Method and system for recognizing a candidate character in a captured image

• US 7,590,275 B2
• Filed: 08/26/2004
• Issued: 09/15/2009
• Est. Priority Date: 08/26/2004
• Status: Expired due to Fees

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: Seiko Epson Corporation (Seiko Group)
• Inventors: Clarke, Ian, Zhou, Hui, Achong, Jeffrey Matthew
• Primary Examiner(s): Werner; Brian P
• Assistant Examiner(s): FUJITA, KATRINA R
• Application Number:
  US10/926,658
• Publication Number:
  US 20060045322A1
• Time in Patent Office: 1,846 Days
• Field of Search: 382/137, 705/45
• US Class Current: 382/137
• CPC Class Codes: G06Q 20/042 characterized in that the payment protocol involves at least one cheque
• CPC Class Codes: G06V 30/2253 Recognition of characters printed with magnetic ink G06V30/2247 takes precedence

Abstract

A system and method of recognizing characters in a captured image, includes selecting a portion of the image encompassing a string of characters in the captured image to be recognized. The selected portion is examined to identify segments that potentially include the characters of the string. During the examining multiple tests on pixels in the selected portion are performed. The identified segments are then examined to recognize the string of characters.

Claims

1. A method of identifying candidate characters of a character string in a captured image, the method comprising using a processor to perform the steps of:

identifying special characters within the captured image of said character string, including;

performing a Fast Fourier Transform (FFT) on a portion of the captured image;

performing an inverse FFT yielding a cross-correlation image representing a cross-correlation between the captured image and a character template image corresponding to the special characters;

box-filtering the cross-correlation image to smooth the cross-correlation image;

subtracting the box-filtered image from the cross-correlation image yielding a modified cross-correlation image; and

based at least in part on the modified cross-correlation image, identifying the special characters;

determining a bounding box generally encompassing said character string based on the identified special characters;

determining segments of said bounding box containing candidate characters;

comparing each determined segment of said bounding box with at least one character template image, each character template image including features associated with a respective character to be recognized; and

identifying a candidate character when said segment includes image information corresponding to one of a set of character template images.

2. The method of claim 1, wherein each determined segment is compared with a plurality of character template images and wherein said features in each character template image include contours of the respective character.

3. The method of claim 2, wherein said contours include pixels forming the outline of said respective character and background pixels adjacent to the outline pixels.

4. The method of claim 2, wherein each character template image has an associated coefficient map assigning coefficients to each pixel in said character template image, coefficients corresponding to pixels of said contour being weighted differently than other pixels in said character template image, during said comparing, pixels of each determined segment being compared with corresponding pixels of each character template image and a correlation result being generated using said coefficients, said correlation result representing the degree of likelihood said segment includes a candidate character represented by a character template image, each correlation result satisfying a threshold being used to identify the segment as potentially including a candidate character represented by the character template image associated with the coefficients used to generate that correlation result.

5. The method of claim 4, wherein the pixels of said contour are weighted higher than other pixels.

6. The method of claim 5, wherein each correlation result above said threshold is used during the identifying.

7. The method of claim 6, wherein each correlation result is generated by summing coefficients assigned to pixels in the character template image that match corresponding pixels in the segment.

8. The method of claim 7, wherein coefficients assigned to pixels in the character template image that do not match corresponding pixels in the segment are subtracted from the summed coefficients.

9. The method of claim 2, wherein said character is an E13B character.

10. The method of claim 9 wherein said captured image is an image of a cheque, wherein said each segment contains an E13B character and wherein said character string is an account identification E13B character string on said cheque.

11. The method of claim 10, wherein said contours include pixels forming the outline of said respective character and background pixels adjacent to the outline pixels.

12. The method of claim 11, wherein each character template image includes a coefficient map assigning coefficients to each pixel in said character template image, coefficients corresponding to pixels of said contour being weighted differently than other pixels in said character template image, during said comparing, pixels of each determined segment being compared with corresponding pixels of each character template image and a correlation result being generated using said coefficients representing the degree of likelihood said segment includes a candidate character represented by a character template image, each correlation result satisfying a threshold being used to identify the segment as potentially including the character represented by the character template image associated with the coefficients used to generate that correlation result.

