Abstract
Single image de-convolution is an ill-posed problem due to the number of unknowns to be estimated is far larger than the number of available observations given by the input data. Experts of this eld have proposed several assumptions on the prior to help solve the problem. Most of the previous approaches concentrate on how to use the heavy-tailed gradient magnitude distribution. Although it seems that previous approaches have already exploited all the information from a single image, there are indeed lots of issues worthy of investigation. In this thesis, we present a new image-based characteristic which is believed to be discriminative in identifying blur and sharp image. We study this assumption with various kinds of natural images, and the criterion used to model it is proven to be useful throughout extensive experiments. The optimization strategy we employed follows the suggestion made by by many researchers in recent years: We separate the optimization into two steps, one for image prediction, and the other for blur kernel estimation. Preliminary results of kernel estimation have shown the potential power of this proposed natural image statistics. The primary contribution of this thesis is hence the introduction and analysis of a new image prior.