Research On High Efficiency Inter Coding In Video Compression

In recent years,the emergence of HD and Ultra HD video makes the second generation video coding standard H.264/AVC and AVS can not meet the needs of practical applications.To improve coding efficiency,JCT-VC has finalized the next generation video coding standard called High Efficiency Video Coding(HEVC)in 2013.HEVC can save about 50% bitrate at the same video quality compared to H.264/AVC.AVS working group has finalized AVS2 in 2014,which has similar coding performance as HEVC.With the increasing demand of video quality and video resolution,higher requirements for the coding efficiency of video coding standard are put forward.Inter prediction plays an important role in most video coding standards since it can improve the coding performance significantly.However,the widely used inter prediction of the video coding standards can not deal with the complex motion effectively or in some cases is not accurate enough.Therefore,it is important to obtain better coding efficiency by enhancing the inter prediction performance based on the current video coding standard.3D video is attracting more and more attention since users can experience realistic 3D scenes and select viewpoints interactively.As an extension of HEVC,3D-HEVC with the primary focus on efficient compression of 3D video.Besides the existing coding tools in HEVC,3D-HEVC introduces some new coding tools to improve the coding performance of multi-view video.The introduction of new coding tools results i n a significant increase of the encoding complexity.Therefore,it is important to optimize the prediction technologies of 3D-HEVC and reduce the encoding complexity of 3D-HEVC.In this thesis,we mainly focus on the research topics about enhancing the inter prediction performance of video coding standard,optimizing the prediction technologies of 3D-HEVC,and reducing the encoding complexity of 3D-HEVC.The contents of the thesis are divided into the following four sections:Firstly,a merge mode for deformable block motion information derivation is proposed.Motion estimation plays an important role in most video coding standar ds since it can improve the coding performance significantly.Block matching motion estimation(BMME)based on translation motion model is widely employed in current video coding standards,which assumes that all the pixels in a block undergo the same motion.However,BMME cannot represent the complex motion accurately,such as rotation,zoom,deformation motion.To resolve the problem,we propose a merge mode for deformable block motion information derivation,which considers bilinear interpolation model,six-parameter affine model,and four-parameter affine model simultaneously.The minimum sum of absolute transformed differences(SATD)criterion is used to select the best model for a block.In the proposed merge mode,the motion information of current block is derived from the motion information of its neighbors.It does not need to perform motion estimation and transmit motion parameters,which effectively avoids the problem of estimating and efficiently coding the increased motion parameters of the complex motion model existing in the previous solutions of deformation motion.The proposed merge mode is integrated into HM14.0.Experimental results show that compared to HM14.0,the proposed merge mode can bring 11%~18% BD bitrate saving for deformation sequences,and 1%~3% BD bitrate saving for test sequences of HEVC on average,while the increased encoding and decoding complexities are moderate.Secondly,a scheme of adaptive Wiener filter to improve the accuracy of inter prediction is proposed.Inter prediction plays an important role in most video coding standards since it can improve the coding performance significantly.The more accurate the inter prediction is,the higher the coding efficiency will be achieved.However,the current inter prediction may not be accurate enough,especially for scenes with significant temporal illumination variations,scene changes or noises.In this paper,an adaptive Wiener filter(AWF)is proposed as an additional module in motion estimation(ME)and motion compensation(MC)process to improve inter prediction performance.In AWF,the filter coefficients of each coding unit(CU)are estimated in both encoder and decoder by referring the reconstructed spatial neighboring pixels of current CU and those of its temporal refere nce block.Only one flag is coded and transmitted to the decoder for a CU to indicate whether AWF is used.The adaptive Wiener filter used in this paper is more powerful for improving the inter prediction accuracy compared with linear regression model and offset model.Experimental results show that the proposed adaptive Wiener filter can improve the coding performance of HEVC,while the increased decoding complexity is tiny.Thirdly,some modifications for optimizing the prediction technologies of 3D-HEVC are proposed.The disparity vector(DV)derivation is critical for inter-view motion prediction,inter-view residual prediction,disparity-compensated prediction(DCP)or any other tools exploiting inter-view correlation.It is used to locate a corresponding block in another view.In HTM-5.0(3D-HEVC test model version 5),the DV is derived from some spatial and temporal neighbors.This paper advocates modifications of the DV derivation to reduce the complexity and to achieve slightly better coding efficiency.Firstly,we propose to remove the additional temporal block,unify the DV searching order for all views,and to impose restrictions on the BR temporal block for memory access bandwidth and complexity reduction in DV derivation.To obtain slightly better coding performance,we propose to search the DV among the temporal neighboring blocks first as the DVs derived from temporal neighboring blocks are more accurate than those derived from spatial neighboring blocks.View synthesis prediction is used to improve the coding performance of multi-view video by mapping the pictures of the adjacent views to current view for predicting the current picture.Backward view synthesis prediction is adopted by 3D-HEVC as it can be implemented with block-based processing more efficiently.To simplify the view synthesis prediction in 3D-HEVC,this paper made two modifications:(1)Unify the depth data accessed by depth oriented neighboring block disparity vector(Do NBDV)and VSP;(2)Reduce memory requirement for neighboring block disparity vector(NBDV)and BVSP mode flag.Both of the modifications introduce no overall coding performance change while the memory access bandwidth and memory requirement are reduced.Lastly,a fast encoder decision algorithm for texture coding i n 3D-HEVC is proposed.The improvement of the coding efficiency of 3D-HEVC is obtained at the expense of a computational complexity increase.Besides the existing coding tools in HEVC,3D-HEVC introduces some new coding tools to improve the coding performance of multi-view video.The introduction of new coding tools results in a significant increase of the encoding complexity.How to reduce the encoding complexity is becoming a critical problem in applications.In this paper,a fast encoder decision algorithm to encode the dependent texture views is proposed,where two strategies to accelerate encoder decision by exploiting inter-view correlations are utilized.The first one is an early merge mode decision algorithm,and the second one is an early CU splitting termination algorithm.The strategy of early merge mode decision includes two conditions:(1)All five inter-view neighbor blocks are coded as merge modes;(2)The RD performance of skip mode is better than 2Nx2 N merge mode for the current CU.If the above two conditions are true,all prediction units(PUs)within the current CU will only need to examine the merge modes instead of examining all the prediction modes.The strategy of early CU splitting termination includes two conditions:(1)The CU splitting depth level of the current CU is equal to or larger than the maximum depth level of the five inter-view neighbor blocks.(2)Skip mode is selected as the best prediction mode for the current CU after checking all the possible prediction modes.If the above two conditions are true,no further CU splitting is needed for the current CU.In this paper,we design a fast encoder decision algorithm of texture coding for 3D-HEVC by combining the inter-view correlation and coding tools of 3D-HEVC.Experimental results show that the proposed algorithm can achieve 47% encoding time reduction with overall 0.1% BD bitrate reduction.