Method and apparatus for implementing a weighted voting scheme for multiple optical character recognition systems

• US 5,519,786 A
• Filed: 08/09/1994
• Issued: 05/21/1996
• Est. Priority Date: 08/09/1994
• Status: Expired due to Term

• Current Assignee: Northrop Grumman Systems Corporation (Northrop Grumman Corporation)
• Original Assignee: TRW Limited (Zeppelin - Stiftung der Stadt Friedrichshafen)
• Inventors: Smith, M'Lissa L., Courtney, William F.
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): Johns, Andrew W.
• Application Number:
  US08/287,999
• Time in Patent Office: 651 Days
• Field of Search: 382/57, 382/62, 382/155, 382/159, 382/161, 382/309, 382/318, 382/304, 382/310
• US Class Current: 382/159
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data

Abstract

A method and apparatus for implementing a weighted voting scheme for reading and accurately recognizing characters in a scanned image. A plurality of optical character recognition processors scan the image and read the same image characters. Each OCR processor outputs a reported character corresponding to each character read. For a particular character read, the characters reported by each OCR processor are grouped into a set of character candidates. For each character candidate, a weight is generated in accordance with a confusion matrix which stores probabilities of a particular OCR to identify characters accurately. The weights are then compared to determine which character candidate to output.

Claims

1. An image processing system for recognizing characters in a scanned image, said system comprising:

a plurality of parallel optical character recognition (OCR) processors, each OCR processor having individual character recognition capabilities for interpreting and reporting given characters in the scanned image;

confusion matrix storage means for storing a plurality of matrices each corresponding to one of the plurality of OCR processors, each of said matrices being defined by a plurality of detection values, each detection value being a function of the probability that a particular character in the image will cause the OCR processor associated with the matrix to report a particular candidate;

said parallel OCR processors each reporting a potential character candidate for each character read from the image, whereby the plurality of OCR processors generate a set of potential character candidates for each character that is read; and

error processing means for;

a. generating character weights for the potential character candidates reported by selected OCR processors in accordance with a plurality of detection values from at least one confusion matrix based upon the reported potential character candidates; and

b. choosing the potential character candidate having the highest weight as the character recognition output of the system.

2. The apparatus as defined in claim 1 wherein said confusion matrix further comprises:

a first index including all characters that can possibly be read by said OCR processor;

a second index including all characters that can possibly be reported by said OCR processor;

whereby a particular reported character of said first index and a particular read character of said second index reference a position in said matrix in which is stored a detection value which defines the probability that said particular reported character is said particular character read.

3. The apparatus as defined in claim 2 further comprising:

a preprocessor, for generating a confusion matrix for each OCR processor said processor including;

a. output text storage means for storing characters reported by said OCR processor, where said OCR processor outputs said reported characters after scanning an image having a test set of characters; and

b. means for generating the detection values of said confusion matrix by comparing characters reported by said OCR processor with ground truth text characters, said ground truth text characters representing an exact reporting of said scanned image.

4. The apparatus as defined in claim 3 wherein the means for generating said confusion matrix increments the cell defined by said reported character and said ground truth character, said ground truth character corresponding to said first index and said reported character corresponding to said second index.

5. The apparatus as defined in claim 4 wherein said means for generating said confusion matrix normalizes the number of occurrences of said reported characters for each of said reported characters.

6. The apparatus as defined in claim 2 wherein at least one of said plurality of OCR processors outputs with each reported character a confidence level defining the confidence with which said reported character is accurately reported.

7. The apparatus as defined in claim 6 wherein said confusion matrix further comprises a third index including all possible confidence levels for said OCR processor, whereby a particular read character of said first index, a particular reported character of said second index, and a particular confidence level of said third index reference a position in said matrix in which is stored a detection value which defines the probability that said particular reported character is said particular character read at a particular confidence level.

8. The apparatus as defined in claim 1 wherein said error processing means applies a plurality of said detection values of said confusion matrix to said potential character candidate to weight potential character candidates, said plurality of detection values chosen in accordance with the potential candidate.

9. The apparatus as defined in claim 8 wherein said error processing means multiplies each of said plurality of detection values to yield said weight for said potential character candidate.

