Merging regions based on the VDM distance
Francisco Abad, J. Garcia-Consuegra, Guillermo Cisneros · 2002
In a per-object classification system, many authors have advised a two-step process: First by obtaining small, homogeneous objects, second by merging those regions. The authors present a new method to decide whether two regions belong to the same class or not. An extension of Chang and Li's method of adaptive region-growing is proposed, in order to take into account the case of multiband images. Region dispersion is calculated by means of the vector degree of match (VDM) distance, proposed by Baraldi and Parmiggiani. This distance measures the similarity degree of two n-dimensional vectors. Each intra-region similarity is computed and then compared with other regions. Instead of defining a similarity degree fixed threshold, obtained by a trial and error process, they propose an adaptive one, based on a global percentage similarity. In short, this work proposes a reduction of the features space into a one-dimensional space which measures the similarity by a non-parametric criterion, thereby avoiding the determination of the density function problem.