System for recognizing handwritten character strings containing overlapping and/or broken characters

• US 5,321,768 A
• Filed: 09/25/1992
• Issued: 06/14/1994
• Est. Priority Date: 09/22/1992
• Status: Expired due to Fees

• Current Assignee: The Research Foundation for The State University of New York (State University of New York)
• Original Assignee: The Research Foundation for The State University of New York (State University of New York)
• Inventors: Fenrich, Richard K., Srihari, Sargur N.
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): PRIKOCKIS, LARRY J
• Application Number:
  US07/951,287
• Time in Patent Office: 627 Days
• Field of Search: 382/9, 382/13, 382/1, 382/57
• US Class Current: 382/178
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/148 Segmentation of character regions

Abstract

A system (method and apparatus) for handwritten character recognition recognizes characters within strings of text despite the presence of overlapping or disjointed characters. A string is segmented into discrete characters by removing noise and punctuation, joining the disconnected components of characters and splitting probable overlapping characters. After joining and splitting (as required), character recognition is performed to recognize characters in the string. Recognized characters are removed and the string is split in order that more characters can be recognized. If unrecognized characters remain in the string, the process is repeated. The joining and splitting operations are repeated until all subsequent characters in the string have been either recognized or determined to be unrecognizable.

Claims

1. A method of recognizing characters in a string of handwritten text, comprising the steps of:

(a) providing an array of pixels corresponding to said string;

(b) estimating the number of characters contained in said string;

(c) partitioning said array into a plurality of groups of rows of said pixels;

(d) vertically partitioning each of said groups of rows in order to separate occupied columns within each of said groups of rows from unoccupied columns, said columns being occupied if they contain a foreground pixel and being unoccupied if they contain no foreground pixel;

(e) designating as a component a plurality of said columns having a foreground pixel if said plurality is contiguous, said component thereby forming a discrete pattern;

(f) removing non-character components from each of said groups of rows;

(g) recognizing said discrete patterns as characters;

(h) removing from said array said recognized discrete patterns having a predetermined confidence level of corresponding to one of said characters; and

(i) reestimating said number of characters contained in said string to obtain a reestimated number of characters.

2. The method according to claim 1, further comprising repeating said steps (c)-(i) a predetermined number of times or until a predetermined percentage of said estimated number of characters is recognized, whichever occurs first.

3. The method according to claim 2, further comprising estimating a number of characters contained in each component and splitting ones of said components not recognized as one of said characters if said splitting is highly probable of yielding a character, said unrecognized components being split such that the resulting number of characters equals said reestimated number of characters.

4. The method according to claim 1, wherein said reestimating is performed using constrained linear regression.

5. The method according to claim 1, wherein said recognizing said discrete patterns is performed using a polynomial discriminant method.

6. The method according to claim 1, further comprising removing ligatures from said string.

7. The method according to claim 1 wherein said providing an array of pixels corresponding to said string comprises:

optically scanning said string to create a pixel image of said string; and

thresholding said pixel image to create a bit map of said string.

8. The method according to claim 1 wherein said estimating the number of characters contained in said string comprises:

determining the vertical length of said string;

determining the horizontal length of said string; and

determining said estimated number of characters from the ratio of said vertical length to said horizontal length.

9. The method according to claim 1, wherein said removing non-character components from each of said groups comprises:

changing to background pixels each of said foreground pixels forming a non character component.

10. The method according to claim 1 further comprising:

after said removing non-character components, grouping components in accordance with predetermined proximity criteria.

11. The method according to claim 1 wherein said splitting comprises:

locating one or more splitting neighborhoods in said components wherein said splitting neighborhoods are one or more points in said string where adjoining characters intersect;

counting the number of said foreground pixels in each of said columns having a foreground pixel in each of said splitting neighborhoods; and

separating each of said components at a column having a predetermined foreground pixel profile.

12. The method according to claim 11, wherein said predetermined pixel profile comprises a right-most column not having two or more of said foreground pixels.

13. The method according to claim 3 wherein said splitting comprises:

locating one or more splitting neighborhoods in said components;

identifying an upper contour of each of said components by locating an uppermost foreground pixel in each column of said component in each of said splitting neighborhoods;

identifying a lower contour of each of said components by locating a lower most foreground pixel in each column of said component in each of said splitting neighborhoods; and

separating each of said components on a line generally between a lower most foreground pixel in said upper contour and an uppermost foreground pixel in said lower contour in each of said splitting neighborhoods.

14. The method according to claim 3 wherein said splitting comprises:

locating one or more splitting neighborhoods in said components;

identifying an upper contour of each of said components by locating an uppermost foreground pixel in each column of said component in each of said splitting neighborhoods; and

separating each of said components on a line including a lower most foreground pixel in said upper contour in said splitting neighborhoods.

15. The method according to claim 3 wherein said splitting comprises:

locating one or more splitting neighborhoods in said components;

identifying a lower contour of each of said components by locating a lower-most foreground pixel in each column of said component in each of said splitting neighborhoods; and

separating said components on a line including an uppermost foreground pixel in said lower contour in each of said splitting neighborhoods.