10. The apparatus as defined in claim 1, wherein said optical character recognition processors are arranged in a parallel configuration.

11. A method for recognizing characters in a scanned image, said method comprising:

providing a plurality of optical character recognition (OCR) processors, each OCR processor having individual character recognition capabilities for interpreting and reporting given characters in the scanned image;

generating a plurality of confusion matrices each corresponding to one of the plurality of OCR processors, each of said matrices being defined by a plurality of detection values, each detection value being a function of the probability that a particular character in the image will cause the OCR processor associated with the matrix to report a particular candidate;

reporting a potential character candidate for each character read from the image for each OCR processor, whereby the plurality of OCR processors generate a set of potential character candidates for each character that is read;

generating character weights for the potential character candidates reported by selected OCR processors in accordance with a plurality of detection values from at least one confusion matrix based upon the reported potential character candidates; and

choosing the potential character candidate having the highest weight as the character recognition output of the system.

12. The method as defined in claim 11 wherein said step of generating said confusion matrix further comprises:

providing a first index including all characters that can possibly be read by said OCR processor;

providing a second index including all characters that can possibly be reported by said OCR processor;

whereby a particular read character of said first index and a particular reported character of said second index reference a position in said matrix in which is stored a detection value which defines the probability that said particular reported character is said particular character read.

13. The method as defined in claim 12 wherein said step of generating said confusion matrix for each OCR processor further comprises:

storing characters reported by said OCR processor as output text, where said OCR processor outputs said reported characters after scanning an image having a test set of characters; and

generating the detection values of said confusion matrix by comparing characters reported by said OCR processor with ground truth text characters, said ground truth text characters representing an exact reporting of said scanned image.

14. The method as defined in claim 13 wherein the step generating said confusion matrix further comprises incrementing the cell defined by said reported character and said ground truth character, said ground truth character corresponding to said first index and said reported character corresponding to said second index.

15. The method as defined in claim 14 wherein said means for generating said confusion matrix further comprises normalizing the number of occurrences of said reported characters for each of said reported characters.

16. The method as defined in claim 12 wherein at least one of said plurality of OCR processors outputs with each reported character a confidence level defining the confidence with which said reported character is accurately reported.

17. The method as defined in claim 16 wherein the step of generating said confusion matrix further comprises providing a third index including all possible confidence levels for said OCR processor, whereby a particular read character of said first index, a particular reported character of said second index, and a particular confidence level of said third index reference a position in said matrix in which is stored a detection value which defines the probability that said particular reported character is said particular character read at a particular confidence level.

18. The method as defined in claim 11 wherein said step of determining said character weight further comprises applying a plurality of said detection values of said confusion matrix to said potential character candidate to weight potential character candidates, said plurality of detection values chosen in accordance with the potential candidate.

19. The method as defined in claim 18 wherein said step of determining said character weight further comprises multiplying each of said plurality of detection values to yield said weight for said potential character candidate.

20. An image processing system for recognizing characters in a scanned image, said system comprising:

a device for receiving data input thereto and outputting an electronic image in a predetermined format;

processing means for receiving said electronic image, said processing means including;

a plurality of optical character recognition (OCR) processors, each OCR processor having individual character recognition capabilities for interpreting and reporting given characters in the scanned image;

confusion matrix storage means for storing a plurality of matrices each corresponding to one of the plurality of OCR processors, each of said matrices being defined by a plurality of detection values, each detection value being a function of the probability that a particular character in the image will cause the OCR associated with the matrix to report a particular candidate;

said OCR processors each normally reporting a potential character candidate for each character read from the image, whereby the plurality of OCR processors generate a set of potential character candidates for each character that is read; and

error processing means for;

a. generating character weights for the potential character candidates reported by selected OCR processors in accordance with a plurality of detection values from at least one confusion matrix based upon the reported potential character candidates; and

b. choosing the potential character candidate having the highest weight as the character recognition output of the system.

21. The apparatus as defined in claim 20 wherein said device for receiving data comprises a telefacsimile machine.

22. The apparatus as defined in claim 20 wherein said device for receiving data comprises a document scanner.

23. The apparatus as defined in claim 20 wherein said processor is a computer.

24. The apparatus as defined in claim 20, wherein said OCR processors are arranged in a parallel configuration.