PDF version of this entire document
The work covered in this thesis can be summarised as follows. Firstly, a novel framework was described which non-rigidly registers images using a model-based similarity measure. This framework is able to deal with any type of images and, while it requires a a sufficiently large number of images in order to become practical (i.e. in order for a sensible model to be built), its performance does not depend on the type of variation that is contained in the set of images. As a result of registering the images using a model-based approach, one also obtains an appearance model, which is progressively refined and whose quality is dependent on the quality of the registration algorithm. This establishes a framework for automatic construction of models that requires a model-based objective function.
The second part of the work is concerned with assessment. Two things are being assessed: the quality of any appearance model (or any generative model) and the quality of a groupwise registration. This opens up the possibilities of comparing NRR algorithms and supports evidence-based development of improved methods.
