PDF version of this entire document
Originally, NRR was based on comparison between pairs of images. When a group of images is presented, it is not clear whether to treat separate pairs from this group separately and repeatedly or to make use of the entire group.
A distinction is made between two approaches to tackling the NRR problem. Some take the approach wherein one image from a set is chosen as a reference (or template) and the remainder of the set is transformed to fit that reference [,,]. It is a repeated pairwise approach and it means that, at the very end, all images will be assimilated to that particular reference, which was arbitrarily and possibly ill chosen. The other approach is based on the idea that the entire set of images should be transformed to minimise a groupwise objective function [,,].
Debates over the validity (or lack thereof) in the pairwise approach are of great relevance to the work presented hereafter.
In repeated pairwise registration, there are various choices one can make as regards a reference image. However, the results of the registration tend to be dependent on this choice. For example, choosing a reference image from the set of original images will in general bias the results, with different choices producing different registrations. This motivates the need to evaluate the results of NRR algorithms [,,].
Another issue emerges when non-deterministic choices are made in the registration algorithm, which can also lead to choice-dependent results.
Because of these problems, the work described in this thesis concentrates on a fully groupwise approach, with the aim of finding an approach that is not subject to these ambiguities and biases.
Roy Schestowitz 2010-04-05
