Bayesian object identification

This paper addresses the image analysis problem of object recognition - locating and identifying an unknown number of objects of different types in a scene. The particular application in mind is the automatic labelling of cells in a microscope slide. High-level statistical image analysis has been the subject of much recent research activity (Baddeley & Van Lieshout, 1993; Grenander & Miller, 1994). The former of these approaches advocates marked point processes as object priors; the latter approach is built around the use of deformable template models. In this paper elements of both approaches are combined to handle scenes containing variable numbers of objects of different types. The complexity of the posterior distribution of interest, together with the variable dimension of the parameter space, mean that reversible jump Markov chain Monte Carlo methods are required (Green, 1995). The naive application of these methods here leads to slow mixing; we propose three strategies to deal with this. This first two expand the model space by introducing an additional "unknown" object type and the idea of a variable resolution template. The third strategy is to include classes of updates which provide intuitive transitions between realisations containing different numbers of cells by splitting or merging nearby objects. A novel point estimator for the number of objects together with their locations, shapes and types is suggested and applied to an example of microscopy data. SOME KEY WORDS: Bayesian inference; Deformable templates; Image analysis; Loss functions; Marked point processes; Markov chain Monte Carlo; Object recognition; Variable dimension distributions.