Mode boundary detection by a morphological watershed transformation for clustering

A new approach to unsupervised pattern classification is developed, based on concepts of mathematical morphology. Regions of high local concentration of observations in the data space are characterised by regional maxima of the underlying probability density function (pdf), which is estimated from the available patterns. The proposed approach assumes that these maxima mark modes of the underlying pdf and gives a new methodology for detecting mode boundaries based upon a morphological watershed transformation. Experiment results demonstrate the effectiveness of the proposed method.