PDF version of this entire document
T his thesis studies the relationship between models and registration. The main contribution is a novel approach to the evaluation of statistical models of appearance [], which can also be used to assess the quality of non-rigid registration (NRR) algorithms. Additionally, a method is presented for registering images, using model complexity as a figure of merit [].
The work is motivated by the observation that, given a set of registered (i.e. fully-aligned) images [], a statistical model of appearance can be built automatically. This is exploited in two ways: first, we propose an objective function for automatic NRR of sets of images, based on an information-theoretic measure of model complexity; second, we show how the quality of NRR of a set of images can be assessed by measuring the quality of the resultant model. Both are important, but the thesis makes a more significant contribution to the latter. The ability to assess NRR algorithms is important for benchmarking and comparative studies, or for quality control in practical applications.
This work thus makes a contribution to both the modelling and non-rigid image registration literature. The approach is generic, but the thesis is focused on 2-D brain images. The extension to 3-D is discussed and, at the `proof-of-concept' stage, 1-D images are also used. These are helpful in validation experiments that exploit synthetic images whose nature is well understood.
