Synthetic On-line Handwriting Generation by Distortions and Analogy

One of the difficulties to improve on the fly writer-dependent handwriting recognition systems is the lack of data available at the beginning of the adapting phase. In this paper we explore three possible strategies to generate synthetic handwriting characters from few samples of a writer. We explore in this paper both classical image distortions and two original ways to generate on-line handwritten characters: distortions based on specificities of the on-line handwriting and a generation based on analogical proportion. The experimentations show that these three approaches generate different distortions which are complementary. Indeed the combination of them allows to achieve using only 4 original characters for the learning phase a mean of 91.3% of recognition rate for 12 writers.