Writer Adaptation for Online Handwriting Recognition System Using Virtual Examples

For an online handwriting recognition system equipped with a writer-independent classifier to progressively improve the recognition performance for a specific writer with an increase in his/her handwriting inputs, the following method is proposed: (1) for a handwriting pattern that causes a recognition error, a two-class classifier of the corresponding class is (re)constructed as a part of a writer-dependent classifier, separately from the writer-independent one, where artificially generated examples are used to compensate for lack of training examples. (2) In the recognition stage, the writer-independent classifier is applied first, and then the constructed writer-dependent classifier is used only in cases in which a result obtained by the writer-independent classifier possesses lower reliability. We examine the effectiveness of the proposed method using 6,000 Japanese Hiragana characters written by 3 users. As a result, an average recognition rate of 98.07% was obtained by the exclusive use of the writer-independent classifier. On the other hand, the rate improved to 99.92% with at most 7 (re)constructions of a writer-dependent classifier.