Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
1999-09-30
2002-11-19
Mancuso, Joseph (Department: 2623)
Image analysis
Pattern recognition
Feature extraction
C382S203000, C382S205000, C382S266000
Reexamination Certificate
active
06483943
ABSTRACT:
FIELD OF THE INVENTION
This invention relates to an image processing method and, more particularly, to a method for extracting features of a picture element (PEL) array pattern by means of software.
BACKGROUND OF THE INVENTION
In pattern recognition such as optical character recognition, various features of patterns to be recognized have been extracted for use in classification of the patterns such as characters based on these features. For example, T. Kawatani, “Handprinted Numerals Recognition by Learning Distance Function,” The Transactions of the Institute of Electronics, Information and Communication Engineers, September 1993, Vol. J76-D-2, No. 9, pp. 1851-1859 (in Japanese) discloses a technique adapted for determining attributes of a point on a contour in a PEL array pattern, such as a designation of concavity/convexity of the point indicating whether it is concave or convex, a direction of the concavity/convexity and a strength of the concavity/convexity (acuteness), whereby they may be utilized for calculating feature values of the PEL array pattern.
In this prior technique, however, a tangent line direction of a point of interest on a contour is simply regarded as a direction of the concavity/convexity of its contour portion, without giving any attribute of direction to a portion that has strong acuteness. While this prior technique enables to obtain a stable value for a direction of the concavity/convexity at a portion where a contour shape is approximate to a straight line as shown in
FIG. 16
, this does not apply to a portion where a contour shape has a strong curvature as shown in FIG.
17
. This is because, in the illustrated situation, a slight positional change of a center point will cause a direction of the concavity/convexity to be radically changed, which will in turn result in an excessive variation of feature values to be extracted from a PEL array pattern. Thus, what has been needed in this technical field is a novel scheme that is capable of extracting features in a more stable manner.
Japanese Patent Publication No. 6-19551 (corresponding to U.S. Pat. No. 5,590,220), assigned to the present assignee, discloses a technique adapted for determining attributes of bending points (identified by the letters “a” through “g” in
FIG. 18
) on a contour in a PEL array pattern, such as a designation of concavity/convexity of a bending point of interest indicating whether it is concave or convex, a direction of the concavity/convexity and a strength of the concavity/convexity (acuteness) as shown in
FIG. 19
, whereby they may be utilized for calculating feature values of the PEL array pattern.
According to this prior technique, a direction of the concavity/convexity is calculated solely at a portion where a contour shape has a strong curvature as shown in
FIG. 18 and
, thus, it avoids said problem of an excessive variation of feature values to be extracted from a PEL array pattern. Nevertheless, this prior technique requires further processing such as dividing contour groups at each of the selected bending points and, hence, it eventually requires a large amount of calculation time for extracting a sufficient quantity of feature values to be used in pattern recognition. Further, this prior technique entails difficulty in taking a strong convexity or concavity in the graphic of “0” and/or “0” that is approximate to a circle, thereby causing an occurrence position of a feature value to be unstable.
SUMMARY OF THE INVENTION
It is, therefore, an object of this invention to provide an image recognition system that is capable of deriving stable feature values, without depending on a contour shape of a graphic.
It is another object of this invention to quickly extract a sufficient amount of feature values for use in effective pattern (character and/or graphic) recognition of high recognition rates.
In accordance with this invention, one or more contours of a PEL array pattern are detected and, then, a strength of concavity/convexity on each contour (acuteness) and a direction of the concavity/convexity are determined. Using an acuteness value of a contour point of interest, its curvature is classified into a group of strong curvatures or another group of weak curvatures such that the definition of a direction of the concavity/convexity (contour direction) is appropriately switched from one to another for calculation. In a preferred aspect of this invention, using an acuteness value of a contour point of interest, its contour curvature is classified into either one of five curvature groups, comprising “strong concavity,” “weak concavity,” “straight line/inflection point,” “weak convexity,” and “strong convexity.” If its contour curvature is classified into either “strong concavity” or “strong convexity,” then calculation of a normal line direction will be made for deriving a contour direction. On the other hand, if its contour curvature is classified into “straight line/inflection point,” then calculation of a tangent line direction will be made. In so doing, it is possible to improve image recognition rates as compared to those of the prior schemes utilizing conventional feature values.
In accordance with one aspect of this invention, there is provided a method for analyzing one or more contours detected from a picture element (PEL) array pattern and extracting feature values of the PEL array pattern, comprising the steps of:
(a) tracing the contours of said PEL array pattern and generating a list of a plurality of contour points that exist on each traced contour;
(b) calculating, for one of said plurality of contour points, an acuteness value representing a contour curvature;
(c) if said acuteness value indicates an acute curve, then using a first calculation equation to calculate a contour direction of said one contour point, and if said acuteness value indicates a straight line or an inflection point, then using a second calculation equation other than the first calculation equation to calculate a contour direction of said one contour point; and
(d) outputting a feature value based on the contour direction of said one contour point.
Note that, in the claims of the present specification, the expression of “first calculation equation” represents a concept covering any equation adapted for calculating a direction that is approximate to a normal line or another direction calculated therefrom, such as a direction of a straight line connecting a contour point of interest and a midpoint between an Nth previous point and an Nth (or Mth) following point of the contour point of interest, as described below for a preferred embodiment. Similarly, the expression “second calculation equation” represents a concept covering any equation adapted for calculating a direction that is approximate to a tangent line or another direction calculated therefrom, such as a direction of a straight line connecting the Nth previous point and Nth (or Mth) following point of the contour point of interest, as described below for a preferred embodiment.
In accordance with another aspect of this invention, there is provided a feature extraction method, wherein said first calculation equation makes a calculation based on coordinate values of said one contour point P
i
, coordinate values of an Nth previous point P
i−N
and coordinate values of an Nth following point P
i+N
of said one contour point, and wherein said second calculation equation makes a calculation based on the coordinate values of the Nth previous point P
i−N
and the coordinate values of the Nth following point P
i+N
of said one contour point.
In accordance with another aspect of this invention, there is provided a method for analyzing one or more contours detected from a picture element (PEL) array pattern and extracting feature values of the PEL array pattern, comprising the steps of:
(a) tracing the contours of said PEL array pattern and generating a list of a plurality of contour points that exist on each traced contour;
(b) calculating, for one contour point P
i
among said plurality of contour points, an a
Dastouri Mehrdad
International Business Machines - Corporation
Mancuso Joseph
Otterstedt Paul J.
Ryan & Mason & Lewis, LLP
LandOfFree
Feature value extraction methods and apparatus for image... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Feature value extraction methods and apparatus for image..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Feature value extraction methods and apparatus for image... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2965351