DATA DRIVEN DESIGN OF HMM TOPOLOGY FOR ON­LINE HANDWRITING RECOGNITION

Jay J. LEE, Jahwan KIM, and Jin H. KIM

Dept. of Electrical Engineering & Computer Science, KAIST,
373­1, Kusong­dong, Yusong­gu, Taejon 305­701, Korea
E­mail: {joony, jahwan, jkim}@ai.kaist.ac.kr

Although HMM is widely used for on­line handwriting recognition, there is no simple and well­established way of designing the HMM topology. We propose a data­driven systematic method to design HMM topology. Data samples in a single pattern class are structurally simplified into a sequence of straight­line segments, and then these simplified representations of the samples are clustered. An HMM is constructed for each of these clusters, by assigning a state to each straight­line segments. Then the resulting multiple models of the class are combined to form an architecture of a multiple parallel­path HMM, which behaves as a single HMM. To avoid excessive growing of the number of the states, parameter tying is applied in that structural similarity among patterns is reflected. Experiments on on­line Hangul recognition showed about 19% of error reductions, compared to the previous intuitive design methods.

In: L.R.B. Schomaker and L.G. Vuurpijl (Eds.)
Proceedings of the Seventh International Workshop on Frontiers
in Handwriting Recognition, September 11-13 2000, Amsterdam,
Nijmegen: International Unipen Foundation,
ISBN 90-76942-01-3
pp. 239-248.