Comparaison de deux critères en classification ascendante hiérarchique sous contrainte de contiguïté, Application en imagerie numérique
Résumé
We analyze an algorithm of ascendant hierarchical classification under contiguity constraint and using the aggregation principle of reciprocal nearest neighbors. This algorithm is situated in the general framework of quick ascendant hierarchical classification algorithms. Two cluster merging criteria are studied. The former is the classical inertia Ward criterion and the latter consists of the maximal likelihood linkage family criteria. A new contiguity version of this criterion proves its efficiency in image segmentation. One major feature of our algorithm is the linear nature of the computational complexity. New strategies concerning multiple aggregation in the class formation and contiguity notion are positively evaluated in terms of quality and efficiency. We establish mathematically and experimentally how the used criterion influences inversion possibility in the tree building. Finally, comparative results of both types of criteria in image segmentation on satellite pictures are discussed.