Browse Prior Art Database

Improved Prototype Establishment in Online Handwriting Recognition

IP.com Disclosure Number: IPCOM000119377D
Original Publication Date: 1991-Jan-01
Included in the Prior Art Database: 2005-Apr-01
Document File: 2 page(s) / 84K

Publishing Venue

IBM

Related People

Chefalas, TE: AUTHOR [+2]

Abstract

Many online, handwriting recognition systems match an unknown character against prototype characters (1-5). Such systems usually represent each way of writing a character by a single prototype that often is one writing of the character. This minimizes the number of prototypes and, therefore, the computation time for matching.

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

Improved Prototype Establishment in Online Handwriting Recognition

      Many online, handwriting recognition systems match an
unknown character against prototype characters (1-5).  Such systems
usually represent each way of writing a character by a single
prototype that often is one writing of the character.  This minimizes
the number of prototypes and, therefore, the computation time for
matching.

      Our recognition system (6,7) collects original character
prototypes from a user's writing samples through a training scenario.
Averaged prototypes are formed by averaging original character
prototypes of the same label and shape (within a match threshold).
For example, similarly shaped 2-stroke A's are averaged to yield an
averaged A prototype.  Averaging previously operated on a global
basis and did not take into account the match distance to neighboring
characters having different labels.

      Disclosed is an incremental-averaging method that improves
shape-recognition accuracy by allowing more than one prototype for
characters that are similar to other characters, like 2-Z and U-V.
The method establishes prototypes by  processing each training
character sequentially.  A character is added to the prototype set if
it is new or is not recognized correctly within a fixed threshold.
Otherwise, the character is averaged into the closest prototype of
its class.

      This procedure was implemented in C-language in our recognition
system and tested on the IBM RT PC.  Results were obtained on the
Schoonard copy-task data (8).  The test data contain 474 characters
from each of nine writers; the alphabet consists of 44 characters of
uppercase, digits, and special symbols.  The system was trained on
965 characters per writer.  In an earlier experiment on these data a
quarter of an error rate of 2.7 percent was judged to be caused by
the shape recognizer, with remaining error attributable to hardware
or the user (8).  The disclosed method decreased the error
attributable to the shape recognizer by 31 perce...