Research And Application Of GPCA In Scalable Video Coding

The development of digital storage, communication and Internet makes possible excessive amount of video data appears in people's life. Video has become one of the most important media where people get information. Although the bandwidth of network, capacity of disk and memory keep growing, video compression is the prerequisite of most multimedia services. Research on efficient compression algorithms is worthy of efforts. Besides, the demand on adaptive coding schemes to fit the varying bandwidth and different terminals, to support reliable wireless transmission leads to the development of scalable video coding, which aims to achieve optimal video quality with available resource. On the other hand, applications of video have extended to areas far beyond video play. Analysis and understanding of video data become very popular these years.The contribution of this paper mainly lies on introducing Generalized Principal Component Analysis (GPCA) to the framework of video coding to replace the traditional Discrete Cosine Transform (DCT) and accomplishing Signal Noise Ratio (SNR) scalability. Besides, motion vectors, the product of motion estimation, are exploited to model the dynamic scenes.This paper is organized as follows: The first chapter illustrates the importance of video coding in multimedia communication and other services, introduces the evolution of two international standards MPEG and H.26x. Chapter 2 briefly describes the framework of video coding, and analyzes some features of the booming technology H.264. Chapter 3 introduces the concept of scalable video coding (SVC), mainly focused on Motion Compensation Temporal Filter (MCTF) and its extension to H.264. Chapter 4 first briefly describes the theory of GPCA, and then introduces it into video coding. The goal of GPCA is to estimate a hybrid linear model for given data. Compared with Principle Component Analysis (PCA), the hybrid model has the advantage that it expresses the data in a more compact way and also exposes the distribution of the data. We use GPCA to transform the residual data, implement a SNR scalable scheme and discuss several means to reduce the computation complexity. In chapter 5, we use motion vectors to model the scenes so as to detect moving objects. The experimental results are given in chapter 6. At the end, we summarize this paper and discuss some future research topics.