Blind deconvolution and super-resolution of low-resolution images and videos.

摘要

This dissertation presents novel approaches for blur deconvolution (BD) and super-resolution (SR) of low-resolution (LR) images and video sequences. SR is the process of reconstructing a high-resolution (HR) image/video by fusing information from one or a series of LR image(s)/video(s) degraded by various artifacts such as aliasing, blurring and noise. Our emphasis for reconstruction is on blind estimation which means that the point spread function (PSF) in each LR input is unknown and should be estimated along with the HR output. Also, SR reconstruction needs that the LR inputs are aligned together using a local/global 2D/3D registration method. In a BD problem, increasing the resolution is intended through deblurring and denoising operations. This means that aliasing removal is not considered in BD and so the input and output data have identical sizes.;In this dissertation, we consider different SR and BD approaches: SR from multiple images, BD from a single image or multiple images, SR/BD from a single video, and SR/BD from multiple videos. SR from a single image is a completely different approach and not studied in this work. Our approach is based on using the maximum a posteriori (MAP) framework to minimize a cost function based on the HR image and the blur(s). Regularization terms are defined in a way to smooth non-edge regions while preserving edges.