Robust industrial optical character recognition

• US 9,396,404 B2
• Filed: 08/04/2014
• Issued: 07/19/2016
• Est. Priority Date: 08/04/2014
• Status: Active Grant

• Current Assignee: Datalogic Adc, Inc. (Datalogic S.p.A.)
• Original Assignee: Datalogic Adc, Inc. (Datalogic S.p.A.)
• Inventors: Goncalves, Luis
• Primary Examiner(s): BHATNAGAR, ANAND P
• Application Number:
  US14/450,394
• Publication Number:
  US 20160034773A1
• Time in Patent Office: 715 Days
• US Class Current: 1/1
• CPC Class Codes: G06V 10/443 by matching or filtering
• CPC Class Codes: G06V 10/50 by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
• CPC Class Codes: G06V 30/224 of printed characters having additional code marks or containing code marks

Abstract

Systems, methods, and articles to provide robust optical character recognition (OCR) for use in industrial environments. One or more implementations include utilizing Histogram of Oriented Gradients (HOG) features with a sliding window approach as a robust and computationally efficient method of OCR. The implementations are relatively simple to use because there are relatively few parameters to adjust, which allows a non-expert user to setup or modify the system and achieve desirable performance. One reason this is possible is because the implementations described herein do not require character segmentation, which can be difficult to optimize.

Claims

1. A method of operation in an image processor system for recognizing one or more characters in an image of a target object, the method comprising:

receiving, by a processor, image data representing an image of the target object;

generating histogram of oriented gradients (HOG) features for the received image data;

accessing, by the processor, a plurality of HOG character models stored in a nontransitory processor-readable storage medium, each of the plurality of HOG character models associated with a respective known character;

for each of a plurality of image locations of the image, comparing each of the plurality of HOG character models to the generated HOG features of the image at the image location;

determining a quality measure based at least in part on the comparing of each of the plurality of HOG character models to the generated HOG features of the image; and

logically associating in a nontransitory processor-readable storage medium at least one known character with at least one image location based on the comparison to generate at least one candidate character location.

2. The method of claim 1, further comprising:

combining two or more candidate character locations to form at least one word including two or more of the known characters; and

storing the at least one word in a nontransitory processor-readable storage medium.

3. The method of claim 2 wherein combining two or more candidate character locations to form at least one word including two or more of the known characters comprises evaluating two or more candidate character locations for at least one of character spacing or character orientation.

4. The method of claim 2, further comprising:

evaluating a spatial distance between two candidate character locations; and

determining that the at least one word represents at least two words based at least in part on evaluating the spatial distance.

5. The method of claim 2, further comprising:

forming at least two words each including two or more of the known characters;

evaluating a spatial distance between the at least two words;

determining that the at least two words represent a single word based at least in part on evaluating the spatial distance; and

combining the two words to form the single word.

6. The method of claim 5, further comprising:

detecting at least one character spaced between two of the at least two formed words.

7. The method of claim 1, further comprising:

dividing the image data into a first portion in which characters are found and a second portion in which no characters are found, wherein generating histogram of oriented gradients (HOG) features for the received image data comprises generating histogram of oriented gradients (HOG) features for only the first portion of the received image data.

8. The method of claim 1, further comprising:

determining that at least one portion of the image data is of uniform color, wherein generating histogram of oriented gradients (HOG) features for the received image data comprises discarding the at least one portion determined to be of uniform color.

9. The method of claim 1 wherein logically associating in a nontransitory processor-readable storage medium at least one known character with at least one image location based on the comparison to generate at least one candidate character location comprises logically associating in a nontransitory processor-readable storage medium two or more known characters with one image location based on the comparison to generate at least one candidate character location associated with the two or more known characters.

10. A method of operation in an image processor system for recognizing one or more characters in an image of a target object, the method comprising:

receiving, by a processor, image data representing an image of the target object;

generating histogram of oriented gradients (HOG) features for the received image data;

accessing, by the processor, a plurality of HOG character models stored in a nontransitory processor-readable storage medium, each of the plurality of HOG character models associated with a respective known character;

for each of a plurality of image locations of the image, comparing each of the plurality of HOG character models to the generated HOG features of the image at the image location; and

logically associating in a nontransitory processor-readable storage medium at least one known character with at least one image location based on the comparison to generate at least one candidate character location,

wherein each of the plurality of HOG character models comprises a plurality of bins each having a spatial dimension, and generating histogram of oriented gradients (HOG) features for the received image data comprises dividing the received image data into a plurality of bins each having a spatial dimension that is equal to the spatial dimension of the bins associated with the HOG character models.

11. The method of claim 10 wherein dividing the received image data into a plurality of bins comprises dividing the received image data into a plurality of bins each having a width of six pixels, seven pixels, or eight pixels, and a height of six pixels, seven pixels, or eight pixels.

