Browse Prior Art Database

Application of Neural Network Technology to Segmentation of Handwritten Characters

IP.com Disclosure Number: IPCOM000116697D
Original Publication Date: 1995-Oct-01
Included in the Prior Art Database: 2005-Mar-31
Document File: 4 page(s) / 138K

Publishing Venue

IBM

Related People

Narasimha, MS: AUTHOR [+2]

Abstract

A string of handwritten characters will often contain elements that are fragmented or have been connected or made to overlap by the writer. In some cases (e.g., the numerals 00), the characters may be recognized successfully in the connected state. In general, however, it is necessary to perform a segmentation step before attempting recognition. When a numeral string contains a combination of connected numerals, fragmented numerals and background noise segmentation can be extremely difficult.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 42% of the total text.

Application of Neural Network Technology to Segmentation of Handwritten
Characters

      A string of handwritten characters will often contain elements
that are fragmented or have been connected or made to overlap by the
writer.  In some cases (e.g., the numerals 00), the characters may be
recognized successfully in the connected state.  In general, however,
it is necessary to perform a segmentation step before attempting
recognition.  When a numeral string contains a combination of
connected
numerals, fragmented numerals and background noise segmentation can
be
extremely difficult.

      Disclosed is a method for applying a computational neural
network in the segmentation of handwritten characters.  In this
method segmentation decisions are based upon functional values
computed by feed-forward neural networks (1).  The single output
value computed by each network represents a fuzzy decision value.
These decision values provide answers to questions such as:
  1.  Is this particular edge vertex formed as a result of the
       touching of vertically oriented strokes in adjacent
characters?
  2.  Is this edge vertex a good location to split two characters
       connected by the extension of a horizontally oriented stroke?
  3.  Should this particular fragment be associated with the
character
       on the left or on the right?
  4.  Does this pair of left and right end points of horizontally
       oriented strokes constitute the ends of a stroke dividing the
       numerator and denominator of a fractional value?
  5.  Is a horizontal stroke near the top of this numeral actually
the
       detached top of a "5" located immediately to the left?

      The input values presented to a neural network are measurements
associated with the physical features of the characters being
segmented.  For example, the features used as inputs to the networks
used in detecting and splitting connected numerals are essentially
those used in the heuristic approach described in (2).  Input feature
measurements are, however, always mapped into the domain (1.0, 1.0)
in a way that  the output decision function is monotonic with respect
to each input.  That is, the partial derivative of the output with
respect to any of n inputs is theoretically positive at any point in
the n-cube.

      The back propagation algorithm is used in training the neural
networks.  Characters used in the training of each network must be
manually selected by a human expert, and the segmentation system must
be able to compute and log the feature measurements of the selected
characters.  After feature measurements for a specific training set
are captured, the associated fuzzy output decision value for each set
of input features must be  manually assigned by the human expert.
When the target output values have been assigned, the back
propagation algorithm is used to derive the interconnection weights
of the network.

 ...