Browse Prior Art Database

Improved Set of Global Features for the Recognition of On-line Hand Writing Using a Statistical Mixture Model

IP.com Disclosure Number: IPCOM000110068D
Original Publication Date: 1992-Oct-01
Included in the Prior Art Database: 2005-Mar-25
Document File: 3 page(s) / 131K

Publishing Venue

IBM

Related People

Bellegarda, EJ: AUTHOR [+4]

Abstract

An improved set of global feature parameters to characterize handwriting, including vertical coordinate, horizontal displacement, and interstroke information is isolated. In particular, it is found that a dynamic definition of the interstroke distance leads to a better recognition accuracy than the previously considered stationary interstroke distance. Background

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

Improved Set of Global Features for the Recognition of On-line Hand Writing Using a Statistical Mixture Model

       An improved set of global feature parameters to
characterize handwriting, including vertical coordinate, horizontal
displacement, and interstroke information is isolated.  In
particular, it is found that a dynamic definition of the interstroke
distance leads to a better recognition accuracy than the previously
considered stationary interstroke distance.
Background

      This article is concerned with the automatic recognition of
handwritten text in any of the following modes: discrete, runon,
cursive, or unconstrained.  We focus here on the signal processing
front end.

      A double codebook processing relying on two sets of prototype
distributions, one for local features and the other for global
features, has been previously considered.  The first codebook is
based on a 6-D feature vector encompassing local features such as
position, direction, and curvature, which was thoroughly analyzed in
(1,2).  The second codebook encompasses more global features such as
the vertical y-coordinate, the horizontal displacement from the
beginning of the stroke, and the interstroke distance if the
character is composed of more than one stroke.  In this article we
concentrate exclusively on the second codebook.  Our goal is to
investigate the effect of the selected parameters on the recognition
rate.

      At each equi-spaced point Pn of coordinates (xn, yn), one forms
a 3-D feature vector of feature elements representing the global pen
trajectory up to Pn.  Those feature elements are given by: (i) the
vertical coordinate yn (height from the baseline), (ii) the
horizontal displacement xn - xi, where xi is the horizontal
coordinate in the first point of the current stroke being processed,
and (iii) the stationary interstroke distance gapdist, which is the
distance between the first point of the current stroke Pi and the
last point of the previous stroke PNk.  Note that gapdist is constant
for all the points belonging to the same stroke and that it is null
for the first stroke in the character, i.e.:

                            (Image Omitted)

      In addition, we consider a dynamic definition of the
interstroke distance.  We define intdist to be the distance between
the current point Pn and in the stroke and the last point PNk in the
previous stroke.  When processing the first stroke in a character,
the parameter will reduce to the norm of the point.  Thus, we have:
The latter definition should carry more information regarding the
character shape.

      Our methodology is first to study one codebook based on a 3-D
feature vector (y, x-xi, gapdist).  We then replace the stationary
in...