Classification-driven stochastic watershed. Application to multispectral segmentation

The aim of this paper is to present a general methodology based on multispectral mathematical morphology in order to segment multispectral images. The methods consists in computing a probability density function pdf of contours conditioned by a spectral classification. The pdf is conditioned through regionalized random balls markers thanks to a new algorithm. Therefore the pdf contains spatial and spectral information. Finally, the pdf is segmented by a watershed with seeds (i.e., markers) coming from the classification. Consequently, a complete method, based on a classification-driven stochastic watershed is introduced. This approach requires a unique and robust parameter: the number of classes which is the same for similar images. Moreover, an efficient way to select factor axes, of Factor Correspondence Analysis (FCA), based on signal-to-noise ratio on factor pixels is presented.