Advanced video coding based on matrix decomposition

• Main Author: Gu, Zhouye
• Other Authors: Lin Weisi
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Data::Coding and information theory
• Online Access: https://hdl.handle.net/10356/55344
• Institution: Nanyang Technological University

Video coding has been widely used in everyday work and life. Most of the video coding technologies are based on a set of principles that reduce the redundancy in digital video, including mainly temporal, spatial and statistical redundancies. In this study, we explore new methodology for further redundancy reduction. Combining the matrix decomposition algorithms, we have proposed three matrix decomposition based techniques that further enhance compression performance, when compared with the relevant state of the art technology. Firstly, we propose a temporal redundancy reduction scheme based upon Two-Dimensional Singular Value Decomposition (2D-SVD) without resorting to motion estimation (ME). Its computational complexity is much lower than that of ME-based video coding scheme. By exploring the energy compaction property of 2D-SVD coefficient matrices, high coding efficiency is achieved compared with other non-ME based methods. Secondly, for spatial redundancy reduction, the existing video codecs use Discrete Cosine Transform (DCT), which is originally designed based on the spatial characteristic of natural image pixels, for predicted residual pixels compression; Since the spatial characteristic of predicted residual pixels differs from that of natural image pixels, we develop a new orthogonal transform—Rotated Orthogonal Transform (ROT) that yields better compression on the prediction residual than the DCT. Lastly, for statistical redundancy reduction, the entropy coders that originally designed for lossy video coding are also used for lossless video coding in the current standard video coding scheme. Since the coefficients generated by lossy and lossless coding processes have different statistical distributions, we therefore propose a new methodology based on Mode Dependent Template (MD-Template) and scan order for better intra lossless video coding. The algorithms proposed in this thesis are developed from three different matrix decomposition problems. The have been validated on a large number of standard video sequences. We have performed careful experimental analysis and demonstrated that the proposed techniques overall perform better than the existing relevant methods.