Research On Blind Image Restoration Based On Improved Regularization Model

As a product of the information on the 21 st century, digital image processing has been widely used. Blind image restoration is an important part, which is to recover the original image from observations of degraded image through related technics. The general method is:establish a mathematical model to describe the degradation process firstly, and then solve the inverse problem, and obtain the reasonable estimate which is the closest to the original image. Since the related process belongs to the two dimensional deconvolution problem, it has the characteristic of solving the inverse problem from the mathematical point. But the inverse problem is often not qualitative, adding regularizations is an effective method.In this paper, in order to ensure that the image can preserve edges well when noises and blurs can be removed, we proposed a new model—modified total variation (MTV) model, which is based on the classical Rudin-Osher-Fatemi (ROF) model, we add the smooth term ||▽u||22 based on the total variation regularization term of the ROF model, so that it can avoid the staircase effect. Later, for removing different kinds of blurs and noises, we generalize the model, and proposed generalized modified total variation (G-MTV) model, in which the fidelity term is the combination of L1 and L2 norms. For the algorithm, in terms of the split Bregman algorithm has the characteristics of good stability and fast convergence rate, we apply this algorithm to construct theoretical conclusion and numerical experiments. Furthermore, as there is little or no knowledge about the point spread function in image blind deconvolution, we have to get the point spread function and the recovered image which is approximate to the original image simultaneously, it will cost much time. In order to fasten the speed of the algorithm, the Fast Fourier transform technical is applied.Different degraded images, blurs and noises are used in the experiments, and we derive some conclusions using MATLAB, and have a comparison between the models MTV and G-MTV, simulation results show the effectiveness of the models MTV and G-MTV. Meanwhile, the estimated PSF are obtained.