Introduction
Segmentation is
one of the most basic steps in a vision system. Unfortunately, although
human beings can perform this step in a split of a second, researchers
have been trying to teach computers to do that for many years, without
satisfactory success until now. Therefore, it is still a very active
research field.
Segmentation is
needed for many kinds of applications, like biomedical applications, which
are of specific interest to this group, and in addition to home
entertainment industry, robotics, security and military etc.
The segmentation
process is usually done by representing the contours as image edges, which
are local image features most commonly extracted from the gradient of the
image. However, in this way, the result is usually fragmented, and
intelligent techniques have to be applied in order to obtain a
non-fragmented segmentation.
A different approach was suggested which evolves a curve from an initial segmentation guess to the actual object boundary. This approach, known as Snakes or Active Contours, suffers from the fact that it cannot handle changes in topology and sharp corners very well.
Recently, Geodesic Active Contours has been proposed. This method is based on Snakes, but uses the Level Sets methods in order to automatically handle sharp corners and changes in topology.
Strangely,
although the Geodesic Active Contour model involves many very complex
mathematical techniques, one of the key issues in the model is the
computation of a partial differential equation, which is usually done by
using divided differences – not a very complex scheme... The order of
approximation in this model is not very good.
In this project we
tried to use advanced approximation techniques, namely the spline
representation for the equations of the Geodesic Active Contours model.
Conclusions
Using splines we can achieve smoother solution,
and avoid almost completely re-initializing the level set function into a
distance map.
The time step is a little bit smaller, however,
since we do not perform so many initializations the overall time is more
or less the same.
Overall, it does not do any huge breakthrough in
the model.
However, visualization can be improved a lot by
using splines, achieving sub-pixel accuracy, and visualization of 3D
functions.
Concerning the Geodesic Active Contours model in
general, it is good for refining segmentation, but not so much for
segmentation without a good initial guess, except for easy images.
|