PDF version of this document
Image registration, or more particularly non-rigid image registration (NRR), is an essential pre-processing step that enables easier interpretation of image sets. Registration of images is a task which involves repeated image transformations, ultimately aiming to make a group of images look more similar. By striving to and thus achieving cross-image similarity, change can be more easily analysed and image structures better understood. Various approaches exist to solving the NRR problem. These different approaches vary in terms of their representation of transformations, the notion of similarity among images, and the method by which good transformations get selected.
In most cases, registration is posed as a pair-wise problem, whereby several images are transformed to fit one single image, commonly known as the reference or template image. A registration that takes into account the entire set and treats images equally, on the other hand, could and should result in a better overall registration.
Due to the arbitrary choice of a reference image, a pair-wise approach is rarely constrained and its results are greatly affected by that subjective choice of a reference. Put differently, depending on which image gets selected as the reference, different results are to be reached. Moreover, the loosely-defined approach by which registration gets solved means that it is difficult to reason about correctness of the result or even quantify it reliably.
In order to assess the power of registration algorithms 'off-line', one can make use of the ground-truth solution, often straying away from that solution by applying perturbation. It is then possible to investigate an algorithm's ability to 'annul' the effect of the perturbation by performing registration. A good registration algorithm will be able to recover the ground-truth solution. The drawbacks of such an approach are that ground-truth solutions must be provided a priori and partial recovery of the solution needs to be evaluated somehow.