A maximal mass confinement principle for rigid and locally rigid image registration

In this paper we propose (1) to set the problem of image registration as a contour/region-template-to-image matching problem using so-called confiners--also called blobs or components--as template regions, (2) to select the confiners of one of the images by passing through the hierarchical structure which they define and registering them successively rigidly form coarse-to-fine to the other image, the target image, and (3) we propose a maximum mass confinement (MMC) principle for contour-to-image registration. This principle allows us to derive a similarity measure assessing how well the confiner fits into the target image simply by calculating the gray value mass confined by its contour. By optimizing this measure for rigid transformations we obtain our MMC algorithm registering a contour locally rigid to the target image. We illustrate that by proceeding based on (1-3) problems can be avoided which were related to previous registration algorithms based on confiners. We compare our MMC algorithm with another template matching algorithm based on normalized mutual information. Equally, we compare our hierarchical image registration strategy with B-Spline based non-rigid registration using normalized mutual information. We performed our evaluation on real and simulated images in terms of robustness, accuracy and computation speed. We show that both, MMC template matching on its own and hierarchical image registration using MMC, in most cases outperform the respective alternative method.