Techniques for reducing the computation effort of matching pursuits

Abstract

Matching pursuits is a method which utilises an iterative procedure to project a given waveform onto a particular dictionary. It has been successfully applied in different areas such as data compression since its introduction in 1993. Recent research has demonstrated that it helps to provide a very good coding performance at very low bit-rate video compression. However, its realization effort is so high that it is nearly impossible for real-time applications. This thesis is primarily concerned with how to realize matching pursuits efficiently and effectively. Since matching pursuits is basically a core module widely used in various applications, a fast matching pursuits algorithm would directly speed up the realization of a number of applications. In this thesis, a technique for convening a matching pursuits problem into a vector quantization problem is presented. This technique allows one to realize matching pursuits with a lot of well-developed fast vector quantization algorithms. As a result, a number of fast matching pursuits algorithms can be obtained. In this thesis, the partial searching algorithm, a classical technique for realizing vector quantization is used as an example to show how to realize matching pursuits and orthogonal matching pursuits, a variant of matching pursuits, with the proposed technique in an efficient way. A fast two-stage algorithm based on a recently proposed two-stage filtering structure for realizing matching pursuits is also presented in this thesis. A complete analysis of the structure is provided and the optimization of the two stages is investigated. Simulation results show that a significant improvement of the performance of the structure can be achieved.