Massively Parallel Image Segmentation on the Connection Machine

the following way: This paper is devoted to a new deterministic and massively parallel alnorithm for combinatorial o~timization in a Markov Random F i e l i First, tbe a posteriori of a tentative labeling, Now, let R ~ M A be defined by: defined in terms of a Markov Random Field is generalized to continuous labelings. This merit function of probagilistic vectors is then convexified by changing its domain. Global optimization is performed, and the maximum is tracked down while the original donlain is restaured. We analyse in details the parallel implementation of this algorithm on the CM2, in terms of speed efficiency, memory and communication requirements. Comparison with other classical algorithms is made on a contextual pixel classification problem.