Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
1999-07-02
2001-05-29
Chang, Jon (Department: 2623)
Image analysis
Pattern recognition
Feature extraction
C382S200000, C382S201000
Reexamination Certificate
active
06240209
ABSTRACT:
A. BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention is in the field of pattern recognition. It relates to a system for recognising characters, such as digits and letters, both in handwritten and printed form, and more particularly to a method for deriving features of a character to be recognised in a character recognition system of this type.
2. Prior Art
Character recognition systems such as those indicated above usually comprise a number of process steps which are described below with reference to a block diagram shown in
FIG. 1
a
. Said process steps are:
(1) optical scanning of a character on the surface of a carrier; such scanning, for example with a video camera, results in a two-dimensional pixel pattern, for example 64×64 pixels, to which a grey value encoded in 5 bits is appended;
(2) binary quantization of the pixel pattern, by subjecting the grey values of all the pixels to a threshold operation, the result, which can be seen as a black/white pattern or a pattern of “zeros” and “ones”, being stored in a memory;
(3) deriving, from said quantized pixel pattern, a description of the character to be recognized;
(4) deriving, from the description, a set of character features of the character to be recognized, such as distance tags or structural characteristics to be discussed in more detail shortly with respect to pertinent prior art references;
(5) checking the set of character features found against results obtained previously with the aid of learning processes on known characters;
(6) deciding, on the basis of the check, which known character the character to be recognized is recognized as.
A recognition technique of this type is known, for example, from U.S. Pat. No. 4,566,124 entitled “Pattern Reading System”. In this technique the quantized pixel pattern is used to derive a contour description. The term ‘contour’ has a well-accepted meaning in the field of character recognition, with the discussion in U.S. Pat. No. 4,566,154 being representative of the connotation of the term ‘contour’ set forth in prior art references in this field. For instance, a ‘contour’, for the case of a black-and-white image pattern in an image plane, is a closed enveloping line encompassing closed areas of the same color (black or white). Accordingly, hereinafter, the term ‘contour’ conforms to the conventional usage of this term in the art, that is, what those of ordinary skill in the art would ascribe to its meaning and usage. As a contour of a digitized pattern of a feature to be recognized is followed, the coordinates of certain points are selected. Those points are selected whose inner product of the point vector corresponding to each point and a directional vector in a number of predefined directions has the highest value. A selected point is characterized as an “extreme point” if the difference between said inner product at that point and the inner product at the following selected point exceeds a previously established limit value. Said extreme points define a polygonal approximation of the contour of the pattern. Said polygon shows a pattern of convex and concave structures. Said pattern is used for comparison with similar patterns, which are characteristic of known characters and stored in a previously compiled “dictionary”. The character to be recognized is read as that known character whose patterns show the best correspondence.
A limitation of the pattern features derived in said known recognition technique is its being based on a structural description of the patterns to be recognized. This, however, produces acceptable recognition results only if the patterns to be recognized have a clear structure. This means that said known technique can be sensibly applied only for the recognition of characters from clear handwriting or typing. Another limitation is the variability of the number of features on which the recognition is based, i.e. the number of convex and concave structures in the pattern. This makes it difficult to apply deciders in the recognition, which work with a fixed number of features, such as, for example, those which make use of standard “feed-forward” neural networks.
U.S. Pat. No. 3,999,161 entitled “Method and Device for the Recognition of Characters Preferably of Figures,” discloses another character recognition system of the above mentioned type. The recognition technique, on which this system is based, derives character features from a number of so-called views of an image, either the complete image or parts thereof, of a binary quantized pixel pattern of a character to be recognized. In this so-called views method, a number of different features of the pixel pattern is derived for each view, from above, from below and from the side, and for each image or part image. For these features, such as jumps, slopes, endpoints and islands, feature values are determined for specific patterns found in the views. During the check phase, the feature values found are weighted using adjustable weighting factors for each class of characters to be recognized. Said weighting factors have been obtained previously from a learning process on known characters. The weighted feature values are used to determine a score for each class. The character to be recognized is recognized as that character which belongs to the class in which the highest score has been determined.
This known technique, using the views method for deriving character features, has been developed in the first instance for recognizing digits. The character features which are derived in this process have been found to be too general for acceptable letter recognition. Moreover, because of the many different features for which a character to be recognized has to be checked, the method for deriving the features is rather complex. Because of the type of the features, which are in fact structural features, a recognition based on said features is furthermore sensitive to breaks in an image pattern of a character to be recognized, and this recognition method is therefore unsuitable for, for example, matrix writing, such as, in particular, letters printed by a matrix printer.
In digital picture processing, it is also known to make use of distance functions. Methods based thereon are sometimes denoted as “distance transform” or “distance mapping”. A known “distance mapping” method of this type is, for example, disclosed by the paper entitled “Euclidean Distance Mapping” as published in Computer Graphics and Image Processing 14, pages 227-248 (1980) and authored by P.E. Danielsson. This involves assigning to each pixel, which may form part of a subset of pixels, in the object (or in the background) of the original binary image, the shortest distance to the nearest pixel in the background (or the object). Corresponding to each point there is thus a distance tag. Further processing is then carried out on the basis of this two-component description, in which processing, inter alia, skeleton structures of the character to be recognized being traced. A distance tag of this type is inadequate per se as a character feature for the purpose of, in particular, recognizing handwritten characters.
B. SUMMARY OF THE INVENTION
The aforementioned limitations as well as other shortcomings and deficiencies of the prior art are obviated, in accordance with the present invention, by a methodology and concomitant circuitry which utilize information about local views of a character description as determined relative to a set of reference points in the image plane.
Broadly, character features are derived for use in a character recognition system to recognize randomly occurring characters, such as letters and digits. Initially, a character description is generated from a recorded image pattern in an image plane wherein the image pattern is presumed to include a character pattern of a character to be recognized. The character description is expressed in terms of a set of boundary points, each boundary point being an image point which has different colors (e.g., black and white) of image points in its direct surroundings
Chang Jon
Koninklijke PTT Nederland N.V.
Michaelson Peter L.
Michaelson & Wallace
LandOfFree
Method for deriving character features in a character... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for deriving character features in a character..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for deriving character features in a character... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2553802