Parameter estimation and applications of a class of Gaussian image models

This paper discusses variations of a model of images and develops algorithms for estimation of all the parameters from the raw image data. The model is suitable for some cases of (1) lossy image compression and realistic reconstruction, (2) texture synthesis and identification, (3) classification of remotely sensed data, and (4) analysis of medical images. Each pixel in the image is modeled as an element of a set of very few known intensity levels (henceforth called pixel-classes) plus an independent zero mean Gaussian random variable. Different statistical structures in the two dimensional lattice of pixel-classes lead to variations in the model. The image representation problem corresponds to estimation of the parameters of the discrete random field formed by the pixel classes, and the parameters of the additive Gaussian field. The authors discuss variations of the model and corresponding applications, and develop convergent estimators for all parameters