A k-nearest neighbour method for managing the evolution of a learning base

Abstract

A character recognition system with continuous learning seeks to constantly enhance its base representation models in order to provide the best recognition rate. The method we are presenting enables the system to enhance its base with models, which are performant in recognition. This method also enables to get rid of models regularly doubtable in efficiency when it comes to interpretation of the characters studied. This rule is similar to the one used in the "Death by suffocation" game of life of Conway. We based ourselves on the theory of k-nearest neighbours to develop a new approach we named /spl epsiv/-adaptive neighbourhood. It makes an adjustment of classes possible, according to confidence rate in each model of the learning base. These rates which are practically represented as weights are taken into account by the stage of the recognition system during the character recognition phase. The use of weight as a model selection factor, useful for recognition, enables the system to manage the evolution of the learning base.