Research On Rate-distortion Optimization Techniques For H.265/HEVC

Recently,with the rapid development of internet technology and multimedia technology,high definition(HD)videos and even ultra-HD videos have been pervasive in our daily life.However,rapid increase of video data brings great challenges to video data storage and transmission.Therefore,the joint collaborative team on video coding(JCT-VC)developed a new generation of video coding standard,i.e.,High Efficiency Video Coding(H.265/HEVC),in January of 2013.Compared with its previous coding standard H.264/AVC,H.265/HEVC achieves equivalent subjective video quality with approximately 50% bitrate reduction.Rate-distortion(R-D)optimization is the key technique for improving coding performance in video coding.Based on the Shannon R-D theory,R-D optimization is always trying to achieve higher reconstructed video quality when satisfying the bit-rates limitation through balancing the coding bit-rates and the coding distortion.In this way,video coding improves its efficiency.This thesis focuses on the R-D optimization in H.265/HEVC and proposes methods to improve subjective quality of the reconstructed videos,to improve R-D performance of video coding,and to reduce the coding complexity.The R-D techniques of the in-loop filtering technique,i.e.,SAO(Sample Adaptive Offset),the Random access(RA)structure,and the quatree partitioning of the coding unit(CU)are studied and analyzed.The main innovations and contributions are concluded as follows:1.A R-D optimization method of SAO in H.265/HEVC which optimizing the perception quality of reconstructed videos is proposed.Considering that,currently,the R-D optimization of SAO in the test model of H.265/HEVC,i.e.,HM,has not yet considered the perception characteristics of human eyes,the subjective quality of the videos after using SAO still has room for improvement.Firstly,a simplified just noticeable distortion(JND)model is established based on the Sobel operators.Secondly,the established JND model is applied into the R-D optimization of SAO.Then offsets are achieved according the JND values of the samples.In this way,the coding bit-rates of the SAO offsets can be assigned to different regions according to the sensitivity of human eyes to the distortion of different regions within one frame.Experimental results show that,the proposed method improves the subjective quality of the reconstructed videos.2.A three-dimensional R-D optimization method of SAO in H.265/HEVC which can suppress the Temporal Flickering Artifact(TFA)in the reconstructed video is proposed.TFA is one of the most common perceived temporal distortions in video coding.Currently,SAO in the HM derives the offsets independently for each frame in the spatial domain without considering temporal frame correlation.As a result,it can reduce the spatial distortion of the pictures,however,the temporal distortion artifacts,such as TFA,cannot be effectively addressed.The R-D optimization of the newly developed SAO considering temporal correlation,referred to as Inter-SAO,is performed to reduce the distortion of the residual samples between adjacent frames.Inter-SAO and SAO in the HM are then combined to form a 3D-SAO filtering method.Experimental results show that,3D-SAO can reduce the TFA,improve the subjective quality of reconstructed videos and improve the R-D performance of video coding effectively.3.An efficient Lagrangian multiplier selection method considering temporal dependency is proposed for the RA structure in H.265/HEVC.RA coding structure is one of the three common test coding structure in the HM.Using the RA coding structure,different frames are assigned to different temporal level,and frames in higher temporal levels use the frames in lower temporal levels as the reference frames.Therefore,frames at different temporal levels have strong temporal inter-layer dependency relationship,i.e.,the dependency between the bit-rates and distortion.In order to improve the R-D performance of the RA structure,the temporal inter-layer dependency of bit-rates and distortion of the RA structure is investigated,respectively.Then an efficient Lagrangian multiplier selection method for the RA structure is proposed.Experimental results show that,the proposed method improves the R-D performance of video coding using RA structure.4.A quantization parameter cascading(QPC)technique considering video content is proposed for the RA coding structure.Firstly,under the objective R-D measurement,the optimal quantization parameter(QP)for each temporal layer of videos with different content is achieved based on the experimental observation.Then,the relationship between the video content characteristics,i.e.,motion and texture features of the videos,and the optimal QP values are established.Finally,the optimal QPC for different video can be achieved based on the established model.Experimental results show that,the proposed method improves the R-D performance of video coding using RA structure.5.A fast SAO R-D method is proposed.During the R-D optimization of SAO,a statistics collection process is needed through looping to collect the distortion and the number of the samples for each SAO mode.And this procedure takes about 90% of the total processing time of SAO.In order to reduce the processing complexity of SAO,the relationship of the SAO parameters between higher temporal levels and the zeroth temporal level is analyzed.Then the fast SAO R-D optimization method is proposed based on the temporal relationship.In the proposed method,the SAO parameters of higher temporal levels use that in the zeroth temporal level for R-D optimization.Experimental results show that,the proposed method reduces complexity of SAO when guaranteeing the R-D performance of video coding.6.A fast coding unit(CU)partitioning scheme is proposed.The CU partitioning method in HM needs to traverse all CU depth.The final partition is determined after obtaining the R-D cost of each CU depth.Therefore,the coding complexity is high.Firstly,the linearly relationships of the bit-rates and distortion between adjacent CU depths are explored,respectively.Secondly,the R-D cost estimation model of adjacent CU depth is proposed.Finally,the R-D cost of the next CU depth is predicted after encoding the current CU depth according to the R-D cost estimation model.The predicted R-D cost is used for judging whether the current CU needs to be split.Experimental results reveal that,the proposed algorithm reduces the CU partitioning complexity when guaranteeing the R-D performance of video coding.