Development of image restoration method using hierarchical MRF model

Abstract

In this thesis, hierarchical Markov random field model-based method for image denoising and restoration is implemented. This method employs a Markov random field(MRF) model with three layers. The first layer represents the underlying texture regions and the second layer represents the noise free image. The third layer represents the observed noisy image. Iterated conditional modes (ICM) algorithm is used to find the maximum a posterior (MAP) estimation of the noise free image as well as the texture region field. This method can effectively suppress additive noise and restore image details. A noise-free gray-scale image is considered. Then Gaussian noise is applied to the image so that the image becomes noisy. The aim is to remove this noise from the image. Image is considered as the combination of disjoint texture regions, and a three-layered hierarchical MRF is used to model the image. The algorithm starts with choosing the number of the regions, iteration time and a MRF neighborhood system. Initially, the local variance of all the pixels is calculated considering a (3*3) window sliding through the image. K-means clustering is applied to the local variance feature image. The MRF parameters are estimated and then the clustered images and the noise-free image are updated using the ICM algorithm and the process is repeated till the MRF parameters become constant. The output obtained is the noise-free image. The method used employs a three-layered MRF model which can express both smooth and texture signals. The advantage of hierarchical MRF model is that the texture information of the image is considered while the process of denoising, so that the edge information and other interesting structures of the image are not lost and the image is restored efficiently.