A way to account for models in image analysis illustrated by motion extraction

Image analysis requires adequate models, i.e. efficacious symbolic representations of a priori knowledge or hypotheses. These are expressed in terms of basic structural entities (grey-values, derivatives, etc.), defined by a conventional preprocessing of image data. That such entities are in turn subject to models is, however, much less in focus. This article argues in favour of a manifest segregation of task-based models versus image-based models, and of maintaining a transparent relation between the latter and the raw data. A possible approach is suggested based on topological duality (to model the data) and gauge invariance (to model the task). The procedure is applied to motion extraction. By virtue of model transparency it has indeed proven possible to obtain results for a (2+1)D benchmark sequence outperforming all existing algorithms that have been reported in a comparative study. The extraction of cardiac wall motion from an MR sequence of a canine heart is illustrated using the same method.