Video Coding Approach Based On Low-rank Matrix Reconstruction

Spatial redundancy and temporal redundancy extensively exist in video, and howto reduce these redundancies is the important part of video coding research. Theexistence of spatial redundancy and temporal redundancy in video makes it possibleto use the low-rank matrix model to present the video signal. Viewing the theory oflow-rank matrix reconstruction as a guide, this paper proposes a lossless intra codingmethod based on low-rank matrix reconstruction and an inter prediction method basedon low-rank matrix reconstruction, which are expected to reduce the spatial andtemporal redundancy further.There is usually a high correlation between pixels that are close to each other in aframe，high correlations should also exist between the current block and its predictionblock which is generated through calculating the adjacent reconstructed pixels of thecurrent block, and these correlative blocks can be presented as a low-rank matrixmodel. In the research of lossless intra coding method based on low-rank matrixreconstruction, this paper proposes an approach for the low-rank matrix generationand transfers the intra prediction problem as a low-rank matrix reconstructionproblem, and then the final prediction can be formed by solving the low-rank matrixreconstruction problem. The new method has the advantage of higher accuracy forintra prediction, which helps to reduce the bit-rate further, and the advantages of thismethod will become more significant when the video has more flat image areas.Experimental results show that the bit-rate saving of the new intra prediction methodcan reach up to4.97%compared with H.264/AVC in lossless coding environment.High correlations exist between frames of video, and it is possible to find someblocks in reference frames which are similar to the current block. In the research ofinter prediction method based on low-rank matrix reconstruction, for exploiting thecorrelations among similar blocks fully, this paper constructs a low-rank matrix modelwhich is based on those similar blocks. Through the construction of the low-rankmatrix model, the inter prediction problem is transferred as a low-rank matrixreconstruction problem, and then the final prediction can be formed by solving thelow-rank matrix reconstruction problem. The new inter prediction method whichbased on low-rank matrix reconstruction need not to code motion information, and atthe same time it has a higher quality of image reconstruction. Experimental resultsshow that the new method can receive rate reduction up to9.91%compared withH.264/AVC, the bit-rate saving achieved from the new method would be particularly significant in the case of high bit-rate.