A Metropolis Sampler for Polygonal Image ReconstructionPeter

We show how a stochastic model of polygonal objects can provide a Bayesian framework for the interpretation of colouring data in the plane. We describe a particular model and give a Markov Chain Monte Carlo (MCMC) algorithm for simulating the posterior distribution of the polygonal pattern. Two important observations arise from our implementation of the algorithm. First, it is computationally feasible to use MCMC to simulate the posterior distribution of a polygonal process for moderately large problems (ie, 10000 data points, with polygonal patterns involving around 120 edges). Our implementation, which we would describe as careful, but unsophisticated, produces satisfactory approximations to the mode of the posterior in about 5 minutes on a SUN Sparc 2. Independent samples from the posterior take a few seconds to generate. The second observation is that the Arak process, a particular type of polygonal process , makes a wonderful debugging tool for testing shape simulation software. This is no mundane observation. Given the usually very complicated nature of such software it is essential to demonstrate that samples genuinely do come from whichever shape distribution is proposed. The Arak process has the advantage that many of its properties can be expressed in precise analytic terms. Computer programmes designed to sample mosaics formed from polygonal shapes can thereby be tested on this process by comparing computer output with exact theoretical expectations.

[1]  N. J. Lennes.  Theorems on the Simple Finite Polygon and Polyhedron , 1911 .

[2]  N. Metropolis,et al.  Equation of State Calculations by Fast Computing Machines , 1953, Resonance.

[3]  T. E. Harris.  The Existence of Stationary Measures for Certain Markov Processes , 1956 .

[4]  W. K. Hastings,et al.  Monte Carlo Sampling Methods Using Markov Chains and Their Applications , 1970 .

[5]  C. Preston.  Spatial birth and death processes , 1975, Advances in Applied Probability.

[6]  B. Ripley,et al.  Markov Point Processes , 1977 .

[7]  Graeme Williams,et al.  Program checking , 1979, SIGPLAN '79.

[8]  Donald Geman,et al.  Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images , 1984, IEEE Transactions on Pattern Analysis and Machine Intelligence.

[9]  J. Banavar,et al.  Computer Simulation of Liquids , 1988 .

[10]  D. Stoyan,et al.  Stochastic Geometry and Its Applications , 1989 .

[11]  Charles J. Geyer,et al.  Practical Markov Chain Monte Carlo , 1992 .

[12]  A. Baddeley,et al.  Stochastic geometry models in high-level vision , 1993 .

[13]  D. Frenkel.  Advanced Monte Carlo techniques , 1993 .

[14]  van Marie-Colette Lieshout,et al.  Stochastic annealing for nearest-neighbour point processes with application to object recognition , 1994, Advances in Applied Probability.

[15]  Ulf Grenander,et al.  General Pattern Theory: A Mathematical Study of Regular Structures , 1993 .

[16]  P. Clifford,et al.  Point-based polygonal models for random graphs , 1993, Advances in Applied Probability.

[17]  Michael I. Miller,et al.  REPRESENTATIONS OF KNOWLEDGE IN COMPLEX SYSTEMS , 1994 .

[18]  Geir Storvik,et al.  A Bayesian Approach to Dynamic Contours Through Stochastic Sampling and Simulated Annealing , 1994, IEEE Trans. Pattern Anal. Mach. Intell..

[19]  C. Geyer,et al.  Simulation Procedures and Likelihood Inference for Spatial Point Processes , 1994 .

[20]  A. Sokal.  Monte Carlo Methods in Statistical Mechanics: Foundations and New Algorithms , 1997 .