A New Framework for Hierarchical Segmentation Using Similarity Analysis

We present a new framework for hierarchical segmentation of multidimensional multivariate functions into homogeneous regions. Homogeneity is defined as constancy of n-th order derivatives (called features) of the function. The degree of similarity (measure of homogeneity) is used as a scale parameter to obtain a stack of segmentations. Hierarchical segmentation is represented as a tree which contains the geometric and topological information about the detected regions. Detected regions preserving their information in the tree over large range of scales are selected into a pyramid representation. Results showing noise robustness and computational efficiency of the proposed method are presented. Experiments to compare the method with three other segmentation techniques and applications to two- and three-dimensional images having one-, three- and six-variate data are described for the zeroth and first order region features.