Segmentation non-supervisée de champs de données multi-composantes

In this paper, we revisit some methods of unsupervised segmentation on multi-component datasets. The first method lies on the Llyod’s algorithm, also known as Kmeans algorithm; we suggest an original approach to initialize this algorithm, by using data obtained during the building of the minimum spanning tree (MST) [7]. The same kind of initialization has been applied for a second method inspired by works of Shi and Malik [11] based upon the Laplacian of a fully connected graph. The very important problem of the choice of the metric for characterizing similarities between two points of multi-dimensionnal datasets is also discussed. We present results obtained with both segmentation methods on multispectral images and on asteroid reflectance spectra.