Approche markovienne bidimensionnelle d'analyse et de reconnaissance de documents manuscrits

In this thesis, we present a general bidimensional markovian approach in theframework of handwritten document analysis and recognition. This approachcalled AMBRES (Bidimensional Markovian Approach for image Recognition andSegmentation) is based on Markov random elds, 2D dynamic programming anda bidimensional analysis of images.AMBRES has been successfully applied to a wide variety of tasks such as hand-written character recognition, handwritten word recognition, layout analysis ofhandwritten documents and logo recognition. It could be also used for other tasksin computer vision.Rigorous protocols have been considered in order to evaluate the system and itsperformances. In particular, AMBRES has been tested in the framework of theRIMES evaluation campaign.