Accuracy of recognition by means of a combination of classifiers

• US 8,660,371 B2
• Filed: 05/06/2010
• Issued: 02/25/2014
• Est. Priority Date: 05/06/2010
• Status: Expired due to Fees

• Current Assignee: ABBYY Development LLC
• Original Assignee: ABBYY Development LLC
• Inventors: Tuganbaev, Diar
• Primary Examiner(s): Bella, Matthew
• Assistant Examiner(s): CONWAY, THOMAS A
• Application Number:
  US12/775,445
• Publication Number:
  US 20110274345A1
• Time in Patent Office: 1,391 Days
• Field of Search: None
• US Class Current: 382/228
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06V 30/10 Character recognition

Abstract

In one embodiment, there is provided a method for an Optical Character Recognition (OCR) system. The method comprises: recognizing an input character based on a plurality of classifiers, wherein each classifier generates an output by comparing the input character with a plurality of trained patterns; grouping the plurality of classifiers based on a classifier grouping criterion; and combining the output of each of the plurality of classifiers based on the grouping.

Claims

1. A method for an optical character recognition (OCR) system, the method comprising:

training each of a plurality of classifiers by selecting separately for each of a plurality of characters weight coefficients for a weighted mean;

recognizing an input character based on the plurality of classifiers;

generating from each classifier an output by comparing the input character with a plurality of training patterns;

grouping the plurality of classifiers based on a classifier grouping criterion; and

combining the output of each of the plurality of classifiers based on the grouping, wherein combining the output of classifiers further comprises;

combining a weight output by each classifier into an interim weight in accordance with a formula for a weighted mean,

using a formula for calculating the weighted mean which is selected for each character on a basis of a preliminary training process based upon the plurality of training patterns, and

subsequently combining the interim weights from the calculated weighted mean with corresponding determined classifiers so as to provide a complete combined final set of weighted classifiers.

2. The method of claim 1, wherein the classifier grouping criterion includes a measure of a similarity between the classifiers.

3. The method of claim 1, wherein the combining a weight output by each classifier of at least one same type group into at least one interim weight in accordance with a formula for a weighted mean is performed as part of a first stage, wherein the interim weight is a final weight in a case of there being only one type group at the first stage, and wherein there is no second stage.

4. The method of claim 1, wherein the combining comprises a second stage in which each interim weight is combined with one of a further weight output by a classifier of a second type and another interim weight, wherein the second stage is performed multiple times in which case for each additional time, and wherein the interim weight is the output of a last performance of the second stage.

5. The method of claim 1, wherein the complete combined final set of weights maps weight pairs (x,x_n+1) to a final weight (w), wherein the final weights are selected such that a combination of weights (w, w) has a same percentage of right and wrong images as a combination (x,x_n+1), and wherein correspondences (x,x₊₁)−

w are selected based on a technique that maintains a percentage of right and wrong images associated with each weight pair;

where x is the interim weight, and x₊₁ the weight output by the classifier of the second type.

6. The method of claim 1, wherein the classifier of the second type is a structure classifier.

7. The method of claim 1, wherein the preliminary training process comprises selecting the weight coefficients which best distinguish images of the character (for which weight coefficients are being selected) from other images.

8. The method of claim 7, wherein the weight coefficients are selected on a basis of searching an extremum of a function, which distinguishes the images as a hyperplane.

9. The method of claim 8, wherein the function is expressed as f=j×

j2+j(1−

yj)2, where xj=i=1naixji, yj=i=1naiyji, ##EQU00008## and a.sub.i are weight coefficients, x.sub.ji is the set of all weights for a character j generated by each classifier i=1 to n by comparing with right character images, and y.sub.ji is the set of all weights for the same character j generated by each classifier i=1 to n by comparing with wrong character images.

10. An optical character recognition (OCR) system comprising:

at least one memory; and

at least one processor controlled by stored programmed instructions in the at least one memory that cause the OCR system to;

train each of a plurality of classifiers by selecting separately for each of a plurality of characters weight coefficients for a weighted mean;

recognize an input character based on the plurality of classifiers;

generate an output from each classifier by comparing the input character with a plurality of trained patterns;

group the plurality of classifiers based on a classifier grouping criterion; and

combine the output of each of the plurality of classifiers based on the grouping, wherein combining the output of classifiers further comprises;

combining a weight output by each classifier of at least one same type group into at least one interim weight in accordance with a formula for a weighted mean,

using a formula for calculating the weighted mean which is selected for each character on a basis of a preliminary training process based upon the plurality of training patterns, and

subsequently combining the interim weights from the calculated weighted mean with corresponding determined classifiers so as to provide a complete combined final set of weighted classifiers.

