High-dimensional mutual information estimation for image registration

We present a new algorithm for mutual information estimation for image registration based on the nearest neighbor entropy estimator of Kozachenko and Leonenko. We modify the algorithm to be numerically robust and computationally efficient, with optimal asymptotic complexity O(N/sub pixels/d/sub dim/). We propose two MI-based criteria exploiting the high-dimensionality of the feature space and show their effectiveness in determining the correct alignment even in difficult cases when classical criteria fail.