A method to accelerate writer adaptation for on-line handwriting recognition of a large character set

An approach to accelerate writer adaptation for on-line handwriting recognition is proposed. It is known that adapting to a writer by learning the writer's own style significantly improves recognition accuracy. However, adapting to a writer can take considerable time until the performance comes up to a satisfactory level, particularly for recognition of a large character set. This paper proposes an adaptation method which uses not only misclassified patterns but also correctly-classified patterns as learning samples. The strategy employed in the method selects acquired prototypes based on their contribution to classification, while treating the misclassified prototypes (i.e. the acquired prototypes that were misclassified before being added) with higher priority when updating the prototype set. The results demonstrate that the proposed method improves the performance and accelerates adaptation especially during the early phase of adaptation. It is also shown that the method yields stable improvement in accuracy over a long period of adaptation with the computational cost acceptable for most real applications.