding the main video Compared to the conventional file format,the proposed method avoid duplicated ERPs'data and can compactly store the ERPs and the main video frames together2.In streaming,we propose a segment index based external dependency description for the synchronization of ERPs and the main video frames.The proposed method extends the conventional segment index by introducing external dependency description and makes it possible for the streaming client to parse the external dependency and request the main video frames and its referenced ERPs simultaneously.Meanwhile,to coop-erate with the external dependency description,we also propose a buffer management for long storage and reuse of ERPs,which enables the streaming client to manage ERPs based on their usage frequency and increase ERPs' hit ratio.The proposed method can reduce the number of re-downloading and re-decoding ERPs,which is more efficient than the conventional streaming scheme which re-download and re-decode ERPs mul-tiple times.Experiment results show that the streaming client using the proposed buffer management can reduce 45%bits in transport and 9%decoding time,compared to the conventional streaming clientWith the above two methods in system layer,CRR coded video can be stored compactly and transported efficiently without introducing extra memory,bandwidth or decoding resources for ERPs,which maintains the high efficiency obtained by CRR coded video coding scheme in the coding layer.