Classifier combination for optical character recognition systems utilizing normalized weights and samples of characters

• US 8,548,259 B2
• Filed: 10/24/2012
• Issued: 10/01/2013
• 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:
  US13/659,289
• Publication Number:
  US 20130044943A1
• Time in Patent Office: 342 Days
• Field of Search: None
• US Class Current: 382/224
• 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

Techniques and methods are disclosed herein for combining and weighting of values from and associated with classifiers. Classifiers are used to recognize characters as part of an optical character recognition (OCR) system. Various methods of normalization facilitate combining of results of classifiers. For example, weight values may be entered into a weight table having two columns, one that includes weights from comparing patterns with images of correct characters, the other column includes weights from comparing patterns with images of incorrect characters.

Claims

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

comparing the character pattern to images of the same character, wherein the images are associated with a database of character images;

comparing the character pattern to images of different characters associated with the database of character images;

calculating a weight value for a first classifier for an image sample of the same character based on comparing the character pattern to images of the same character;

calculating a weight value for the first classifier for an image sample of a different character based on comparing the character pattern to images of the different characters;

calculating a weight value for a second classifier for an image sample of the same character based on comparing the character pattern to images of the same character;

calculating a weight value for the second classifier for an image sample of a different character based on comparing the character pattern to images of the different characters;

for each weight value for the first classifier, calculating a corresponding normalized weight for the first classifier, based on a correlation between same character and different character samples of characters which received a same weight value by comparing the said calculated character pattern weight values;

for each weight value for the second classifier, calculating a corresponding normalized weight for the second classifier, based on a correlation between same character and different character samples of characters which received a same weight value by comparing the said calculated character pattern weight value;

calculating corresponding normalized weight values; and

combining the normalized weight values for the first classifier with those normalized weight values associated with the second classifier.

2. The method of claim 1, wherein the method further comprises:

repeating for other image samples said calculating the weight value for the image sample of the same character based on comparing the character pattern to images of the same character; and

repeating for other image samples said calculating the weight value for the image sample of the different characters based on comparing the character pattern to images of the different characters.

3. The method of claim 1, wherein calculating a corresponding normalized weight for each weight value includes creating a weight normalization data structure.

4. The method of claim 1, wherein the method is repeated for each character pattern used by the OCR system.

5. The method of claim 1, wherein the method is repeated for each type of classifier used by the OCR system.

6. The method of claim 1, wherein calculating the weight value for each image sample includes applying a modifier based on a scale associated with a classifier.

7. The method of claim 1, wherein the first classifier and the second classifier are a set of nearest neighbor classifiers.

8. The method of claim 1, wherein calculating weight values comprises:

creating a data structure with a first portion assigned to weight values that correspond to images of the same character and another portion assigned to weight values that correspond to images of the different characters.

9. The method of claim 8, wherein the method further comprises:

for each weight value, determining a number of image samples with a weight equal to or better than a corresponding weight value from the first portion and from the other portion; and

wherein calculating corresponding normalized weight is based on a function, wherein the function includes use of a formula that is substantially proportional to a ratio of the number of samples of different characters to a sum of samples of different and same characters.

10. An electronic device for training an optical character recognition (OCR) classifier, the device comprising:

a processor;

a memory in electronic communication with the processor, the memory configured with instructions that cause the electronic device to;

compare a character pattern to images of the same character, wherein the images of the same character are part of a collection of character images;

compare the character pattern to images of different characters, wherein the images of the different characters are part of a collection of character images;

calculate a weight value for an image sample of the same character based on comparing the character pattern to images of the same character;

calculate a weight value for an image sample of a different character based on comparing the character pattern to images of the different character;

for each weight value, calculate a corresponding normalized weight, based on a correlation between same character samples and different character samples for the character samples that received a same weight value by comparing the said calculated character pattern weight values; and

store in said memory said normalized weights for use by the OCR classifier.

11. The electronic device of claim 10, wherein the instructions further cause the electronic device to:

repeat for other image samples said calculating the weight value for the image sample of the same character based on comparing the character pattern to images of the same character; and

repeat for other image samples said calculating the weight value for the image sample of the different character based on comparing the character pattern to images of the different characters.

12. The electronic device of claim 10, wherein calculating weight values includes:

creating a data structure with a first portion assigned to weight values that correspond to images of the same character and another portion assigned to weight values that correspond to images of the different characters.

13. The electronic device of claim 10, wherein calculating the weight value for each image sample includes applying a modifier based on a scale associated with the OCR classifier.

14. The electronic device of claim 10, wherein calculating the weight values for the image sample is associated with a first classifier, and wherein the instructions further cause the electronic device to:

calculate weight values for a second classifier;

calculate corresponding normalized weight values; and

combine the normalized weight values for the first classifier with those normalized weight values associated with the second classifier, wherein the first classifier and the second classifier are a set of nearest neighbor classifiers.

15. One or more physical computer-accessible media encoded with instructions for performing a method, the method comprising:

comparing a character pattern to images of the same character, wherein the images of the same character are part of a collection of character images;

comparing the character pattern to images of different characters, wherein the images of the different characters are part of a collection of character images;

calculating a weight value for an image sample of the same character based on comparing the character pattern to images of the same character;

calculating a weight value for an image sample of a different character based on comparing the character pattern to images of the different characters;

for each weight value, calculating a corresponding normalized weight, based on a correlation between same character samples and different character samples for the character samples that received a same weight value by comparing the said calculated character pattern weight values; and

storing in a computer-accessible memory said normalized weights for use by the OCR classifier.

16. The one or more physical computer-accessible media of claim 15, wherein the method further comprises:

repeating for other image samples said calculating the weight value for the image sample of the same character based on comparing the character pattern to images of the same character; and

repeating for other image samples said calculating the weight value for the image sample of the different character based on comparing the character pattern to images of the different characters.

17. The one or more physical computer-accessible media of claim 15, wherein calculating weight values includes:

creating a data structure with a first portion assigned to weight values that correspond to images of the same character and another portion assigned to weight values that correspond to images of the different characters.

18. The one or more physical computer-accessible media of claim 15, wherein calculating the weight value for each image sample includes applying a modifier based on a scale associated with the OCR classifier.

19. The one or more physical computer-accessible media of claim 15, wherein calculating the weight values for the image sample is associated with a first classifier, and wherein the method further comprises:

calculating weight values for a second classifier;

calculating corresponding normalized weight values; and

combining the normalized weight values for the first classifier with those normalized weight values associated with the second classifier, wherein the first classifier and the second classifier are a set of nearest neighbor classifiers.