Research On Key Technology For AVS Video Coding And The Realization On DaVinci Platform

AVS is a new generation of video coding standard, which is formulated by Chinaand has the independent intellectual property rights. It is based on the framework ofH.264and emphasizes the independent intellectual property rights. Theimplementation complexity is taken into fully account in AVS, so that thecomputational complexity of AVS is only70%of H.264, while coding efficiency isequal to that of H.264. AVS has the following characteristics:(1) high-performance,its coding efficiency is2times more than MPEG-2, equal to that of H.264;(2) lowcomplexity, the algorithmic complexity is significantly lower than H.264resulting inlower costs of hardware and software implementation;(3) mastering the mainintellectual property rights, patent licensing model is simple and low cost. By virtue ofthese advantages, the AVS has gradually become a global scale standard, which ismost likely to be the defacto standard for second-generation audio and video codingstandard, and it will play an extremely important role in the future of videoapplications.However, due to the relatively late proposal, its application is not very mature,and still in the initial stage. With the vigorous support by national departmentsconcerned and many manufacturers, a complete industrial chain, from AVS encoder,AVS-IPTV system to the AVS decoder and the AVS decoder chip, is constructed. Along with development of relative industry, AVS has a broad application prospect.Though encoding complexity of AVS is lower than H.264, its computationalcomplexity is still high. So the real-time coding is facing challenges, and the efficientoptimization algorithm is needed to reduce the huge amount of calculation. In order toput AVS into the real-time applications, some procedures in AVS coding process areinvestigated that includes: intra prediction, inter prediction, motion estimation,transform and quantization, as well as its realization on the DaVinci platform. Themain works and achievements are as follows:First, in order to improve coding efficiency, SSIM is introduced into Lagrangiancost function to correct expression of distortion metric. Based on a large amount ofexperiments, empirical formula of Lagrangian parameter is established. The bit rate isaveragely reduced by13.22%. Especially, when the sequences have massivestationary blocks, the bite rate is saved more than30%with QP=10. At the same time,quality of reconstructed image is decreased by0.14%which could be ignored.Then, after the theory of the intra prediction and the process of the intra modedecision are analyzed, the fast intra prediction algorithm based on sum of absolutetransformed differences (SATD) criterion and spatial correlation is proposed tooptimize intra mode decision and reduce the complexity. First, some candidate modesare selected based on block's SATD which decreases the number of intra predictionmode. Then, mode correlation of neighbor blocks is used to reduce the complexity ofluminance blocks. Finally, early termination principle based on all zero blockdetection is adopted in process of intra mode decision for improving the speed ofmode decision. Encoding time of the proposed approach is averagely reduced by54.34%.Third, in order to reduce the calculation complexity of matching criterion,improved histogram based partial distortion search (IHPDS) algorithm is presented. Inthis algorithm, block is firstly divided into1×4sub-blocks and the complexity ofsub-blocks is analyzed by histogram. Then the calculation order of sub-blocks isdetermined based on the complexity. Finally, the threshold of accumulated partialdistortion is adaptively changed according to the complexity of sub-blocks to increase the probability of early termination of distortion calculation and effectively reduce thecomplexity of motion estimation. The computational costs of IHPDS are averagelyreduced by94%, taking no account of overhead computation. The actual motionestimation time of IHPDS is saved by68%on average.Furthermore, according to the characteristics of DCT transform coefficients thata substantial number of transform coefficients are quantized to zeros in video coding,in order to improve the detection rate of the all-zero blocks, all-zero block detectionthreshold based on the statistical characteristics of all-zero blocks is obtained. Andthen zero quantized DCT (ZQDCT) coefficients detection threshold is derived basedon Gaussian model, meanwhile ZQDCT coefficients detection algorithm based onGaussian model is proposed by combining the all-zero block detection threshold andZQDCT coefficients detection threshold. The DCT, Q, IQ, IDCT calculation isreduced and its time is averagely saved by20%.Finally, the video encoding hardware platform is designed based on DM6446,and AVS encoder is successfully ported to the DM6446. At the same time, thepackaging and integration of the algorithm based on DaVinci software framework hasbeen done, AVS video coding is realized on the DM6446.