Although region and boundary techniques have been used for a wide
variety of image understanding tasks, they rarely can be applied
without a high-level agent to guide their operation. Consider for
example, the generation of spurious boundaries in locations with steep
gradients. These false boundaries can occur in locations where
undersampling produces a stair-step or aliased image. Even if
aliasing can be corrected, two- and three-dimensional segmenters are
still susceptible to boundary tracking problems (boundary tracking is
necessary in order to aggregate primitive elements, such as lines and
surfaces, into complete anatomical structures or landmarks). Boundary
tracking errors can occur at locations where edges intersect or where
edges break off from one another. Another major limitation of these
methods is that they fail to utilize or extract any shape information
from the image objects (Section will describe a
technique for understanding shape information). In general, some sort
of high-level interaction is required to (1) provide a context for the
operators, (2) bound the search space, (3) disambiguate conflicting
results, and (4) account for missing information.
Section
will describe some
approaches to deal with these limitations.