11. The OCR system of claim 10, wherein the classifier grouping criterion includes a measure of a similarity between the classifiers.

12. The OCR system of claim 10, wherein the classifier of the second type is a structure classifier.

13. The OCR system of claim 10, wherein the combining comprises a second stage in which each interim weight is combined with one of a further weight output by a classifier of a second type and another interim weight.

14. The OCR system of claim 13, wherein the second stage is performed multiple times in which case for each additional time, the interim weight is the output of a last performance of the second stage.

15. The OCR system of claim 13, wherein the complete combined final set of weights maps weight pairs (x,x_n+1) to a final weight (w), wherein the final weights are selected such that a combination of weights (w, w) has a same percentage of right and wrong images as a combination (x,x_n+1), and wherein correspondences (x,x_n+1)−

w are selected based on a technique that maintains a percentage of right and wrong images associated with each weight pair;

where x is the interim weight, and x_n+1 is the weight output by the classifier of the second type.

16. The OCR system of claim 10, wherein the preliminary training process includes selecting the weight coefficients which best distinguish images of the character (for which weight coefficients are being selected) from other images.

17. The OCR system of claim 16, wherein the weight coefficients are selected on a basis of searching an extremum of a function, which distinguishes the images as a hyperplane.

18. The OCR system of claim 17, wherein the function is expressed as f=j×

j2+j(1−

yj)2, where xj=i=1naixji, yj=i=1naiyji, ##EQU00009## and a.sub.i are weight coefficients, x.sub.ji is the set of all weights for a character j generated by each classifier i=1 to n by comparing with right character images, and y.sub.ji is the set of all weights for the same character j generated by each classifier i=1 to n by comparing with wrong character images.

19. A computer readable medium having stored thereon a sequence of instructions which when executed by a computer, causes the computer to perform a method for an Optical Character Recognition (OCR) system, the method comprising:

training each of a plurality of classifiers by selecting separately for each of a plurality of characters weight coefficients for a weighted mean;

recognizing an input character based on the plurality of classifiers;

generating from each classifier an output by comparing the input character with a plurality of trained patterns;

grouping the plurality of classifiers based on a classifier grouping criterion; and

combining the output of each of the plurality of classifiers based on the grouping, wherein combining the output of classifiers comprises;

combining a weight output by each classifier of at least one same type group into at least one interim weight in accordance with a formula for a weighted mean,

using a formula for calculating the weighted mean which is selected for each character on a basis of a preliminary training process based upon the plurality of training patterns, and

subsequently combining the interim weights from the calculated weighted mean with corresponding determined classifiers so as to provide a complete combined final set of weighted classifiers.

20. The computer readable medium of claim 19, wherein the classifier grouping criterion includes a measure of a similarity between the classifiers.

21. The computer readable medium of claim 19, wherein the combining a weight output by each classifier of at least one same type group into at least one interim weight in accordance with a formula for a weighted mean is performed as part of a first stage, wherein the interim weight is a final weight in a case of there being only one type group at the first stage and no second stage.

22. The computer readable medium of claim 19, wherein the complete combined final set of weights maps weight pairs (x,x_n+1) to a final weight (w), wherein the final weights are selected such that a combination of weights (w, w) has a same percentage of right and wrong images as a combination (x,x_n+1), and wherein correspondences (x,x_n+1)−

w are selected based on a technique that maintains a percentage of right and wrong images associated with each weight pair;

where x is the interim weight, and x_n+1 is the weight output by the classifier of the second type.

23. The computer readable medium of claim 19, wherein the combining comprises a second stage in which each interim weight is combined with one of a further weight output by a classifier of a second type and another interim weight.

24. The computer readable medium of claim 23, wherein the second stage is performed multiple times in which case for each additional time, the interim weight is the output of a last performance of the second stage.

25. The computer readable medium of claim 24, wherein the classifier of the second type is a structure classifier.

26. The computer readable medium of claim 19, wherein the preliminary training process comprises selecting the weight coefficients which best distinguish images of the character (for which weight coefficients are being selected) from other images.

27. The computer readable medium of claim 26, wherein the weight coefficients are selected on a basis of searching an extremum of a function, which distinguishes the images as a hyperplane.

28. The computer readable medium of claim 27, wherein the function is expressed as f=j×

j2+j(1−

yj)2, where xj=i=1naixji, yj=i=1naiyji, ##EQU00010## and a.sub.i are weight coefficients, x.sub.ji is the set of all weights for a character j generated by each classifier i=1 to n by comparing with right character images, and y.sub.ji is the set of all weights for the same character j generated by each classifier i=1 to n by comparing with wrong character images.