12. A method of operation in an image processor system for recognizing one or more characters in an image of a target object, the method comprising:

receiving, by a processor, image data representing an image of the target object;

generating histogram of oriented gradients (HOG) features for the received image data;

accessing, by the processor, a plurality of HOG character models stored in a nontransitory processor-readable storage medium, each of the plurality of HOG character models associated with a respective known character;

for each of a plurality of image locations of the image, comparing each of the plurality of HOG character models to the generated HOG features of the image at the image location, wherein comparing each of the plurality of HOG character models to the generated HOG features of the image at the image location comprises, for each of a plurality of image locations of the image, correlating each of the plurality of HOG character models with the HOG features at the image location to generate a correlation value for each of the HOG character models; and

logically associating in a nontransitory processor-readable storage medium at least one known character with at least one image location based on the comparison to generate at least one candidate character location;

for each of the plurality of image locations,

selecting one of the plurality of HOG character models having the highest correlation value with the image location;

logically associating the highest correlation value of the selected HOG character model with the image location in a nontransitory processor-readable storage medium; and

logically associating the known character associated with the selected HOG character model with the image location in a nontransitory processor-readable storage medium.

13. The method of claim 12, further comprising:

for each of the plurality of image locations,

comparing the highest correlation value of the selected HOG character model at the image location to the highest correlation value for at least one neighboring image location.

14. The method of claim 12, further comprising:

for each of the plurality of image locations,

comparing the highest correlation value of the selected HOG character model at the image location to the highest correlation value for each image location that is a 1-pixel neighbor of the image location.

15. The method of claim 14, further comprising:

for each of the plurality of image locations,

selecting the image location as a candidate character location if the highest correlation value of the selected HOG character model at the image location is equal to or greater than the highest correlation value for each image location that is a 1-pixel neighbor of the image location.

16. The method of claim 15, further comprising:

for each of the image locations selected as a candidate character location,

comparing the highest correlation value of the selected HOG character model at the image location to at least one correlation value for each image location that is a 2-pixel neighbor of the image location.

17. The method of claim 15, further comprising:

for each of the image locations selected as a candidate character location,

comparing the highest correlation value of the selected HOG character model at the image location with at least one of;

a mean of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location, a median of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location, or a maximum of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location.

18. The method of claim 17, further comprising:

for each of the image locations selected as a candidate character location,

retaining the image location as a candidate character location if the highest correlation value of the selected HOG character model at the image location is greater than or equal to twice at least one of;

the mean of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location, the median of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location, or the maximum of the correlation values of the selected HOG character model at the image locations that are 2-pixel neighbors of the image location.

19. An image processor system, comprising:

at least one processor; and

at least one nontransitory processor-readable medium, communicatively coupled to the at least one processor and which stores histogram of oriented gradient (HOG) character models of a set of known characters and at least one of processor-executable instructions or data, each of the plurality of HOG character models comprises a plurality of bins each having a spatial dimension, wherein in use the at least one processor;

receives image data that represents an image of a target object;

divides the received image data into a plurality of bins each having a spatial dimension that is equal to the spatial dimension of the bins associated with the HOG character models;

generates histogram of oriented gradients (HOG) features for the received image data;

accesses the plurality of HOG character models stored in the at least one nontransitory processor-readable storage medium, each of the plurality of HOG character models associated with a respective known character;

for each of a plurality of image locations of the image, compares each of the plurality of HOG character models to the generated HOG features of the image at the image location; and

logically associates in the at least one nontransitory processor-readable storage medium at least one known character with at least one image location based on the comparison to generate at least one candidate character location.

20. The image processor system of claim 19 wherein the at least one processor:

divides the received image data into a plurality of bins each having a width of six pixels, seven pixels, or eight pixels, and a height of six pixels, seven pixels, or eight pixels.

21. The image processor system of claim 19 wherein, for each of a plurality of image locations of the image, the at least one processor:

correlates each of the plurality of HOG character models with the HOG features at the image location to generate a correlation value for each of the HOG character models.

22. The image processor system of claim 21 wherein, for each of a plurality of image locations of the image, the at least one processor:

selects one of the plurality of HOG character models having the highest correlation value with the image location;

logically associates the highest correlation value of the selected HOG character model with the image location in a nontransitory processor-readable storage medium; and

logically associates the known character associated with the selected HOG character model with the image location in a nontransitory processor-readable storage medium.

23. The image processor system of claim 22 wherein, for each of a plurality of image locations of the image, the at least one processor:

compares the highest correlation value of the selected HOG character model at the image location to at least one correlation value for at least one neighboring image location.

