Browse Prior Art Database

More Robust Descriptor of the Spatial Relationships between Strokes for Statistical Handwriting Recognition

IP.com Disclosure Number: IPCOM000110593D
Original Publication Date: 1992-Dec-01
Included in the Prior Art Database: 2005-Mar-25
Document File: 3 page(s) / 137K

Publishing Venue

IBM

Related People

Bellegarda, EJ: AUTHOR [+4]

Abstract

Proposed here is an improvement on the definition of the formerly defined dynamic interstroke distance by (i) normalizing for stroke direction, and (ii) determining a stroke-dependent origin to handle more appropriately the first stroke in a character. The new definition provides a more robust description of the spatial relationships between strokes, as measured by the error rate of a statistical handwriting recognition system.

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

More Robust Descriptor of the Spatial Relationships between Strokes for Statistical Handwriting Recognition

       Proposed here is an improvement on the definition of the
formerly defined dynamic interstroke distance by (i) normalizing for
stroke direction, and (ii) determining a stroke-dependent origin to
handle more appropriately the first stroke in a character.  The new
definition provides a more robust description of the spatial
relationships between strokes, as measured by the error rate of a
statistical handwriting recognition system.

      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.

      Previously in [*] a global feature parameter encompassing
information regarding the (dynamic) interstroke distance was defined.
This parameter, found adequate to achieve character shape tracking,
was defined as:

                            (Image Omitted)

where Pn is the current point in the stroke and d(Pn,PN) is the
distance between Pn and the last point PN in the previous stroke.
Note that when processing the first stroke in a character, the
parameter reduces to the norm of the point.

      There are two potential shortcomings with such a definition.
First, in the case of multi-stroke characters, intdist relies too
heavily on the last point of each stroke.  This point may drastically
vary in location depending on the stroke direction chosen by the
writer.  This is best illustrated in the figure which shows a
2-stroke lower case p. In (a) and (b) the values of d(Pn,PN) are very
different although dealing with the same shape.

      Second, there is a potential disproportion between the values
of intdist corresponding to the first stroke and its value for
subsequent strokes.  In a typical situation, the norm of Pn is an
order of magnitude greater than d(Pn,PN).

      Proposed is an alternative definition called middist, which
relies on the distance between the current point in the stroke and
the center of gravity of the previous stroke, in such a way that its
value is similar for cases (a) and (b) of the figure.  Also, to
remedy to the disproportion predicament when dealing with the first
stroke of a character, we propose to replace Pn by the distance
between Pmin and Pn, where Pmin is positioned by the minimum x and
minimum y in the first stroke.

      Thus, at each equispaced point Pn of coordinates (xn,yn) one
forms a 3-D feature vector of feature elements representing the
global pen trajectory up to Pn, and composed of (i) the vertical
coordinate yn (height from the baseline), (ii) the horizontal
displacement xn...