Browse Prior Art Database

Front-End Paradigm for On-Line Handwriting Recognition

IP.com Disclosure Number: IPCOM000111137D
Original Publication Date: 1994-Feb-01
Included in the Prior Art Database: 2005-Mar-26
Document File: 4 page(s) / 216K

Publishing Venue

IBM

Related People

Bellegarda, EJ: AUTHOR [+2]

Abstract

A new front-end processing is introduced for on-line handwriting recognition. The method relies on local curvature information to flexibly segment the handwriting into meaningful, morphologically consistent units. The new front-end is used in conjunction with a bank of multilayer neural networks to perform on-line character recognition.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 26% of the total text.

Front-End Paradigm for On-Line Handwriting Recognition

      A new front-end processing is introduced for on-line
handwriting recognition.  The method relies on local curvature
information to flexibly segment the handwriting into meaningful,
morphologically consistent units.  The new front-end is used in
conjunction with a bank of multilayer neural networks to perform
on-line character recognition.

      Existing front-end approaches to the automatic recognition of
handwriting are based on either of the following: (i) an entire
stroke from pen down to pen up [1 ]; (ii) data-driven frames built
from a statistical analysis of the handwriting [2 ]; and (iii)
sub-stroke units similar to frames, but incorporating some a priori
information regarding the mechanics of the handwriting process, e.g.,
ballistic strokes [3 ].

      The stroke-based approach suffers from inherent difficulties
such as (a) a large set of strokes is needed to cover the entire
handwriting alphabet, and (b) stroke lengths and shapes vary
differently for different writing modes (handprint, cursive).  In a
frame-based front-end processing, the body of data is looked at as a
whole and elementary units of handwriting (frames) are chosen
systematically without any consideration for the mechanics of
handwriting.  Namely, at selected set of points in the pen
trajectory, a fixed number of parameters vectors are concatenated  to
form a basic vector representative of the handwriting in this
vicinity.  In such an approach the elementary units of handwriting
tend to be more robust and less sensitive to noise than in a
stroke-based approach, since they allow for different writing styles
and deviations thereof [2 ].  On the other hand, they do not utilize
any prior knowledge about the handwriting process.  In prior art, we
have tried to incorporate such knowledge by deriving a front-end
based on ballistic strokes.  A ballistic stroke corresponds to the
portion of handwriting resulting from one complex muscular motion.
Mathematically, a ballistic stroke corresponds to the writing between
two consecutive minima in the y-velocity.

      Here we present an alternative front-end processing based on
the concept of a  segment.  While a ballistic stroke reflects
handwriting production, a segment is closer in spirit to handwriting
perception.  Thus, it relies on local curvature information as
opposed to (possibly spurious) velocity minima.  This allows (i) more
flexibility in the unit definition by introducing a threshold
curvature as a variable, and (ii) more robustness in the resulting
segmentation by adjusting the threshold for the expected level of
noise.  As a result, each handwritten character can be described by a
small number of representative segments (building blocks), where all
points within a segment share somewhat similar characteristics.
These segments are then used as inputs to a bank of neural networks
to perform handwriting recognition.

      S...