Nonstationary composite image modeling and two-source coding

A second-generation coding algorithm based on a nonstationary composite model for images is presented. The authors consider as sources the gross information describing the underlying structure of the image, and its fine or textural information. Median and range filters are used to segment the image into homogeneous and transition regions. The mean value function in each of these regions is estimated to provide the gross information. Estimation in the transition regions is done by means of space-varying unbiased estimators based on Legendre orthogonal polynomials. The textural information is characterized by means of linear prediction with space-varying coefficients. Parametrizing these coefficients in terms of known basis makes possible a recursive solution of the prediction problem. The coding procedure is based on coding separately the gross and the textural information. Examples illustrate the efficiency of the proposed coding procedure