24. The image processor system of claim 22 wherein the at least one processor:

compares the highest correlation value of the selected HOG character model at the image location to at least one correlation value for each image location that is a 1-pixel neighbor of the image location; and

selects the image location as a candidate character location if the highest correlation value of the selected HOG character model at the image location is equal to or greater than the at least one correlation value for each image location that is a 1-pixel neighbor of the image location.

25. The image processor system of claim 24 wherein the at least one processor:

compares the highest correlation value of the selected HOG character model at the image location to at least one correlation value for each image location that is a 2-pixel neighbor of the image location.

26. The image processor system of claim 19 wherein the at least one processor:

combines two or more candidate character locations to form at least one word that includes two or more of the known characters; and

stores the at least one word in the at least one nontransitory processor-readable storage medium.

27. The image processor system of claim 26 wherein the at least one processor:

evaluates two or more candidate character locations for at least one of character spacing or character orientation.

28. The image processor system of claim 26 wherein the at least one processor:

evaluates a spatial distance between two candidate character locations; and

determines the at least one word represents at least two words based at least in part on evaluation of the spatial distance.

29. The image processor system of claim 26 wherein the at least one processor:

forms at least two words each including two or more of the known characters;

evaluates a spatial distance between the at least two words;

determine the at least two words represent a single word based at least in part on evaluation of the spatial distance; and

combines the two words to form the single word.

30. The image processor system of claim 29 wherein the at least one processor:

detects at least one character spaced between two of the at least two formed words.

31. The image processor system of claim 19 wherein the at least one processor:

determines a quality measure based at least in part on the comparison of each of the plurality of HOG character models to the generated HOG features of the image.

32. The image processor system of claim 19 wherein the at least one processor:

divides the image data into a first portion in which characters are found and a second portion in which no characters are found; and

generates histogram of oriented gradients (HOG) features for only the first portion of the received image data.

33. The image processor system of claim 19 wherein the at least one processor:

determines that at least one portion of the image data is of uniform color; and

generates histogram of oriented gradients (HOG) features for the received image data, wherein the at least one processor discards the at least one portion determined to be of uniform color.

34. A method of operation in an image processor system for recognizing one or more words in an image of a target object, the method comprising:

receiving, by at least one processor, training image data representing a plurality of training characters;

generating a plurality of histogram of oriented gradients (HOG) character models, each of the plurality of HOG character models associated with one of the training characters;

receiving, by at least one processor, image data representing an image of the target object;

generating HOG features for the received image data;

determining correlations between each of the HOG character models and the HOG features for the received image data;

determining a plurality of candidate character locations based on the determined correlations; and

assembling at least one word using at least some of the determined plurality of candidate character locations.

35. The method of claim 34 wherein determining a plurality of candidate character locations based on the determined correlations comprises:

for each image location of a plurality of image locations,

determining a maximum correlation value at the image location, the maximum correlation value being a correlation value of one or more of the plurality of HOG character models having the highest correlation value with the HOG features at the image location;

determining whether the image location is a local peak by determining whether the maximum correlation value at the image location is greater than or equal to maximum correlation values at image locations that are 1-pixel neighbors of the image location;

selecting the image location as a prospective candidate character location upon determining that the image location is a local peak;

for each image location selected as a prospective candidate character location,

determining whether the image location is a strong local peak by comparing the maximum correlation value at the image location with correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location; and

selecting the image location as a candidate character location upon determining that the image location is a strong local peak.

36. The method of claim 35 wherein determining whether the image location is a strong local peak by comparing the maximum correlation value at the image location with correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location comprises:

comparing the maximum correlation value at the image location with at least one of;

a mean of the correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location;

a median of the correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location;

or

a maximum of the correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location.

37. The method of claim 34 wherein determining a plurality of candidate character locations based on the determined correlations comprises:

for each image location of a plurality of image locations,

for each HOG character model of the generated plurality of HOG character models,

determining whether the image location is a local peak by determining whether the correlation value between the HOG character model and the HOG features at the image location is greater than or equal to the correlation values between the HOG character model and the HOG features at image locations that are 1-pixel neighbors of the image location;

selecting the image location as a prospective candidate character location upon determining that the image location is a local peak;

for each image location selected as a prospective candidate character location,

determining whether the image location is a strong local peak by comparing the correlation value between the HOG character model and the HOG features at the at the image location with correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location; and

selecting the image location as a candidate character location upon determining that the image location is a strong local peak.

38. The method of claim 34 wherein determining a plurality of candidate character locations based on the determined correlations comprises:

for each image location of a plurality of image locations,

for each HOG character model of the generated HOG character models,

determining whether the image location is a strong local peak by comparing the correlation value between the HOG character model and the HOG features at the at the image location with correlation values between the HOG character model and the HOG features at image locations that are 2-pixel neighbors of the image location; and

selecting the image location as a candidate character location upon determining that the image location is a strong local peak.