Structural skeleton representation is an efficient method to represent the general structure of a shape. Many ways exitst to produce skeletons, but a raw skeleton contains many branches that are either redundant or the product of noise along the shape's contour. The process of skeleton pruning reduces the number of branches to obtain a much more compact tree. However, most modern methods of pruning suffer from two problems: noise and topological alteration.
Noise along the contour of a shape produces many spurrious branches in the intial tree. By current methods, some of these noisy branches are erroneously retained after pruning. This leads to skeletal structures that are inaccurate and that contain more branches than they need to have to represent the shape.
Topological alteration occurs during the pruning process. Branches that are not pruned are often times reduced in length, cut short by the pruning algorithm. In cutting the branch short, the endpoint of the branch is moved inward toward the center of the shape, off of the contour it is meant to represent. Sometimes in cutting a branch short, parts of a skeleton actually get disconnected, creating two separate skeletons out of one.
In the proposed method, we use contour partitioning to solve both of these problems. The contour partitions are obtained using Discrete Curve Evolution (DCE), an algorithm that progressively filters out noise while retaining points of visual significance along the contour. Using these points, a skeleton can be constructed that is topologically stable to the original shape, and is mostly free from noise distortions.
The skeletons we obtain represent the most significant aspects of the original shape. We are then able to use these skeletons to perform effective shape similarity comparisons.
| <- Prev | Next -> |