System and method for increasing the accuracy of optical character recognition (OCR)

• US 9,152,883 B2
• Filed: 11/02/2009
• Issued: 10/06/2015
• Est. Priority Date: 11/02/2009
• Status: Active Grant

• Current Assignee: Hyland Switzerland, Sarl
• Original Assignee: Johannes Hausmann, Ralph Meier, Thorsten Wanschura
• Inventors: Urbschat, Harry, Meier, Ralph, Wanschura, Thorsten, Hausmann, Johannes
• Primary Examiner(s): ENTEZARI, MICHELLE M
• Application Number:
  US12/588,928
• Publication Number:
  US 20110103688A1
• Time in Patent Office: 2,164 Days
• Field of Search: 382/155, 382/156, 382/157, 382/158, 382/159, 382/160, 382/161, 382/187, 382/321
• US Class Current: 1/1
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/12 Detection or correction of errors, e.g. by rescanning the pattern
• CPC Class Codes: G06V 30/1914 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries, e.g. user dictionaries

Abstract

A system and/or method for increasing the accuracy of optical character recognition (OCR) for at least one item, comprising: obtaining OCR results of OCR scanning from at least one OCR module; creating at least one OCR seed using at least a portion of the OCR results; creating at least one OCR learn set using at least a portion of the OCR seed; and applying the OCR learn set to the at least one item to obtain additional optical character recognition (OCR) results.

Claims

1. A method for increasing the accuracy of optical character recognition (OCR) for at least one item, comprising:

obtaining OCR results of OCR scanning from at least one OCR module;

creating at least one OCR seed using at least a portion of the OCR results, the at least one OCR seed comprising a plurality of imagelets corresponding to each character identified in the at least a portion of the OCR results, wherein the at least one OCR seed is cleaned by selecting imagelets similar to one another for each character identified in the at least a portion of the OCR results;

creating at least one OCR learn set using at least a portion of the OCR seed;

comparing the at least one OCR learn set to each imagelet to create at least one mismatch distribution of the at least one OCR learn set compared to each imagelet, the at least one mismatch distribution comprising at least one confidence rating including a confidence score for the imagelet compared to at least one possible character; and

applying the OCR learn set and the at least one mismatch distribution to the at least one item to obtain additional OCR results such that only possible characters having a confidence score higher than a threshold are considered when applying the at least one mismatch distribution to obtain the additional OCR results.

2. The method of claim 1, wherein image statistics are deducted based on the at least one OCR learn set.

3. The method of claim 1, wherein the at least one item comprises:

at least one page;

at least one document;

at least one part of at least one document;

at least one zone;

at least one paragraph;

at least one line;

at least one word;

or

at least one character;

or

any combination thereof.

4. The method of claim 1, wherein the at least one OCR learn set is used on different items.

5. The method of claim 1, wherein different OCR modules and/or different parameter sets and/or different preprocessing steps are used.

6. The method of claim 1, wherein the additional OCR results take into account:

font size;

font type;

or

font properties;

or

any combination thereof.

7. The method of claim 1, wherein whitespace characteristics are taken into account in creating the at least one OCR learn set.

8. The method of claim 1, wherein the method is performed iteratively to allow for a refinement of the method.

9. The method of claim 1, wherein the at least one OCR learn set comprises stored imagelets and/or new imagelets.

10. The method of claim 1, wherein stored imagelets are used to increase the at least one OCR learn set.

11. The method of claim 1, wherein the at least one OCR learn set and related statistics are used to derive at least one image filtering technique.

12. The method of claim 1, wherein the OCR learn set is created by obtaining information related to the image representation for each character and the variability of the similar imagelets from the image representation for each character in the at least one OCR cleaned seed.

13. The method of claim 12, wherein the image representation is a bitmap.

14. The method of claim 13, wherein the bitmap is grayscale or color.

15. The method of claim 1, wherein statistical measurements and/or cluster analysis is used to select similar imagelets.

16. The method of claim 15, wherein the statistical measurements and/or the cluster analysis comprises:

at least one Bayes classifier;

at least one neural networks classifier;

at least one support vector classifier;

or

any combination thereof.

17. A system for increasing the accuracy of optical character recognition (OCR) for at least one item, comprising:

at least one processor, wherein the at least one processor is configured to perform;

obtaining OCR results of OCR scanning from at least one OCR module;

creating at least one OCR seed using at least a portion of the OCR results, the at least one OCR seed comprising a plurality of imagelets corresponding to each character identified in the at least a portion of the OCR results, wherein the at least one OCR seed is cleaned by selecting imagelets similar to one another for each character identified in the at least a portion of the OCR results;

creating at least one OCR learn set using at least a portion of the OCR seed;

comparing the at least one OCR learn set to each imagelet to create at least one mismatch distribution of the at least one OCR learn set compared to each imagelet, the at least one mismatch distribution comprising at least one confidence rating including a confidence score for the imagelet compared to at least one possible character; and

applying the OCR learn set and the at least one mismatch distribution to the at least one item to obtain additional OCR results such that only possible characters having a confidence score higher than a threshold are considered when applying the at least one mismatch distribution to obtain the additional OCR results.

18. The system of claim 17, wherein image statistics are deducted based on the at least one OCR learn set.

19. The system of claim 17, wherein the at least one item comprises:

at least one page;

at least one document;

at least one part of at least one document;

at least one zone;

at least one paragraph;

at least one line;

at least one word;

or

at least one character;

or

any combination thereof.

20. The system of claim 17, wherein the at least one OCR learn set is used on different items.

21. The system of claim 17, wherein different OCR modules and/or different parameter sets and/or different preprocessing steps are used.

22. The system of claim 17, wherein the additional OCR results take into account:

font size;

font type;

or

font properties;

or

any combination thereof.

23. The system of claim 17, wherein whitespace characteristics are taken into account in creating the at least one OCR learn set.

24. The system of claim 17, wherein the method is performed iteratively to allow for a refinement of the method.

25. The system of claim 17, wherein the at least one OCR learn set comprises stored imagelets and/or new imagelets.

26. The system of claim 17, wherein stored imagelets are used to increase the at least one OCR learn set.

27. The system of claim 17, wherein the at least one OCR learn set and related statistics are used to derive at least one image filtering technique.

28. The system of claim 17, wherein the OCR learn set is created by obtaining information related to the image representation for each character and the variability of the similar imagelets from the image representation for each character in the at least one OCR cleaned seed.

29. The method of claim 28, wherein the image representation is a bitmap.

30. The method of claim 29, wherein the bitmap is grayscale or color.

31. The system of claim 17, wherein statistical measurements and/or cluster analysis is used to select similar imagelets.

32. The system of claim 31, wherein the statistical measurements and/or the cluster analysis comprises:

at least one Bayes classifier;

at least one neural networks classifier;

at least one support vector classifier;

or

any combination thereof.