EUCLIDEAN SKELETONS

An approach for the skeletonization of two-dimensional (2-D) or 3-D ob jects is presented. Two local measures, phi and d, are introduced to c haracterize skeleton points in n-D, whose good localization is ensured by Euclidean distance mapping techniques. These measures allow the le vel of detail in the resulting skeleton to be controlled. Thresholding these measures does not generally yield a well-defined skeleton: a lo w threshold preserves the original object's topology but produces a no ise sensitive skeleton, while a larger threshold produces a more robus t skeleton but it is generally not homotopic with the original object. To overcome these drawbacks, functions of these measures can be intro duced. Although they generally yield convincing experimental results, they are still sensitive to noise. Instead, a novel global step for 2- D and 3-D images called topological reconstruction is introduced, that will provide the skeleton with robustness with respect to noise and e nsure homotopy with the original object. Moreover, this method is not iterative (like thinning approaches) and hence has reasonable computat ional time for 3-D objects. Results on synthetic 2-D patterns and on r eal 3-D medical objects are presented. (C) 1998 Elsevier Science B.V. All rights reserved.