Decision Support with Belief Functions Theory for Seabed Characterization
The seabed characterization from sonar images is a very hard task because of the produced data and the unknown environment, even for an human expert. In this work we propose an original approach in order to combine binary classifiers arising from different kinds of strategies such as one-versus-one or one-versus-rest, usually used in the SVM-classification. The decision functions coming from these binary classifiers are interpreted in terms of belief functions in order to combine these functions with one of the numerous operators of the belief functions theory. Moreover, this interpretation of the decision function allows us to propose a process of decisions by taking into account the rejected observations too far removed from the learning data, and the imprecise decisions given in unions of classes. This new approach is illustrated and evaluated with a SVM in order to classify the different kinds of sediment on image sonar.
Computer Science - Artificial Intelligence, Computer Science - Information Theory, I.4, I.5