Cursive script recognition based on a combination of AI techniques

A system for the recognition of on-line cursive script, called NuScript, is under development at the University of Nottingham. Most previous recognition systems have been based on models of the writing process, in contrast, NuScript is based mainly on the human reading process. The realisation of this model has been achieved by integrating multiple artificial intelligence techniques and other more traditional pattern recognition methods via a blackboard framework. Techniques investigated so far include feed-forward neural networks, a constraint propagation system and genetic algorithms, as well as more traditional pattern recognition algorithms based on statistical and syntactical methods. This paper first briefly describes the human reading models on which the system is based. It then gives an overview of NuScript, the object-oriented blackboard system which implements the model at the highest level. This is followed by a more detailed description of some of the component knowledge sources which implement different subsystems in a variety of ways; of particular note is a new technique based on feature moment vectors (FMVs). This description includes some work using constraint propagation and genetic algorithms to obtain the values of parameters used by the knowledge sources. The use of FMVs for both lexicon reduction and for candidate confidence adjustment is described. Finally some preliminary results are described and conclusions about the system drawn