Parsimonious texture representation employing binary Gaussian processes and binomial linear prediction

An unsupervised texture synthesis method operating in a small dimensionally parameter space is presented. It consists of two stages. In the first stage, the morphological structure of a given prototype is reproduced using a hard limited Gaussian random field. The second stage is a simple low-pass filter for texture reconstruction from a binary input. The model identification process is described and synthetic texture examples are reported. (C) 2001 Elsevier Science B.V. All rights reserved.