Efficient proposals in reversible jump MCMC for disk detection

Bayesian image analysis is a popular choice to address the uncertainty in an image. Statistical models in this context tend to be complex and require simulation based inference usually using Markov chain Monte Carlo (MCMC). However, MCMC algorithms for such models often suffer from long run-times. In this paper we develop an adaptive MCMC method which is guided by information extracted from the image via a generalized Hough transform. We illustrate the method for a simple disk recognition problem and show how it outperforms a standard MCMC approach.