4 - Champs de Markov cachés et estimation conditionnelle itérative

This work deals with the parameter estimation problem in hidden Markov fields . The principal goal is the comparison of methods deriving from a recent general procedure of estimation in the case ofhidden data, Iterative Conditional Estimation (ICE), with some existing algorithms. The paper starts with the recall of the importance of the hidden Markov fields estimation problem in unsupervised image segmentation. Then we compare ICE methods with sonie existing algorithms at different levels : principle, generality of models, difficulty of its implementation. The Stochastic Gradient of L. Younes and the Gibbsian EM of B . Chalmond are compared with ICE methods in some detail . The principle of ICE, different from principles of all existing methods, allows the conception of algorithms applicable in a quite general framework. Furthermore, the EM formulae can be obtained by ICE and, in this sense, it can be seen as particular ICE case .