Co-articulation Modeling for Handwriting Synthesis

Summary form only given. Personal style handwriting synthesis aims to produce texts in the writer's style from a few handwriting samples. A handwriting synthesis system usually focuses on the production of the glyph, a handwriting unit separated by a space such as English word, Tamil word, Chinese character, Korean character, etc. Since, the glyph consists of one or more of components, called graphemes, and a grapheme is an atomic unit of glyph that has a relatively unique shape, grapheme is a most desirable unit of handwriting generation. However, handwritten grapheme generation is not simple as it sounds, because a grapheme is written differently depending on its neighboring graphemes, which is called co-articulation effect. Noticing the value of statistical structure analysis in character recognition, KAIST has used Bayesian network to model strokes and stroke relationships of handwritten characters. Bayesian network modeling overcomes the crudeness of naive Bayesian approach as well as the complexity of brute force Bayesian approach. From the given handwritten samples, we have constructed writer independent Bayesian network models utilizing the hierarchy of pixel-stroke-grapheme-glyph. The proposed system is applied to the Korean character synthesis. Experimental results demonstrate a high degree of visual plausibility