A Contextual Maximum Likelihood Framework for Modeling Image Registration

Abstract

We introduce a novel probabilistic framework for image registration. This framework considers, in contrast to previous ones, local neighborhood information. We integrate the neighborhood information into the framework by adding layers of latent random variables, characterizing the descriptive information of each image. This extension has multiple advantages. It allows for a unified description of geometric and iconic registration, with the consequential analysis of similarities. It enables to arrange registration techniques in a continuum, limited by pure intensity-and feature-based registration. With this wide spectrum of techniques combined, we can model hybrid registration approaches. The probabilistic coupling allows further to deduce optimal descriptors and to model the adaptation of description layers during the process, as it is done for joint registration/segmentation. Finally, we deduce a new registration algorithm that allows for a dynamic adaptation of the description layers during the registration. Excellent results confirm the advantages of the new registration method, the major contribution of this article lies, however, in the theoretical analysis.