 |
| Illustration of competition mediated by symmetry. |
Back Top Next Kolloquium - DVI - F.Leymarie
Benjie Kimia, Thomas Sebastian & Huseyin Tek
Skeletally coupled deformable models (SCDM) is a segmentation technique, that implements a "skeletally" mediated competition between growing seeds. The initialized seeds grow by a combination of statistic and smoothing forces. In addition, seeds that are adjacent to each other (with no other seeds in between), compete for pixels in between them.
This competition enables the seed that better represents a pixel to capture it irrespective of when the seed arrived. This "back and forth" competition is mediated by the inter-seed skeleton (the skeleton of the background).
The skeleton is interpreted as the "predicted" boundary resulting from the growth of the seeds and the desirability of this prediction is used to modulate the growth of seeds.
SCDM combines the advantages of seeded region growing which implements a global competition among all initialized seeds, region competition which implements a local competition between adjacent regions, once they contact each other, and curve evolution based bubbles.
|
| Illustration of competition mediated by symmetry. |
Seed on the right is closer to the edge than the one on the left. Hence, the seed on the right has to be "slowed down" to allow the one on the left to "catch up". This is achieved by making the seeds compete for the pixels in between them. This competition is mediated by the inter-region skeleton. Specifically, if the skeletal point is more likely to belong to one region compared to the other, the former should grow faster to capture it. In the above illustration, the skeletal point A is more likely to be part of the region on the left, and so that region has to grow faster relative to the region on the right.
 |
 |
 |
| Original image | Seeds initialized | Converged segmentation |
 |
| The 3D visualization of the carpal bones, segmented via SCDM. |
3D model....
Probabilistic model (gray-level statistics):
Currently we use a Gaussian model for the evolution. We assume that the intensities are normally distributed, and use a force proportional to log(P), where P(I, µ, sigma) is the probability that the pixel I belongs to a Gaussian distribution of mean µ and variance sigma².