Abstract
In this thesis, we propose a novel image restoration framework for restoring images degraded by unknown motion blurs from a single image. Our approach takes advantage of the bi-level image patches to estimate the blur kernel. The patches which contain only two-grayscale regions are first selected. Then we propose a method to find an initial guess for our algorithm. This method includes searching a threshold value based on a cost function and utilizing an illumination model to account for small illumination variations. Afterwards, we propose a probabilistic model which combines both blur kernel estimation and non-blind bi-level image deconvolution into a single maximum a posteriori (MAP) formulation. An alternating minimization algorithm is developed to iteratively refine both the blur kernel and the bi-level blurred patches. Finally, we apply Richardson-Lucy (RL) deconvolution to restore the entire image by using the estimated blur kernel. Some experimental results on the deblurring of simulated and real blurred images are given to demonstrate the performance of the proposed blind motion deblurring algorithm.