Blind Image Deconvolution: Model and Computation

Abstract

This dissertation aims at answering some fundamental questions in blind image deconvolution, and presenting state-of-the-art computational methods. The first part focuses on the mathematical models and computational methods fundamental to blind image deconvolution. We firstly introduced a new image prior for characterizing clear images with sharp edges based on wavelet frame and $\ell_1$ norm related sparsity-prompting prior. Secondly a robust non-blind image deconvolution method is developed to handle the errors in kernel which could cause great ringing artifacts if not handled properly.

The second part is devoted to the development of practical image deblurring systems. An efficient two-stage approach to remove spatially-varying motion blurring from a single photo is designed to solve the non-stationary image blurring. For motion blurring with non-linear camera response function, we developed a dual-image approach using a single-shot mode available in commercial cameras: a high-resolution image in JPEG format and a low-resolution image in RAW format.