| Light field can record both the intensity and direction of light rays,and this capability breaks the bottleneck of traditional photography which captures just the limited information of light in a scene.Furthermore,the advent of light field technology made a revolutionary change in traditional image and video processing,so it has become one of the research hotspots in the field of digital media recently.However,the massive amount of data usage for light field information puts high pressure on its storage and transmission,which is also one of the biggest obstacles to its development.To mitigate the above problems,the research on efficient light field compression designs is of urgency and interest.Current work on light field compression mainly focuses on light field images,while light field video compression remains a relatively under-explored field.Light field videos,as the latest format of light field content,contains multidimensional information.The rich information enables increasingly immersive 3D visual experiences to viewers.Thus,light field videos have a positive impact on the popularity of light field technology,which means the research value for light field videos appears more self-evident.The time dimension has been introduced in light field videos,making the storage and transmission challenges yet more urgent and difficult when compared to light field images.Given this information,an efficient coding strategy for such a considerably large volume of data will be a pivotal factor in opening the door to new market opportunities.There have been several multiview-based light field video compression methods focusing on exploring efficient prediction structures to exhibit data redundancy reported in the literature.Although these methods have advanced light field video compression to some extent,there is still space for improvement.Towards such research issues,this thesis carries out the studies on light field video compression in terms of prediction structure,sparse sampling coding,user-dependent interactive application by taking into account all the intricacies of light field videos.The major contributions of this work are summarized as follows:(1)Considering that existing prediction structures cannot fully remove the inter-view redundancy in light field data,this thesis proposes a similarity-assisted prediction structure for light field video compression by exploiting the intrinsic spatial self-similarity.Firstly,much more efficient compression can be achieved by computing similarities among all the views in a light field video and using them for view type selection.Furthermore,the best trade-off between compression efficiency and random access complexity is obtained by leveraging horizontal,vertical and diagonal references.Experimental results show that the proposed strategy can achieve about 7-35% compression efficiency improvements while retaining better randomaccess efficiency,compared to the existing inter-view prediction structures.(2)Considering that a large number of views in a light field video bring great pressure to the encoder,this thesis proposes a view synthesis-based sparse coding strategy for light field videos by exploiting their inherent geometrical structure.Firstly,the number of views that need to be encoded can be reduced by utilizing sparse sampling,which is more beneficial to making use of inter-view correlations.Then the prediction coding is improved by a view synthesis algorithm.The multidimensional data redundancies of light field videos can be decreased effectively due to incorporating the advantages of both sparse coding and prediction coding.Finally,the light field video is reconstructed at the decoder side.Experimental results show that the proposed strategy can achieve about 35-50% bitrate savings and up to 2.4d B PSNR gains against the existing compression methods.(3)Considering that existing light field video compression methods are inapplicable to low-latency interactive applications,this thesis proposes a user-dependent interactive light field video compression strategy.The proposed strategy contains two components: a userdependent view selection scheme and an adaptive compression design for selected view sequences.Firstly,by predicting trajectories and using projection models,the selected views of users for streaming are determined by the presented view selection method.Then the strong dependency between different types of views is reduced by the proposed coding method,which results in a relatively good balance between view access and compression efficiency.Experimental results illustrate that the proposed strategy can achieve the best performance compared with other traditional light field video compression methods in user-dependent interactive scenarios. |
