PDF version of this document
Interpretation of images is a rudimentary task which can be tackled by various different approaches. One such approach involves modelling objects which are expected to appear in images. Having modelled objects in some form or another, it is then possible to use this model to analyse new images describing similar objects.
Another problem which is associated with interpretation is image registration. Registration of images is an essential step which enables one to compare images at greater ease. It does so by aligning an arbitrary number of images so that homologous elements described in the image eventually overlap. This alignment of images is performed by transforming the space of images. This means that each image pixels is projected onto a new location and this results in images which are deformed. The ultimate aim is to deform all images until they all appear alike.
An existing method of solving the registration problem is to choose one common reference (or target) frame. There is a need to define a target according to which images are transformed, yet the flawed approach defines only one such target. The choice of this target affects the results, hence there is no unique solution to the problem.
The study we conduct addresses this issue by expressing the registration task in term of the entire set of images. We transform all the images at the same time and evaluate the transformation using a model representative of all the images. We define a principled cost function for this model and the images it encapsulates. Using an optimiser, we then strive to minimise that cost. Since we optimise over a single representation that accounts for all images, we obtain one unique solution.
This research work unifies several different strands. It takes a method which is commonly used to analyse images and indirectly uses it to guide image registration. What we also achieve as a result of the process is a description of transformations from which observed image variation can be derived. Essentially this produces one mean image and operations that deform it to assimilate to plausible, yet unseen images. Even more usefully, by registering images, correspondences can be identified automatically. This resolves the need for manual landmark which is otherwise necessary for the construction of statistical models. Such models can regenerate images similar to the ones which were registered; they can also be used to perform various measurements over objects in the images.