Estimating Parameters of Gibbs Models for Image Processing and Analysis

V.N. Vasyukov and D.V. Goleshchikhin (Russia)


Image processing and analysis, Gibbs random fields,estimating parameters, texture image, line process


A regular method is presented for estimating potentials of finite-state Gibbs models, used in image processing and analysis. The method is based on using sufficient statistics of Gibbs distribution, describing finite-state random fields. A finite-state field describing texture images, and a line process describing hidden level of a compound hierarchical Gibbs model, used for gray and color image restoration, are examples of finite-state fields. Estimation algorithms are based on solving high-dimensional linear equation systems. Amount of all possible different field configurations on a current site neighborhood defines quantity of equations in the system. In texture analysis, configurations of the image observed on the current site neighborhood are used in estimation. Estimated values are used for texture analysis. In restoration of images based on compound model, realizations of the line process are obtained by contour detection on any typical images of the class, and those realizations are used as source material for potential estimation. Estimated values are used for adjustment of the model used for restoration. Restoration quality is enhanced by using the hidden level, and estimation of this level obtained as a by-product, may be used for contour analysis of images. Experimental results are presented that illustrate efficiency of the algorithms.

Important Links:

Go Back