Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Application of Artificial Neural Networks to Fixed Font Character Recognition

IP.com Disclosure Number: IPCOM000121599D
Original Publication Date: 1991-Sep-01
Included in the Prior Art Database: 2005-Apr-03
Document File: 4 page(s) / 144K

Publishing Venue

IBM

Related People

Rawson, AR: AUTHOR

Abstract

Disclosed is a method of automatic character recognition suitable for machine-printed fixed-font characters. This method extends the prior art in the field of automatic character recognition by eliminating the highly time and labor intensive process of feature design normally required in development of an automatic character recognition system. Additionally, the implementation of the character recognition system is aided by bypassing the processing required for feature extraction.

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

Application of Artificial Neural Networks to Fixed Font Character
Recognition

      Disclosed is a method of automatic character recognition
suitable for machine-printed fixed-font characters.  This method
extends the prior art in the field of automatic character recognition
by eliminating the highly time and labor intensive process of feature
design normally required in development of an automatic character
recognition system. Additionally, the implementation of the character
recognition system is aided by bypassing the processing required for
feature extraction.

      Refer to Figures 1a and 1b for the following discussion. The
traditional method is diagrammed in Figure 1a, and the proposed
method is shown in Figure 1b.

      Both methods begin (1) with the building of a labeled training
set. This is a database containing exemplars of the sampled and
binarized images of the characters within a fixed set, each of which
is correctly labeled with a code which identifies the symbol. At this
point the two processes diverge.

      The next step (2) in the traditional method is feature/feature
extraction design. Feature design and feature extraction design are
interrelated subprocesses. A feature is really defined by the
algorithm which has been designed to extract it from the character
image.  Feature design is the definition of a specific set of salient
shape descriptors which are to be used by the classification process
within a character recognition system. Feature extraction design is
the specification of the algorithm applied to an input image.  The
output of this algorithm is a "yes/no" decision as to the presence of
a given feature. Practical feature design can not be carried out in
isolation from the task of designing the algorithm to extract the
features.

      Largely due to this dependency between feature design and
feature extraction, the process is an extremely labor intensive one,
often requiring a team of engineers to complete the task for a given
font.  To save processing time in the feature extraction process, the
number of bits tested in the input character image matrix must be
kept to a minimum.  This necessitates a careful trade-off between
reliability of the feature extraction process and its speed of
execution. Although aided by some batch-oriented computer programs,
the task of picking which bits to test in the input character image
matrix falls to the recognition system development engineer.

      After feature/feature extraction design this process calls for
the training of a classifier (3). This is an automated process. The
next task (4) is the execution of the present version of the
recognition system against a labeled test set which is independent of
the or...