Abstract

Super-Resolution (SR) is a generic term, referring to a series of digital image processing techniques in which a high resolution (HR) image is reconstructed from a set of low resolution (LR) video frames or images. In other words, a HR image is obtained by integrating several LR frames captured from the same scene within a very short period of time. Constructing a SR image is a process that may require a lot of computational resources. To solve this problem, the SR reconstruction process involves 3 steps, namely image registration, degrading function estimation and image restoration. In this thesis, the fundamental process steps in SR image reconstruction algorithms are first introduced. Several known SR image reconstruction approaches are then discussed in detail. These SR reconstruction methods include: (1) traditional interpolation, (2) the frequency domain approach, (3) the inverse back-projection (IBP), (4) the conventional projections onto convex sets (POCS) and (5) regularized inverse optimization. Based on the analysis of some of the existing methods, a Wavelet-based POCS SR image reconstruction method is proposed. The new method is an extension of the conventional POCS method, that performs some convex projection operations in the Wavelet domain. The stochastic Wavelet coefficient refinement technique is used to adjust the Wavelet sub-image coefficients of the estimated HR image according to the stochastic F-distribution in order to eliminate the noisy or wrongly estimated pixels. The proposed SR method enhances the resulting quality of the reconstructed HR image, while retaining the simplicity of the conventional POCS method as well as increasing the convergence speed of POCS iterations. Simulation results show that the proposed Wavelet-based POCS iterative algorithm has led to some distinct features and performance improvement as compared to some of the SR approaches reviewed in this thesis.