Browse Prior Art Database

Evaluation of the Curvature Information for the Recognition of On Line Handwriting Using a Statistical Mixture Model

IP.com Disclosure Number: IPCOM000110133D
Original Publication Date: 1992-Oct-01
Included in the Prior Art Database: 2005-Mar-25
Document File: 3 page(s) / 133K

Publishing Venue

IBM

Related People

Bellegarda, EJ: AUTHOR [+4]

Abstract

Since its measurement is inherently noisy, it is unclear how much the instantaneous curvature contributes to the recognition of on-line handwriting. In this article, we find that the information provided by the curvature has no significant effect on the recognition rate. Both positional and directional information appear to be more reliable features to achieve a reasonably good description of the handwriting data.

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

Evaluation of the Curvature Information for the Recognition of On Line Handwriting Using a Statistical Mixture Model

       Since its measurement is inherently noisy, it is unclear
how much the instantaneous curvature contributes to the recognition
of on-line handwriting.  In this article, we find that the
information provided by the curvature has no significant effect on
the recognition rate.  Both positional and directional information
appear to be more reliable features to achieve a reasonably good
description of the handwriting data.

      This article is concerned with the automatic recognition of
handwritten text in any of the following modes: discrete, runon,
cursive, or unconstrained.

      Handwriting has previously been characterized by a set of
feature parameters inferred from pen trajectory, and encompassing
position, direction, and curvature information.  In a study aimed at
determining the relative contributions of the selected feature
parameters [*], we found that omitting the curvature is not
deleterious to the error rate, and therefore suggested that the
curvature is not as fundamental as position and direction in
positively shaping the recognition rate.

      With the benefit of hindsight, however, it appears that this
conclusion may have been influenced by the definition used in the
implementation (*).  Since the curvature in [*] was defined in a
linear fashion as differentials in cosine and sine, any subsequent
linear operation aiming at reducing redundancy would likely eliminate
this second order information.  Since, in fact, such a linear
operation was previously performed on the spliced data, this looks
like a foregone conclusion.

      To ascertain the matter, here we investigate a nonlinear
implementation of the curvature, based on the cosine and sine of the
differential in angle.  Again we postulate that the curvature,
regardless of its linear or nonlinear implementation, is not
essential to the present setup in characterizing handwriting.

      Our methodology is first to replace the last two elements of
the 6-D feature vector by its nonlinear counterpart, i.e., the cosine
and sine of the differential in angle.  This will help us to test our
postulate on the curvature.  To challenge the postulate even further,
we then separate out the 6-D codebook into multiple codebooks.  Thus,
the original 6-D feature vector is split into three 2-D vectors
(position, angle, curvature) each assigned to different codebook.
Two cases will be considered depending on whether one deals with
linear or nonlinear curvature.  Then, finally we shall present
results when only two codebooks (position and angle) are active and
the curvature has been discarded.
Results

      We run our experiments on an 81-character vocabulary task
involving discretely written letters, digits, and symbols
(mathematical, punctuation).  The data was provided by 8 writers,
among them one left-h...