The Study Of Decorrelation Techniques For Video Coding On Heterogeneous Multi-Core Processor

Video coding is currently the worldwide hotspot of research institutes and industrial applications. In video coding, by decorrelating redundancy in video contents, video compression can be well achieved. Therefore, research on decorrelation techniques of video coding has great theoretical value and practical significance. Heterogeneous multi-core processor (HMP) gradually becomes the mainstream of future processors. Efficient video coding on heterogeneous multi-core processor (HMP) has also drawn considerable attention. In this work, we focus on improving rate-distortion (RD) performance of video coding on HMP. A novel intra prediction strategy as well as several transform tools is presented in this work. Moreover, parallel motion estimation algorithm is also studied.Considering that DC (Direct current) prediction mode in intra prediction is unsuitable for smoothly-varying area, a distance based weighted prediction (DWP) is proposed. A linear prediction model is built based on the inverse relation of correlation to distance between two pixels. To reduce computational complexity, especially the cost of hardware implementation, integer DWP (iDWP) is further proposed. Experiments demonstrate that significant improvements on RD performance can be achievable by DWP and iDWP with a small amount of computation complexity increase.Since the residual signals of different intra prediction modes exhibit dissimilar energetic distribution, it is difficult for DCT (Discrete cosine transform) with fixed transform matrix to achieve ideal decorrelation performance. Though KLT (Karhunen-Loève transform) is optimal for transform coding, its coding performance is data-dependent. Due to the relatively accordant energetic distribution of residual signals of the same prediction mode, an optimal frequency matching (OFM) algorithm is proposed to train a unique KLT matrix for each intra prediction mode, which can avoid the extremely high computational complexity of real-time KLT matrix training. The experiments show that the coding performance of the trained KLT matrices is stable and superior to DCT.Variable block size prediction is an important coding tool adopted in video coding. However, the mismatch between prediction block and DCT matrix not only degrades decorrelation performance but also brings blocky artifacts inside the block bigger than 4×4. To overcome these defects, M-channel filter bank (MCFB) is proposed to decompose residual blocks after prediction. Decomposition using M-channel filter bank possesses the following three merits: firstly, blocky artifacts inside the prediction block can be alleviated due to the inherent property of convolution; secondly, block-based RD optimization can be performed; finally, the frequency characteristics of the transformed coefficients are similar to that of DCT. The experimental results show that MCFB can significantly outperform DCT in both subjective and objective video quality. To reduce the computational complexity of transformation using MCFB, an integer M-channel filter bank is also constructed.Motion estimation in video coding is of enormous computational complexity and the GPU (Graphics processing unit) in HMP is utilized to accelerate motion estimation. However, the parallel computing mode of GPU makes current macroblock unable to take neither motion nor mode information of its neighbor macroblocks as reference for motion estimation, which accordingly disables RD optimization for best motion vector selection. As a compromise, SAD (Sum of absolute difference) is the only criterion to determine the best motion vectors. In this study, Collocated macroblock based motion estimation (CMME) algorithm is presented, which takes the motion information of the collocated macroblock in the previous frame as reference to estimate the motion vector cost. Extensive experiments demonstrate that up to 0.8dB PSNR improvements can be achieved by using CMME with little compleixity increase. And the CMME algorithm is very suitable for video sequences with slower motion and low bitrate coding.In a word, through the thorough study of decorrelation methods, novel transform and intra prediction techniques as well as related integer implementation scheme are proposed to improve RD performance. In addition, motion estimation under parallel computing mode is also studied in this work.