Research Of Image Model Based On Natural Image’s Statistal Priors And Sparse Priors

In the recent years, the natural image prior model has been a research focusin the field of image processing. Natural image has many features, such as localsmoothing, nonlocal similarity, statistical property, sparsity. There have manyways to extract prior information of natural image. The two striking methord ofthem are Fields of Experts (FoE) model based on the statistical properties ofnatural images and K-SVD method based on the sparsity of natural images.FoE model is a higher-order Markov random field model (MRF) establishedon the basis of Product of Experts (PoE) model, and it extracts image prior withfilters from lots of natural images. K-SVD is a sparse representation modelwhich makes use of the characteristics of that the natural images can berepresented in a sparse way, and it captures image prior with redundantdictionary from the image to be processed. There is no literature explain therelationship between these two methods. The paper focuses on the relationshipbetween the FoE filters and sparse bases (redundant dictionary) of K-SVD whichis a typical method of the sparse representation.FoE’s filters capture the component of natural images that should not appear,while the sparse bases of K-SVD grab the most likely component of the naturalimage. FoE’s filters and the sparse bases of K-SVD describe the natural imageprior information from two different directions, and contaion the different priorinformation. Comprehensive utilization of these two kind of prior information,can better to extract the prior information of natural images. The thesis that thisis two kind of different prior information is verified by experiment. Then a newmodel, which joints statistic prior and sparsity prior, has been put up. Theexperimental results of image denoising show that the new method gain goodresults than the FoE and K-SVD.