ASYMMETRIC SCORE NORMALIZATION FOR HANDWRITTEN WORD SPOTTING SYSTEM

• US 20090180695A1
• Filed: 01/15/2008
• Published: 07/16/2009
• Est. Priority Date: 01/15/2008
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Perronnin, Florent, Rodriguez Serrano, Jose A.
• Granted Patent: US 8,027,540 B2
• Time in Patent Office: Days
• US Class Current: 382/186
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06V 30/1918 Fusion techniques, i.e. combining data from various sources, e.g. sensor fusion
• CPC Class Codes: G06V 30/2276 with probabilistic networks, e.g. hidden Markov models

Abstract

A method begins by receiving an image of a handwritten item. The method performs a word segmentation process on the image to produce a sub-image and extracts a set of feature vectors from the sub-image. Then, the method performs an asymmetric approach that computes a first log-likelihood score of the feature vectors using a word model having a first structure (such as one comprising a Hidden Markov Model (HMM)) and also computes a second log-likelihood score of the feature vectors using a background model having a second structure (such as one comprising a Gaussian Mixture Model (GMM)). The method computes a final score for the sub-image by subtracting the second log-likelihood score from the first log-likelihood score. The final score is then compared against a predetermined standard to produce a word identification result and the word identification result is output.

Claims

1. A method comprising:

receiving an image of a handwritten item;

performing a word segmentation process on said image to produce a set of sub-images;

extracting a set of feature vectors from each sub-image;

computing a first log-likelihood score of said feature vectors using a word model having a first structure;

computing a second log-likelihood score of said feature vectors using a background model having a second structure different than said first structure;

computing a final score for said sub-image by subtracting said second log-likelihood score from said first log-likelihood score;

comparing said final score against a predetermined standard to produce a word identification result; and

outputting said word identification result.

2. The method according to claim 1, all the limitations of which are incorporated herein by reference, wherein said model comprises a Hidden Markov Model (HMM).

3. The method according to claim 2, all the limitations of which are incorporated herein by reference, wherein said first structure comprises a different number of states than said second structure.

4. The method according to claim 2, all the limitations of which are incorporated herein by reference, wherein said computing of said first log-likelihood score and said computing of said second log-likelihood score comprise one of continuous training and semi-continuous training.

5. The method according to claim 4, all the limitations of which are incorporated herein by reference, wherein said continuous training uses different sets of Gaussian functions for said word model and said background model, and

wherein said semi-continuous training uses the same set of Gaussian functions for said word model and said background model.

6. The method according to claim 2, all the limitations of which are incorporated herein by reference, wherein said computing of said first log-likelihood score and said computing of said second log-likelihood score have relatively lower computational costs when said second structure is simpler than said first structure, and relatively higher computational costs when said second structure is the same as said first structure.

7. A method comprising:

receiving an image of a handwritten item;

performing a word segmentation process on said image to produce a sub-image;

extracting a set of feature vectors from said sub-image;

computing a first log-likelihood score of said feature vectors using a word model having a first structure comprising a Hidden Markov Model (HMM);

computing a second log-likelihood score of said feature vectors using a background model having a second structure comprising a Gaussian Mixture Model (GMM);

computing a final score for said sub-image by subtracting said second log-likelihood score from said first log-likelihood score;

comparing said final score against a predetermined standard to produce a word identification result; and

outputting said word identification result.

8. The method according to claim 7, all the limitations of which are incorporated herein by reference, wherein said first structure comprises a different number of states than said second structure.

9. The method according to claim 7, all the limitations of which are incorporated herein by reference, wherein said computing of said first log-likelihood score and said computing of said second log-likelihood score comprise one of continuous training and semi-continuous training.

10. The method according to claim 9, all the limitations of which are incorporated herein by reference, wherein said continuous training uses different sets of Gaussian functions for said word model and said background model, and

wherein said semi-continuous training uses the same set of Gaussian functions for said word model and said background model.

