Abstract: The iterated conditional expectations (ICE) procedure is studied under the Ising model. If the parameter in the Ising model is not well chosen, experiments show that the ICE suffers from problems of over- and under-smoothing. To detect and avoid these problems, a criterion is proposed in which the performance of the iterations is predicted. Based on this criterion, we have an algorithm that eliminates the over- and under-smoothing iterations, adjusts the parameter in the Ising model, and finally leads to the ``right'' parameter and the ``right'' number of iterations. The ICE procedure is successfully modified by the proposed algorithm according to the experiments.
Key words and phrases: Markov random field (MRF), iterated conditional expectations (ICE), Ising model, image restoration, stopping rule.