Statistical trajectory of an approximate em algorithm for probabilistic image processing

Kazuyuki Tanaka, D. M. Titterington

Research output: Contribution to journalArticlepeer-review

15 Citations (Scopus)


We calculate analytically a statistical average of trajectories of an approximate expectation-maximization (EM) algorithm with generalized belief propagation (GBP) and a Gaussian graphical model for the estimation of hyperparameters from observable data in probabilistic image processing. A statistical average with respect to observed data corresponds to a configuration average for the random-field Ising model in spin glass theory. In the present paper, hyperparameters which correspond to interactions and external fields of spin systems are estimated by an approximate EM algorithm. A practical algorithm is described for gray-level image restoration based on a Gaussian graphical model and GBP. The GBP approach corresponds to the cluster variation method in statistical mechanics. Our main result in the present paper is to obtain the statistical average of the trajectory in the approximate EM algorithm by using loopy belief propagation and GBP with respect to degraded images generated from a probability density function with true values of hyperparameters. The statistical average of the trajectory can be expressed in terms of recursion formulas derived from some analytical calculations.

Original languageEnglish
Article number007
Pages (from-to)11285-11300
Number of pages16
JournalJournal of Physics A: Mathematical and Theoretical
Issue number37
Publication statusPublished - 2007 Sept 14


Dive into the research topics of 'Statistical trajectory of an approximate em algorithm for probabilistic image processing'. Together they form a unique fingerprint.

Cite this