The Application Of The Sparse Representation In Image Recognition

Sparse representation, also known as compressive sensing, is a hot topic in image recognition, computer vision and numerical calculation. Since sparse representation is in widespread use while it is hard to be solved effectively, more and more algorithms against sparse representation are proposed. This thesis introduces the definition and background of sparse representation, and point out some problems about the current sparse representation algorithms, including high cost of computation, low convergence speed and difficult to adapt to the scale ratio of the dictionary matrix. Besides, the current algorithms do not consider about the parallelism, which causes the waste of computing resource on parallel platform. We focus on the computing problems in sparse representation and bring out solutions to the algorithms and implementations to optimize solving sparse representation. The contributions of this thesis include:1. We propose an optimized algorithm, which has faster converged speed, less iteration and good adaptability to the scale ratio. Furthermore, we divide the sparse problem into multi-variable and make the subproblems parallel. This new method suits the parallel platform commendably.2. We realize our algorithms both on Matlab and Graphic Processing U-nit platform, and design a series of experiment, including numeric data and realistic data, to prove the advantages of our algorithms on running speed and parallelism.3. In view of some expanded form of sparse representation, we discuss and derivate them in detail, and provide the algorithms to solve them.