Browse Prior Art Database

Pattern Recognition Method for Personal Computers

IP.com Disclosure Number: IPCOM000114745D
Original Publication Date: 1995-Jan-01
Included in the Prior Art Database: 2005-Mar-29
Document File: 4 page(s) / 225K

Publishing Venue

IBM

Related People

Linsker, R: AUTHOR [+2]

Abstract

Disclosed is a pattern recognition technique, for use in Personal Computers (PCs) applications, where the characters being read are considered to be of low quality. The technique utilizes methods such as: The Gabor Wavelet Convolution; The Neural Network Pattern Recognition; and The Character String Location methods (1,2,3).

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 25% of the total text.

Pattern Recognition Method for Personal Computers

      Disclosed is a pattern recognition technique, for use in
Personal Computers (PCs) applications, where the characters being
read
are considered to be of low quality.  The technique utilizes methods
such as: The Gabor Wavelet Convolution; The Neural Network Pattern
Recognition; and The Character String Location methods (1,2,3).

      The technique uses character string location methodology for
automatically locating alpha-numeric character strings in noisy
images.  It is an improvement over both text character location
methods and window variance methods neither of which contain
direction sensing features.  It is also an improvement over
conventional algorithmic methods in that the technique used can
tolerate noisy images and incomplete characters.

      The main feature of this concept is the use of Gabor wavelet
convolution and neural network pattern recognition to locate
alpha-numeric characters in a specific range.  In prior art, known
implementations of the Gabor wavelet convolution and neural network
pattern matching were limited to image texture segmentation and did
not include alpha-numeric characters.

      Optical Character Recognition (OCR) of standard readable
characters in various sizes, fonts, and colors on varied backgrounds
exposed to surface dirt, rust and scratches in harsh environments,
typically involves the following steps:
  o  First, the object must be illuminated satisfactorily and the
      image captured by a camera, or linear array scanner, must be
      converted to digital brightness levels.  After an image is
      digitally formed, it is pre-processed to adjust contrast and
      average brightness.
  o  Next, the locations of relevant identification characters are
      tentatively located, in order, from the most likely to the
least
      likely.  Then the individual characters are segmented from each
      other within the text rectangles.  Character recognition is
      performed after segmentation.
  o  Finally, the recognized characters are checked by contextual
      cues, or against a data base, to see if they represent a
      meaningful identification string.

      Often, recognition of a corporate identifier provides the
simplest indication of the validity of the text location operation.
Item type suffixes, such as a "U" character, as used for intermodal
shipping containers, provide simple validation.  Identification
numbers may have check characters to improve reliability.  If the
recognized characters are not part of an identification string, the
system selects another tentative character string location, and the
process is repeated until an identification string is recognized.

      Gabor Wavelet Convolution - Computational models of biological
vision, in conjunction with artificial neural network simulations,
have suggested methods of use in machine v...