Advanced Sequence Classification Techniques Applied to Online Handwriting Recognition

摘要

The term handwriting recognition (HWR) denotes the process of transforming a language, which is represented in its spatial form of graphical marks, into its symbolic representation. Online HWR performs this task concurrently to the writing process. The present thesis studies high-accuracy recognition methods applied to online HWR. Those methods have been implemented within the writer independent online HWR system frog on hand (f reiburg recognition of on-line handwriting). In online HWR, data are typically represented as vector sequences. In addition to HWR, vector sequence data appear in a number of additional pattern recognition problems, for instance, in speech recognition, genome processing, financial and medical applications, and robotics. For those problems, designing classifiers that directly address the data’s natural representation can greatly improve the recognition accuracy, compared to a potential pre-applied transformation to vector space data. Beside introducing novel online HWR approaches, a concern of this thesis is also to develop broadly applicable pattern recognition techniques, which are generic to this bouquet of sequence data problems. Emphasis is placed on classification. This thesis describes two complementary classification methods, one of them (CSDTW) falling into the so-called generative, the other one (SVMGDTW) into the so-called discriminative classification category. The generative CSDTW (cluster generative statistical dynamic time warping) is a scalable sequence classification, which aims at holistically combining sequence cluster analysis and statistical modeling. Contrary to previous approaches, these two aspects are embedded in a single feature space and use a closely related distance measure. As will be shown, this combined modeling leads to very accurate HWR results. Particularly interesting in the context of statistical classification, like CSDTW, is the modeling of so-called directional data (i.e., data which corresponds to a direction, thus, in 2D is distributed on the unit circle; opposed to directional data, linear data is distributed along the real line). In online HWR directional data appear as a valuable feature by means of the angular pen trace direction. This thesis describes a unified modeling of directional and linear data within