Research On Image Restoration Based On Multivariate Finite Mixture Model

Image restoration technology has become a hot issue discussed by the academic area and the industrial area in the era of information.Image denoising methods based on multivariate finite mixture model have attracted much attention form experts in recent years.multivariate finite mixture model is a semi-parametric model between parametric model and non-parametric model based on mathematical theory that has a better robustness.It can be seen as a combination of multiple single models,and may approach any distribution while the quantity of single models is large enough.Gaussian Mixture Model is most widely used among them.Expected Patch Log Likelihood algorithm constructs the denoising model by considering GMM as patches' prior.Though this method performs well in noise removal,it still has room for improvement.The research carried out in this paper mainly includes the following parts:(1)An edge restoration regularization term is proposed to overcome the weakness that original EPLL method ignores structural information.Our proposed method divides the image into flat regions and edge regions by local variance,and measures the degree of edges being restored by local variance's difference between the restored image and the degraded image.Moreover,we add an edge restoration regularization term as the constraint condition of EPLL denoising methods.The experiment shows that proposed method can not only preserve more edge information,but also achieve advance in numerical value(PSNR).(2)A new adaptive parameter and a gradient fidelity term are proposed to overcome the weakness that edges and details are oversmoothed.The original restoration method considers regularization parameter as a constant so that we will remove noise with the same strength on the whole image,leading to the loss of edges and details.Then,we propose a restoration method combining the adaptive parameter and the gradient fidelity term.The proposed parameter removes noise as much as possible at flat regions while denoises as a weak strength at edge regions.Meanwhile the added gradient fidelity term can preserve more details and release staircase effect.