Techniques of deconvolution, interpolation and super-resolution for high-resolution image reconstruction

Abstract

This thesis investigates how to produce a high quality, high-resolution image from low quality, low-resolution images. The generic name of high-resolution image reconstruction covers related subjects of deconvolution, interpolation, and super-resolution. In the first part, we attempt to address blind deconvolution by assessing the relevance of parametric blur information, and incorporating the knowledge into the parametric double regularization scheme. Further, an iterative algorithm based on multichannel recursive filtering is proposed to address multichannel image deconvolution. The second part of this thesis deals with image resolution enhancement from single/several low-resolution observations. The image interpolation is formulated as a regularized least squares solution of a cost function. We derive the optimal solution using a combined framework of Kronecker product to reduce the computational cost greatly. The proposed bispectrum algorithm utilizes the characteristics of higher-order statistics to suppress Gaussian noise for subpixel image registration. The main contribution of blind super-resolution is the development of multichannel blind deconvolution to estimate the unknown point spread functions, and its integration into the super-resolution scheme to render high-resolution images.