16. The method according to claim 11 wherein said locating comprises:

dividing the number of pixel columns in said component by said estimated number of characters to form a neighborhood multiplier;

designating as a splitting neighborhood a group of said pixel columns centered at a pixel column whose location corresponds to a multiple of said neighborhood multiplier;

counting foreground pixels in each column of pixels within said component; and

assigning as a split point of column of pixels having the fewest number of foreground pixels.

17. The method according to claim 13 wherein said locating further comprises:

dividing the number of pixel columns in said component by said estimated number of characters to form a neighborhood multiplier;

designating as a splitting neighborhood a group of said pixel columns centered at a pixel column corresponding to a multiple of said neighborhood multiplier;

identifying an upper contour of said foreground pixels in said components;

identifying a lower contour of said foreground pixels in said components

locating a low point on said upper contour in each of said neighborhoods;

locating on said lower contour a high point in each of said neighborhoods;

separating said components generally along a line between said high and low point in one of said neighborhoods.

18. The method according to claim 14 wherein said locating further comprises:

dividing the number of pixel columns in said component by said estimated number of characters to form a neighborhood multiplier;

designating as a splitting neighborhood a group of said pixel columns centered at a pixel column corresponding to a multiple of said neighborhood multiplier;

identifying an upper contour of said foreground pixels;

locating a low point on said upper contour in each of said neighborhoods, said low point being a split point.

19. The method according to claim 15 wherein said locating further comprises:

dividing the number of pixel columns in said component by said estimated number of characters to form a neighborhood multiplier;

designating as a splitting neighborhood a group of said pixel columns centered at a pixel column corresponding to a multiple of said neighborhood multiplier;

identifying a lower contour of said foreground pixels; and

locating a high point on said lower contour in each of said neighborhoods, said high point being a split point.

20. The method according to claim 1, further comprising:

when said estimation of the number of characters is incorrect, changing said estimation to be one less than said incorrect estimation and repeating the steps of claim 1.

21. The method according to claim 1, further comprising:

when said estimation of the number of characters is incorrect, changing said estimation to be one more than said incorrect estimation and repeating the steps of claim 1.

22. The method according to claim 1, further comprising grouping ones of said components if they satisfy a predetermined proximal relationship.

23. Apparatus for recognizing characters in a string of handwritten text, comprising:

means for providing a patterned array of pixels corresponding to said string;

means for estimating the number of characters contained in said string;

means for partitioning said array into a plurality of groups of rows of said pixels;

means for vertically partitioning each of said groups of rows to separate occupied columns within each of said groups of rows from unoccupied columns, said columns being occupied if they contain a foreground pixel and being unoccupied if they do not contain a foreground pixel;

means for removing non-character components from each of said groups of rows;

means for designating as a component a plurality of said columns having a foreground pixel if said plurality is contiguous, said component thereby forming a discrete pattern;

means for recognizing said discrete patterns as characters;

means for removing from said array ones of said components having a predetermined confidence level of forming a character; and

means for reestimating said number of characters contained in said string to obtain a reestimated number of characters.

24. The apparatus according to claim 23, wherein said means for reestimating uses constrained linear regression.

25. The apparatus according to claim 23, wherein said means for recognizing uses a polynominal discriminant method.

26. The apparatus according to claim 23, further comprising:

means for estimating the number of characters in one of said components;

means for splitting into a predetermined number of characters ones of said components not recognized as one of said characters if said unrecognized components can be split to yield highly probable characters, said predetermined number of characters depending on the estimated number of characters in said string;

means for recognizing said split components as characters;

means for removing from said array each of said split components having a predetermined confidence level of forming one of said characters; and

means for grouping ones of said components satisfying a predetermined proximal relationship.

27. Apparatus for recognizing characters in a string of handwritten text, comprising:

means for providing a patterned array of pixels corresponding to said string;

means for estimating the number of characters contained in said string;

means for partitioning said array into a plurality of groups of rows of said pixels;

means for vertically partitioning each of said groups of rows to separate occupied columns from unoccupied columns, said columns being occupied if they contain a foreground pixel and being unoccupied if they do not contain a foreground pixel;

means for removing non-character components from each of said groups of rows;

means for designating as a component a plurality of said columns having a foreground pixel if said plurality is contiguous, said component thereby forming a discrete pattern;

means for recognizing said discrete patterns as characters;

means for removing from said array ones of said components having a predetermined confidence level of forming a character;

means for reestimating said number of characters contained in said string to obtain a reestimated number of characters; and

means for recognizing a word comprising said recognized characters.

28. The apparatus according to claim 27, further comprising:

means for estimating the number of characters in one of said components;

means for splitting into a predetermined number of characters ones of said components not recognized as one of said characters if said unrecognized components can be split to yield highly probable characters, said predetermined number of characters depending on the estimated number of characters in said string;

means for recognizing said split components as characters;

means for removing from said array each of said split components having a predetermined confidence level of forming a character; and

means for grouping ones of said components satisfying a predetermined proximal relationship.

29. The apparatus according to claim 27, further comprising means for sorting a mail piece on which said words are written in response to the results of said recognition.