Indexation de structures de documents par réseaux bayésiens

Our objective is to study the contribution of naive increased Bayesian networks in problems of image classification. The images used in this study represent the structure of a document containing text blocks and graphics. We proposed three variants of Bayesian networks. First naive Bayesian networks RN who, despite their simple structure and strong assumption on independence have given very good results. Secondly, the naive Bayesian networks augmented by a tree TAN. Indeed, the assumption of independence among attributes is in general false. Thus, there are different techniques to relax this assumption. Thirdly, the naive Bayesian networks augmented by a forest called FAN who they are rather well known classification problems have not been investigated to our knowledge in image classification. The results showed a marked improvement over the FAN network RN and TAN.