Browse Prior Art Database

Flexible Adaptation Method with Radii Training for the Recognition Library Management

IP.com Disclosure Number: IPCOM000105020D
Original Publication Date: 1993-Jun-01
Included in the Prior Art Database: 2005-Mar-19
Document File: 4 page(s) / 153K

Publishing Venue

IBM

Related People

Mohiuddin, M: AUTHOR [+2]

Abstract

Disclosed is a system for making a recognition library (e.g. printed recognition library, such as a library for printed Optical Character Recognition (OCR)), to easily adapt to the change of fonts (environment) that are significantly different from those in the generation phase of the initial master library. It uses radii for normalization of distance calculation and generates application specific libraries besides the master library. The specific libraries have only three values (unique identification numbers, ASCII code, radius) for each entry, so this system can easily adapts to the actual environment with a small additional cost.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 38% of the total text.

Flexible Adaptation Method with Radii Training for the Recognition Library Management

      Disclosed is a system for making a recognition library (e.g.
printed recognition library, such as a library for printed Optical
Character Recognition (OCR)), to easily adapt to the change of fonts
(environment) that are significantly different from those in the
generation phase of the initial master library.  It uses radii for
normalization of distance calculation and generates application
specific libraries besides the master library.  The specific
libraries have only three values (unique identification numbers,
ASCII code, radius) for each entry, so this system can easily adapts
to the actual environment with a small additional cost.

      Clustering methods are used for generation of prototype
libraries in OCR, voice recognition, etc.  To illustrate the need for
clustering, consider omnifont machine-print OCR.  Several styles of
A's, B's, etc., may need to be recognized in the target documents.
Hence, during a training phase a number of samples of different fonts
of various characters are collected and grouped into clusters based
on their similarity.  During the recognition phase, characters from
the target document are compared with representatives of the various
clusters and identities assigned based on similarity.  Unfortunately,
the target documents may contain fonts that were not present in the
training set, or their frequencies of occurrence may be quite
different from the training set.  These can cause the recognition
accuracy to be seriously impaired.

      The key for improving recognition rate is the adaptation of a
recognition library.  Even if there are a variety of fonts, almost
arbitrarily-shaped aggregate clusters (combining sub-clusters for
single category) can be formed by adding new clusters and tuning the
radius sizes.

      First, the initial generation of a recognition library using a
large quantity of sample data is described.  Secondly, the addition
of new templates and their adaptation by radii-training is described.

      Initial library generation - The recognition library generation
consists of the following steps:

1.  Collect as many sample character images as possible that
    preferably reflect the frequency of fonts in actual large
    applications.

2.  Execute some clustering (most existing methods are applicable) on
    the sample characters for each category independently from other
    categories, and generate all the templates (clusters).  This is
    done in n-dimensional feature space.  Initialize the radius of
    each of the clusters to the same value.

3.  Execute the recognition for all the training samples.  In the
    recognition, the distance (e.g., Euclidean) between an input
    feature and all the templates are calculated and normalized by
    the radius of the clusters.  The template ASCII code associated
    with the shortest norma...