Browse Prior Art Database

Verifier for Similarly Shaped Characters

IP.com Disclosure Number: IPCOM000108743D
Original Publication Date: 1992-Jun-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 5 page(s) / 210K

Publishing Venue

IBM

Related People

Kamon, Y: AUTHOR [+2]

Abstract

Disclosed is a verification method for on-line Japanese handwriting recognition. In the recognition system, a discrimination step will output candidate characters sorted in ascending order of their distance from the input character. The first candidate, which has the shortest distance, is usually the correct result. However, as there are many similarly shaped characters, an additional verification mechanism is often required to improve the recognition accuracy. This article describes a simple verification mechanism.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 28% of the total text.

Verifier for Similarly Shaped Characters

       Disclosed is a verification method for on-line Japanese
handwriting recognition.  In the recognition system, a discrimination
step will output candidate characters sorted in ascending order of
their distance from the input character.  The first candidate, which
has the shortest distance, is usually the correct result.  However,
as there are many similarly shaped characters, an additional
verification mechanism is often required to improve the recognition
accuracy.  This article describes a simple verification mechanism.

      Generally, lists of ten candidates are sent from the
discriminator to the verifier.  In the verifier, two candidates are
selected and make a pair to be compared.  For every pair, the
probability of the correctness of each candidate is calculated by
using statistical information contained in a feature dictionary.  If
the difference between the probabilities is large enough (that is, if
the features are clearly distinguished), the verifier changes the
candidate order and improves the recognition accuracy.

      This article does not describe how pairs of candidates are
selected from the candidate lists.  The tournament method, in which
the winning candidate makes a new pair with a higher-ranked
candidate, is not a new idea, and is often implemented in commonly
used verifiers.  The following description is based on this idea.
1. Choosing Distinguishing Features

      Over a dozen features are used in the verifier.  All of them
are extracted from the input character.  Effective features that
clearly distinguish between a pair of candidates should be chosen
automatically.  The selection of effective features is described in
the next section. However, as these features are selected full
automatically, no human-assisted dictionary is required.  The
effective features are newly selected for every input character
according to the actual feature value of the character.  One merit of
the system is that only features that are actually effective for the
current input character are used in the verifier decision (not
features from a fixed list).  Another merit is that no change in the
selection mechanism is necessary when features are improved or new
features are added.
2. Calculating the Probabilities of Candidates

      When the number of features is n, the feature vector of the
input pattern can be described in the form (v1, v2, ... vR).  If a
pair of candidates, A and B, are chosen from the candidate lists, we
can find the probability of A being correct and that of B being
correct: the former is (PA1, PA2, ... PAR) and the latter is (PB1,
PB2, ... PBR), where Pxk is the probability when vk is applied as
candidate X.

      Before effective features are chosen, the probabilities are
scaled.  The following is a pseudo-code for the process of selecting
effective features:
DO i=1 to n
    IF ( the probabilities of both candidates are too s...