Improvements to sparse signal processing in compressive sensing and other methods

Compressive Sensing (CS), as a newly developed branch of sparse signal processing and representation approaches, has quickly found various applications in a large number of research topics in modern digital signal processing area. This work is devoted to investigating some effective CS and sparse si...

Full description

Saved in:
Bibliographic Details
Main Author: Huang, Honglin
Other Authors: Anamitra Makur
Format: Theses and Dissertations
Language:English
Published: 2012
Subjects:
Online Access:https://hdl.handle.net/10356/50358
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Compressive Sensing (CS), as a newly developed branch of sparse signal processing and representation approaches, has quickly found various applications in a large number of research topics in modern digital signal processing area. This work is devoted to investigating some effective CS and sparse signal processing schemes and approaches, and also to adapting the sparse signal processing idea and CS framework to some specific applications such as object-based surveillance video compression. In compressive sensing, the sampling strategy and reconstruction algorithms are two major components. In this thesis, we first investigate the optimization of the sampling matrix for effective CS performance. An optimizing method, by using a simple polynomial shrink function and a detector to estimate the convergence to stop the iterations early, is proposed to provide better CS performance. A number of experimental simulations are presented to demonstrate the optimized measurement matrix’s effectiveness. Then, a novel Backtracking-based Adaptive Orthogonal Matching Pursuit (BAOMP) method is proposed to effectively reconstruct or approximate the sparse solutions for CS and other sparse representation problems. Different from other Orthogonal Matching Pursuit (OMP)-type algorithms, the proposed method incorporates a backtracking step to more carefully choose the reliable support set, and at the same time, it does not require the signal’s sparsity level to be known before reconstruction. Various experiments on exact sparse signal reconstruction case, noisy signal or noisy measurement approximation case, and two-dimensional (2-D) compressible signal approximation case are illustrated to show the better performance than that of other known OMP-type methods. Furthermore, as an application of CS approach and sparse signal processing idea, an object-based surveillance video compression system based on CS is proposed, in which the sparse object error after motion compensation is coded by using CS scheme. In the proposed system, first the front-moving objects are segmented from background, then these are object-based compensated from previous reconstructed frames, and finally the object error blocks are encoded by CS approach and quantized to transmit or store. Extensive experiments show the proposed system’s efficiency. Finally, a simple iterative reconstruction method based on Projection Onto Convex Sets (POCS) is designed to effectively encode the object error. Firstly, we consider the reconstruction performance of one-dimensional (1-D) Autoregressive (AR) source signal, and then use a natural image as the input signal to do some experiments to illustrate the proposed method’s performance on image reconstruction. Finally, as an application of the iterative reconstruction method, a novel sparse signal compression scheme based on this iterative method is presented.