Article Preview
TopSeveral computational approaches have been implemented during the last decades for skeleton extraction. Skeletonization provides an effective and compact representation reducing a 3D form to a surface and a 2D form to a one-dimensional structure. There are different classes of methods to compute skeleton of bounded objects. Methods based on Distance Transform (DT) generate a distance map, a graylevel image representing the distance to the closest boundary from each point of the shape. In this framework, the skeleton is described as being the locus of local maxima of a distance map (Blum,1967), (Montanari,1969). If this function is visualized in the three-dimensional space it appears as a not differentiable surface showing ridges formed by points that, when projected onto the image plane, define the skeleton structure. The topological thinning methods work eroding iteratively the shape until the skeleton is obtained (Ammann & Sartori-Angus, 1985), (Zhang & Wan, 1996).
In most cases the criteria used to delete a point are local, whereas the skeleton allows to capture global geometric features of a given shape (Pavlidis,1980, Lam, Lee & Suen,1992). Skeleton may be computed by Voronoi diagrams created using boundary points as anchor points (Ogniewicz, & Ilg, 1992). The main drawback of this approach is that a great number of anchor points are not relevant for skeleton generation and therefore additional skeleton branches are frequently introduced. The Field-based approaches evaluate skeletons recurring to potential functions, derived by the Electrostatic or Gravitational Theory. Boundary pixels are considered behaving like point sources of a potential field, as a consequence these methods require a reliable contour point localization. The resulting fields are diffused introducing an edge-strength function. In this context the skeleton is extracted through the level curves of the strength function (Grigorishin, Abdel-Hamid, & Yang, 1996).