Blind Image Restoration Based On Inhomogeneous Gaussian Markov Random Fields Model

In this paper, we study a new blind image deconvolution method based on the inhomogeneous Gaussian Markov random field (IGMRF) model of image, which is made up of two blocks.In the first block, a modified generalized Cross validation blur identification method was proposed. Specifically, the Discrete Periodic Radon Transform (DPRT) was brought into the image restoration problem. So the original two dimensional autoregressive moving average (ARMA) model of degraded image was converted to its one dimensional counterpart, and only the parameters of 1-D ARMA process was estimated by means of Generalized Cross Validation (GCV). Compared with its 2-D counterpart, the 1-D problem is easier to solve. Then, Inverse Discrete Periodic Radon Transform (IDPRT) is employed to get the blur parameters which are included in the parameters of ARMA model, so the blur identification problem was solved.The second component is a new de-convolution method based on IGMRF model of image, which can adapt to the various characteristics of different regions in images. So the de-convolution method can suppress more noise in homogeneous region of image and recover more image detail in inhomogeneous regions. Because human visual characteristic shows that it is more sensitive to noise in flat region than to noise in inhomogeneous regions where there are full of details. So the results obtained are appreciating. Besides, because there are many parameters in IGMRF model, they were recommended to not be estimated in the conventional method which is by means of Gibbs sampler, but to be done approximately. The approximated estimation method is efficient and can produce good results as well. The speed of overall algorithm is very fast.At last, the simulation results have shown that the new image restoration method is very validity and that de-convolution result are very delicate to the parameters in the algorithm which should be estimated properly.