A data-driven Bayesian sampling scheme for unsupervised image segmentation

A Bayesian scheme for fully unsupervised still image segmentation is described. The likelihood function is constructed by assuming that the grey level at each pixel site is a realization of a Gaussian random variable of unknown parameters, there being an uncertain number of distinct Gaussian classes in the image. Spatial connectivity between pixels is encouraged via a Markov random field prior. The task of identifying the model parameters and recovering the underlying class label at each site (i.e. segmentation) is accomplished using a novel reversible jump Markov chain Monte Carlo (MCMC) scheme. This scheme explores the space of possible segmentations via proposals that are driven by the actual image realization-so-called data-driven proposals. The aim is to (i) induce good mixing in regions of high probability, and (ii) to optimize the acceptance probability of the proposals. A key development is a stochastic version of a recursive labeling algorithm which has been used in previous work for fast image region splitting. In the current stochastic context, it yields fast and effective split and merge proposals. The performance of the novel MCMC scheme is illustrated in simulation