13. The method of claim 12 wherein pixels of said contours are weighted higher than other pixels and wherein each correlation result above the threshold is used during the identifying.

14. The method of claim 13 wherein each correlation result is generated by summing coefficients assigned to pixels in the character template image that match corresponding pixels in the segment and by subtracting coefficients assigned to pixels in the character template image that do not match corresponding pixels in the segment.

15. The method of claim 10 wherein said comparing and identifying is performed for each segment of said cheque that encompasses a candidate E13B character in said account identification E13B character string.

16. The method of claim 15 further comprising examining each identified segment to recognize the candidate E13B character encompassed thereby.

17. The method of claim 16 wherein said examining is performed using a trained neural network.

18. The method of claim 2 further comprising the step of examining each identified segment to recognize said candidate character.

19. The method of claim 18 wherein said examining is performed using a trained neural network.

20. The method of claim 19, wherein each character template image includes a coefficient map assigning coefficients to each pixel in said character template image, coefficients corresponding to pixels of said contour being weighted differently than other pixels in said character template image, during said comparing, pixels of each determined segment being compared with corresponding pixels of each character template image and a correlation result being generated using said coefficients representing the degree of likelihood said segment includes a candidate character represented by a character template image, each correlation result satisfying a threshold being used to identify the segment as potentially including the character represented by the character template image associated with the coefficients used to generate that correlation result.

21. The method of claim 20 wherein pixels of said contours are weighted higher than other pixels and wherein each correlation result above the threshold is used during the identifying.

22. The method of claim 21 wherein each correlation result is generated by summing coefficients assigned to pixels in the character template image that match corresponding pixels in the segment and by subtracting coefficients assigned to pixels in the character template image that do not match corresponding pixels in the segment.

23. The method of claim 1, wherein the location of the special characters is used to determine the bounding box generally encompassing said character string by estimating the location of the limits of the character string using the locations of the special characters and using a known size of characters of the character string.

24. The method of claim 1, wherein identifying special characters of said character string further includes performing a FFT on the character template image corresponding to the special characters.

25. The method of claim 24, wherein identifying special characters of said character string further includes:

locating two peak values in an image resulting from the subtraction of the box-filtered image from the inverse FFT of the cross-correlation image;

performing a spatial analysis of the two peak values; and

confirming that the peaks match the special characters based on the spatial analysis of the two peak values.

26. A method of recognizing characters of a character string in an image, comprising using a processor to perform the steps of:

identifying designated characters of said character string comprising;

performing a Fast Fourier Transform (FFT) on a portion of the captured image;

performing an inverse FFT yielding a cross-correlation image representing a cross-correlation between the captured image and a character template image corresponding to the designated characters;

box-filtering the cross-correlation image to smooth the cross-correlation image;

subtracting the box-filtered image from the cross-correlation image yielding a modified cross-correlation image; and

based at least in part on the modified cross-correlation image, identifying the designated characters;

using the location of the designated characters to determine a bounding box generally encompassing said character string;

determining segments of said bounding box containing candidate characters;

performing character recognition using a neural network by comparing each determined segment of said image containing a candidate character to be recognized with a plurality of character template images representing recognizable characters;

when the candidate character of a determined segment corresponds to one of said character template images, recognizing the candidate character as that represented by said one character template image; and

when the candidate character of a determined segment corresponds to a plurality of said character template images, using a correlation between the determined segment and each character template image of the plurality to recognize the candidate character as that represented by one character template image of said plurality.

27. The method of claim 26 wherein the correlation representing the character template image of the plurality that is deemed most likely to represent the candidate character encompassed by the determined segment is used.

28. The method of claim 27 wherein the candidate character is not recognized when the correlations are below a threshold strength.

29. The method of claim 28 wherein the characters are E13B characters.