An investigation into the representation of feature vectors and their stochastic modeling by stroke-based HMM for on-line handwriting character recognition

摘要

Feature-parameter representation of on-line hand-written characters and probabilistic-distribution modeling of feature parameters, that affect the performance of stroke-HMM based character recognition, are investigated in this paper. Experiments of on-line hand-written Kanji character recognition with a lexicon of 1016 elementary characters revealed that velocity vectors of pen movement represented in polar-coordinate system reduced the error rate by 41% (single Gaussian), 16% (3 mixtures) compared with those in the Descartes system. Furthermore, it was shown that the continuous density models with Gaussian mixtures outperformed the discrete density models that have been commonly employed in the conventional systems.