Browse Prior Art Database

Hidden Markov Model Handwriting Recognizer

IP.com Disclosure Number: IPCOM000108627D
Original Publication Date: 1992-Jun-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 3 page(s) / 121K

Publishing Venue

IBM

Related People

Tappert, CC: AUTHOR

Abstract

Hidden Markov Models (HMMs) have been applied with success to speech recognition (1). Here, we apply HMM to handwriting recognition. In particular, we disclose a particular HMM that, as a special case, reduces to a standard elastic-matching, handwriting recognizer. Because the HMM optimizes for variance and transition parameters, it is more powerful and yields a higher recognition accuracy compared to elastic matching.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Hidden Markov Model Handwriting Recognizer

       Hidden Markov Models (HMMs) have been applied with
success to speech recognition (1).  Here, we apply HMM to handwriting
recognition.  In particular, we disclose a particular HMM that, as a
special case, reduces to a standard elastic-matching, handwriting
recognizer.  Because the HMM optimizes for variance and transition
parameters, it is more powerful and yields a higher recognition
accuracy compared to elastic matching.

      In handwriting recognition, following appropriate signal
processing  and feature extraction, a writing unit (character,
stroke, etc.) is usually represented by a sequence of feature vectors
S = V1, V2, ..., VN, where feature vector Vi = (v1i, ..., vri)
corresponds to coordinate point i , N is the number of feature
vectors (length) of the writing unit, and r is the number of
features.  Commonly used features are simply the normalized x and y
coordinate values.  Prototypes provide references against which the
unknown is matched.  In matching an unknown against a prototype, a
point distance is used to compare an arbitrary point i in the unknown
to a point j in the kth prototype.  The point distances are summed to
obtain an overall distance of an unknown to a prototype.

      Consider the finite-state-machine model of a prototype
character shown in the figure.

      In this model the observable output from transitions between
states is a feature vector that is a probabilistic function of the
origin state  of the transition.  In traversing each arc in this
model a feature  vector is produced with assumed underlying
multivariate normal distribution

                            (Image Omitted)

where Prob(i,j) is the probability of producing feature vector Vi =
(v1i, v2i, ... , vri) on an arc originating from state j
characterized by a mean vector Mj = (m1j, m2j, ... , mrj) and a
covariance matrix Sj = {s2mnj} (m,n = 1, 2, ..., r); r is the number
of features.  This model performs nonlinear warping in that some
transitions cause stretching while others cause compression of the
sequence of feature vectors produced by the model.

      Traversing the arcs of the model produces the sequences of
feature vectors that represent a character.  In order to find the
optimal overall probability of the model generating the unknown, it
is necessary to estimate the  maximum value of the cumulative
probability over the possib...