The uses of texture in image analysis are widespread, ranging from remotely
sensed data to medical imaging to military applications. Image processing
tasks that use texture characteristics include classification, region segme
ntation, and synthesis of data. While there are several approaches availabl
e for texture modeling, the research presented here is concerned with stoch
astic texture models. Stochastic approaches view a texture as the realizati
on of a random field and are most useful when the texture appears noisy or
when it lacks smooth geometric features. The model introduced in this paper
is a subclass of Markov random fields (MRFs) called partially ordered Mark
ov models (POMMs). Markov random fields are a class of stochastic models th
at incorporate spatial dependency between data points. One major disadvanta
ge of MRFs is that, in general, an explicit form of the joint probability o
f the random variables describing the model is not obtainable. However, a p
opular subclass of MRFs, called Markov mesh models (MMMs), allows the expli
cit description of the joint probability in terms of spatially local condit
ional probabilities. We show how POMMs are a generalization of MMMs and dem
onstrate the versatility of POMMs to texture synthesis and pattern recognit
ion in imaging. Specifically, we give a fast, one-pass algorithm for simula
ting textures using POMMs, and introduce examples of heterogeneous models t
hat suggest potential applications for object recognition purposes. Then we
address an inverse problem, where we present results from a series of stat
istical experiments designed to estimate parameters of stochastic texture m
odels for both binary and gray value data. Although the applications in thi
s paper focus on imaging, in their most general form, POMMs can be found in
areas such as probabilistic expert systems, Bayesian hierarchical modeling
, influence diagrams, and random graphs and networks. (C) 1999 Pattern Reco
gnition Society. Published by Elsevier Science Ltd. All rights reserved.