Core Points - Variable and Reduced Parameterization for Symbol Recognition

Abstract: Recent research in the field of on-line handwriting recognition has been focused on statistical systems such as Hidden Markov Models, Neural Networks or a combination of these. There are however merits of employing an approach based on template matching. The first part of this thesis presents a new strategy for parameterization of on-line handwritten character samples. A novel efficient template matching method enabled by this parameterization is also proposed. In consecutive chapters of the thesis it is also shown that the proposed structural parameterization enables an effective application of template matching methods to the recognition of cursive script. Ambiguity of the shapes of individual characters in unconstrained cursive handwriting necessitates dictionary interaction for real applications. A fast technique for applying dictionary information to the language independent graph approach has also been developed. A large data set of on-line cursive writing has been collected and the developed system for mixed and cursive on-line handwriting recognition has been shown to produce state of the art results on this data set. One of the obvious potential weaknesses of a structural parameterization technique such as the one presented in this thesis is its sensitivity to digital noise in the form of superfluous coordinates. Possible remedies to deal with such effects have also been studied.