Skip to Main content Skip to Navigation

Reconnaissance des formes évolutives par combinaison, coopération et sélection de classifieurs

Abstract : In the field of Pattern Recognition (PR), imperfect information handling is often carried out by appling several theories dealing with its imprecision, its uncertainty, its incompletion and/or its contradictions. The decision making would then be influenced by this imperfection. An alternative to this handling consists of reducing the effect of this unperfectness by associating several classifiers through several strategies. Combination is one of the possible strategies for the association of several classifiers. However, it is also possible to imagine two other strategies: the first one is to involve several classifiers into cooperation in order to make a final decision, with the possibility that the decision of a classifier could be influenced by another classifier decision. The second consists in selecting the decisions of one or several classifiers according to the situation and/or to time. We propose a {\em classifier system} approach which can involve one or several of these strategies. In the PR framework, we propose, in this thesis, some approaches and some solutions involving the combination, the cooperation and/or the selection of classifiers strategies, while bearing in mind the temporal aspect of the evolving natural objects features. Through the study of the static and dynamic aspects of PR, we suggest that to recognize dynamic classes, there are two possible approaches. When the trajectories of the classes do not intersect and these classes are multi-modal, our approach consists of transforming these dynamic classes into static ones. The classes have then complex shapes. To deal with such classes, an algorithm of classifiers cooperation is proposed. It involves an unsupervised PR method which is able to carry out an adaptive classifier selection and several supervised PR methods. When there are no intersections and the classes evolve in time in a continuous way, the proposed approach consists in transforming the PR system in a dynamic one. A method, based on the modelisation of the system states changes with a fuzzy Petri net, is proposed. The fuzzy specification of these changes allows the prediction of the system states which are the most adapted to the PR problem, at the involved instant.
Complete list of metadata
Contributor : Veyis Gunes <>
Submitted on : Wednesday, October 12, 2011 - 7:14:48 PM
Last modification on : Wednesday, October 14, 2020 - 3:55:18 AM
Long-term archiving on: : Friday, January 13, 2012 - 3:16:18 AM


  • HAL Id : tel-00631621, version 1



Veyis Gunes. Reconnaissance des formes évolutives par combinaison, coopération et sélection de classifieurs. Apprentissage [cs.LG]. Université de La Rochelle, 2001. Français. ⟨tel-00631621⟩



Record views


Files downloads