Une méthode générique pour la conception de moteurs de reconnaissance de symboles manuscrits en ligne
This paper presents a generic approach for designing on-line handwritten shapes recognizers. Our approach allows designing very different recognition engines that correspond to various needs in pen-based interfaces. In particular, it allows dealing with a wide class of symbols and characters. We present in detail our system and make the link between our models and more standard statistical models such as Hierarchical Hidden Markov Models and Dynamic Bayesian Networks. We then evaluate fundamental properties of our approach: learning from scratch any symbol, learning from very few training sample. We show experimentally that, using our approach, one can learn both a state-of-the-art writerindependent recognizer for alphanumeric characters, and a writer-dependent recognizer working with any two dimensional shapes that learns a new symbol with a few training samples and requires very few machines resources.