Abstract
A new information measure for probability distributions is presented; based on it, a similarity measure between images is derived, which is used for constructing a robust image registration algorithm based on random sampling, similar to classical approaches like mutual information. It is shown that the registration method obtained with the new similarity measure shows a significantly better performance for small sampling sets; this makes it specially suited for the estimation of non- parametric deformation fields, where the estimation of the local trans- formation is done on small windows. This is confirmed by extensive com- parisons using synthetic deformations of real images.