11. The method according to claim 7, all the limitations of which are incorporated herein by reference, wherein said computing of said first log-likelihood score and said computing of said second log-likelihood score have relatively lower computational costs when said second structure is simpler than first structure, and relatively higher computational costs when said second structure is the same as said first structure.

12. A system comprising:

an input/output device adapted to receive an image of a handwritten item;

a word segmenter operatively connected to said input/output device, wherein said word segmenter is adapted to segment said image to produce a sub-image;

an extractor operatively connected to said word segmenter, wherein said extractor is adapted to extract a set of feature vectors from said sub-image;

a processor operatively connected to said extractor, wherein said processor is adapted to;

compute a first log-likelihood score of said feature vectors using a word model having a first structure;

compute a second log-likelihood score of said feature vectors using a background model having a second structure different than said first structure; and

compute a final score for said sub-image by subtracting said second log-likelihood score from said first log-likelihood score; and

a comparator operatively connected to said processor, wherein said comparator is adapted to compare said final score against a predetermined standard to produce a word identification result,

wherein said input/output device is further operatively connected to said comparator and is adapted to output said word identification result.

13. The system according to claim 12, all the limitations of which are incorporated herein by reference, wherein said model comprises a Hidden Markov Model (HMM).

14. The system according to claim 13, all the limitations of which are incorporated herein by reference, wherein said first structure comprises a different number of states than said second structure.

15. The system according to claim 13, all the limitations of which are incorporated herein by reference, wherein during said computing of said first log-likelihood score and said computing of said second log-likelihood score said processor is further adapted to perform one of continuous training and semi-continuous training.

16. The system according to claim 15, all the limitations of which are incorporated herein by reference, wherein said continuous training uses different sets of Gaussian functions for said word model and said background model, and

wherein said semi-continuous training uses the same set of Gaussian functions for said word model and said background model.

17. The system according to claim 13, all the limitations of which are incorporated herein by reference, wherein during said computing of said first log-likelihood score and said computing of said second log-likelihood score said processor has relatively lower computational costs when said second structure is simpler than said first structure, and relatively higher computational costs when said second structure is the same as said first structure.

18. A system comprising:

an input/output device adapted to receive an image of a handwritten item;

a word segmentator operatively connected to said input/output device, wherein said word segmentator is adapted to segment said image to produce a sub-image;

an extractor operatively connected to said word segmentator, wherein said extractor is adapted to extract a set of feature vectors from said sub-image;

a processor operatively connected to said extractor, wherein said processor is adapted to;

compute a first log-likelihood score of said feature vectors using a word model having a first structure comprising a Hidden Markov Model (HMM);

compute a second log-likelihood score of said feature vectors using a background model having a second structure comprising a Gaussian Mixture Model (GMM); and

compute a final score for said sub-image by subtracting said second log-likelihood score from said first log-likelihood score; and

a comparator operatively connected to said processor, wherein said comparator is adapted to compare said final score against a predetermined standard to produce a word identification result,

wherein said input/output device is further operatively connected to said comparator and is adapted to output said word identification result.

19. The system according to claim 18, all the limitations of which are incorporated herein by reference, wherein said first structure comprises a different number of states than said second structure.

20. The system according to claim 18, all the limitations of which are incorporated herein by reference, wherein during said computing of said first log-likelihood score and said computing of said second log-likelihood score said processor is further adapted to perform one of continuous training and semi-continuous training.

21. The system according to claim 20, all the limitations of which are incorporated herein by reference, wherein said continuous training uses different sets of Gaussian functions for said word model and said background model, and

wherein said semi-continuous training uses the same set of Gaussian functions for said word model and said background model.

22. The system according to claim 18, all the limitations of which are incorporated herein by reference, wherein during said computing of said first log-likelihood score and said computing of said second log-likelihood score said processor has relatively lower computational costs when said second structure is simpler than said first structure, and relatively higher computational costs when said second structure is the same as said first structure.