Research On Content-aware Optimization Algorithm For High Efficiency Video Coding(HEVC)

High Efficiency Video Coding(HEVC)is the latest video coding standard jointly developed by the International Telecommunication Union and the International Organization for Standardization.After several years of development,HEVC has become the world’s largest video coding standard with the largest number of companies and scientific research institutions participating in and the largest international influence.Most of the current optimizations for HEVC do not consider the differences between video contents,so that there is considerable visual redundancy and time-domain redundancy in the encoded bit-stream.This dissertation studies content-aware optimization algorithm for HEVC.Specifically,the differences between video contents are utilized in coding unit(CU)splitting,transform/quantization,rate control and rate distortion optimization to further remove visual redundancy and time-domain redundancy in videos and optimize the encoding speed.The main contributions of this dissertation are as follows:1)A fast coding algorithm based on visual saliency is proposed,aiming at the problem of high complexity of CU splitting and mismatch between quality distribution and human vision.Firstly,a saliency detection algorithm is designed by using coding features.Then,more proper quantization parameter(QP)is selected for CU according to saliency,which makes the quality distribution of video content more consistent with the human vision characteristics.Finally,according to the average saliency of 32×32 CU and the distribution of saliency,a fast CU splitting algorithm is designed,which effectively reduces the complexity of CU splitting.Experimental results show that the proposed fast encoding algorithm can reduce the encoding time by an average of 42.9% and reduce the bitrate by an average of 12.67%(QP = 32)compared with HEVC test model HM-13.0.2)A coding tree unit(CTU)level rate control algorithm based on visual saliency is proposed,aiming at the problem of mismatch between video quality distribution and human vision.Firstly,saliency is used as the weight of distortion,thereby improving the ability of salient areas to obtain coding bits.Then,the CTU-level optimal coding bit allocation scheme is built according to the weighted distortion and solved by the Lagrangian optimization.Finally,coding parameters for CTUs are solved based on weighted cumulative error allocation.Experimental results show that the proposed rate control algorithm can effectively improve the quality of salient areas and improve the weighted peak-signal-to-noise-ratio by an average of 0.4999 d B and 0.2431 d B with low-delay configuration and randomaccess configuration respectively compared with HM-16.14.3)An all zero block detection algorithm based on the quantization level of the maximum magnitude(QLMM)is proposed,aiming at the problem that frequency domain distribution of transform coefficients is not fully considered in current all zero block detection algorithms.Firstly,it is proved that "QLMM equals to 0" is the(approximate)sufficient and necessary condition for “a transform block is an all zero block”.Based on this condition,a recursive coefficient estimation method based on Walsh transform is designed at first.During coefficient estimation,most of the high-frequency coefficients are ignored due to the fact that the human vison is not sensitive to high-frequency distortion.Then,the method to detect all zero blocks in the uniform quantizer is designed according to the threshold of quantization level.Finally,the method to detect all zero blocks in the rate distortion optimized quantizer is designed according to the difference of rate distortion cost between zero QLMM and non-zero QLMM.Experimental results show that 94.6% 16×16 AZB and 90.4% 32×32 AZB in the uniform quantizer and 95.9% 16×16 AZB and 95.4% 32×32 AZB in the rate distortion optimized quantizer can be detected by the proposed algorithm.4)A sub-frame layer rate distortion optimization algorithm based on inter-frame dependency of contents is proposed for low-delay hierarchical prediction structure,aiming at the problem that the difference of inter-frame dependency between contents within a frame is not distinguished.Firstly,the rationality of sub-frame layer rate distortion optimization is demonstrated by analyzing the difference of inter-frame dependency between sub-frames.Then,the inter-frame dependency model suitable for frames and sub-frames is established based on Laplace distribution and distortion propagation chain.Based on the established inter-frame dependency model,the rate distortion optimization algorithm for sub-frame layers in L1-L4 is designed using Lagrangian optimization.In addition,to solve the problem that there is only one L0 frame in the low-delay hierarchical prediction structure,an adaptive L0sub-frame rate distortion optimization algorithm is proposed.Experimental results show that the proposed sub-frame layer rate distortion optimization algorithm can reduce the bitrate by7.0% on average compared with HM-16.14.