Automatic Semantic Network Construction for Multi-Layer Annotation of Satellite Images

Jean-Baptiste Bordes (1, 2) , Henri Maître (1, 3)

1 LTCI - Laboratoire Traitement et Communication de l'Information
2 DGA/CTA/DT/GIP - CEP Arcueil
3 TSI - Département Traitement du Signal et des Images

A novel method is presented for annotating satellite images. The labels used for annotation are given by a user with a set of example images. A learning step is then applied to learn the model. The originality of the method is to formulate the problem of semantic annotation to a further extent than a mere probabilistic classification task. The method takes into account the semantical relationships between the concepts by considering a duality between the structure of the model and the structure of the set of labels. The semantical structure of the labels is represented by a semantic network containing three semantical relationships: synonymy, meronymy, and hyponymy. The semantic network is constrained in a hierarchy induced by the links of hyponymy and meronymy. By a procedure of MDL model selection, it is possible to find the optimal semantical structure of the set of labels.

Mots clés: remote sensing; indexing; statistical modelling; semantic indexing

Domaines: Traitement des images [eess.IV]

Jean-Baptiste Bordes, Henri Maître. Automatic Semantic Network Construction for Multi-Layer Annotation of Satellite Images. 2008. ⟨hal-00315531⟩