PDF version of this document
Registration has become a vital pre-processing step in the analysis of bio-medical data where unpredictable yet restrained variations are inherent. In recent years there have been attempts to find an alternative to the arbitrary choice of a reference to carry out registration [1] and such endeavours presently continue. We base our research on the hypothesis that better paradigms exist for finding solutions to the registration problem and not only will these solutions be more precise, but also they will be globally optimal, suggesting that they may represent the correct answer.
Utilisation of registration methods for the establishment of overlap is made ever more appealing as it allows the construction of appearance models automatically. Thus far, good appearance model formations had the pre-requisite that good landmark points needed to be identified in the whole training data, thereby indicating where analogous regions lie. Following the successful work of Davies [2] on shape models, it is known that the construction of optimal statistical models can be conveniently treated as an optimisation problem. An objective function is defined which aims to find a model that is most concise and still fits all of its training data. The inference of the ``goodness'' of a model is assisted by the Minimum Description Length (MDL) criterion, relying on the assumption that simple descriptions are most preferable. In line with this approach, we will define a distinct way of representing an appearance model and the data encoded using that model. Once such messages can be composed, we shall seek a message whose length is minimal. Lower message lengths imply lower variation in the model and hence increased similarity across the whole data set. The optimisation problem itself is guided by a reparameterisation that sets the regime by which the shortest message is to be found.