Contextual estimators of mixing probabilities for Markov chain random fields

Abstract

This paper discusses the estimation of proportions of classes from an image. Classification methods are compared with likelihood methods and the importance of contextual information is discussed. The joint distribution of classes at neighbouring sites is modelled by a Markov chain random field. The class attributes are estimated from training sets and unclassified observations. The effect of biased class means is reduced with a stochastic model of the bias. Contextual likelihood methods yield better results than non-contextual methods.