Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract


Bayesian inference by solving fixed-point problems: Markov random field models and Bethe tree approximations


Professor Peter C. Doerschuk
Electrical Engineering and Computer Science, Purdue University


One approach to stochastic pixel-level models for images is Gibbsian Markov random fields. Such models are suitable for describing a label, gray level, or hyperspectral vector at each pixel. Optimal estimators for segmentation or restoration problems based on such Gibbsian MRF models and noisy measurements at each pixel usually must be computed by simulation. A non-simulation approach based on Bethe tree approximations to the graph underlying the Gibbsian MRF will be described. The relationship between this approach and Belief Propagation on loopy graphs will be discussed. Use of this approach for spatial pattern classification of passive multispectral optical agricultural remote sensing data will be described.



Problems with this site should be emailed to jonesb@mit.edu