Open set recognition using transduction

• US 20060093208A1
• Filed: 03/10/2005
• Published: 05/04/2006
• Est. Priority Date: 10/29/2004
• Status: Active Grant

• Current Assignee: George Mason Intellectual Properties, Inc.
• Original Assignee: George Mason Intellectual Properties, Inc.
• Inventors: Wechsler, Harry, Li, Fayin
• Granted Patent: US 7,492,943 B2
• Time in Patent Office: Days
• US Class Current: 382/159
• CPC Class Codes: G06F 18/24147 Distances to closest patterns, e.g. nearest neighbour classification
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 40/10 Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
• CPC Class Codes: G06V 40/172 Classification, e.g. identification
• CPC Class Codes: G06V 40/58 Solutions for unknown imposter distribution

Abstract

Disclosed is an open set recognition system that utilizes transductive inference. One embodiment of this open set recognition system includes capture device(s), a basis, quality checker(s), feature extractor(s), a gallery, a rejection threshold, a storage mechanism, and a recognition stage. The basis is configured to encode the sample(s) and is derived using representative training samples. The feature extractor(s) generates signature(s) from sample(s) using the basis. The rejection threshold may be created using a rejection threshold learning mechanism configured to calculate the rejection threshold using the sample(s) by: swapping one of the sample identifier with other possible sample identifier(s); computing a credibility value (p) for each of the swapped sample identifiers; deriving a peak-to-side ratio distribution using a multitude of the credibility values; and determining the rejection threshold using the peak-to-side ratio distribution. The open set recognition stage authenticates or reject as unknown the identity of unknown sample(s) by: deriving a set of credibility values by iteratively assigning each of the gallery identifiers to the unknown sample; deriving a peak-to-side ratio for the unknown sample using the set of credibility values; comparing the peak-to-side ratio for the unknown sample to the rejection threshold; rejecting the unknown sample as unknown if the peak-to-side ratio is less than or equal to the rejection threshold; and finding the closest of the at least one gallery sample if the peak-to-side ratio is greater than the rejection threshold.

Claims

1. A pattern recognition system comprising:

a) at least one capture device configured to acquire at least one sample, each of said at least one sample associated with a sample identifier;

b) a basis configured to encode at least one of said at least one sample, said basis derived using a multitude of representative training samples;

c) at least one quality checker configured to evaluate the quality of at least one of said at least one sample;

d) at least one feature extractor configured to generate at least one signature from at least one of said at least one sample using said basis;

e) a gallery, said gallery including at least one gallery sample, each of said at least one gallery sample being one of said at least one signature;

f) a rejection threshold, said rejection threshold created using a rejection threshold learning mechanism, said rejection threshold learning mechanism configured to calculate said rejection threshold using at least one of said at least one sample by;

i) swapping one of said sample identifier with other possible said sample identifier;

ii) computing a credibility value (p) for each of the swapped sample identifiers;

iii) deriving a peak-to-side ratio (PSR) distribution using a multitude of said credibility value; and

iv) determining said rejection threshold using said peak-to-side ratio distribution;

g) a storage mechanism configured to store at least one of said at least one gallery sample;

and

h) an open set recognition stage configured to authenticate or reject as unknown the identity of at least one unknown sample, by;

i) deriving a set of credibility values by iteratively assign each of the gallery identifiers to the unknown sample and calculating a credibility value;

ii) deriving a peak-to-side ratio for said unknown sample using said set of credibility values;

iii) comparing said peak-to-side ratio for said unknown sample to said rejection threshold;

iv) rejecting said unknown sample as unknown if said peak-to-side ratio is less than or equal to said rejection threshold; and

v) finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold.

2. A system according to claim 1, wherein said step of finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold includes calculating a credibility value for the closest of said at least one gallery sample.

3. A system according to claim 1, wherein said step of finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold includes calculating a confidence value for the closest of said at least one gallery sample.

4. A system according to claim 1, further including an error analysis stage configured to identify difficult to recognize samples.

5. A system according to claim 4, wherein difficult to recognize samples are processed using data fusion techniques.

6. A system according to claim 4, wherein difficult to recognize samples are processed using multiple representations.

7. A system according to claim 1, wherein at least one of said at least one sample is a biometric sample.

8. A system according to claim 1, wherein said basis is used to derive standard PCA and/or Fisherfaces coefficients.

9. A system according to claim 1, wherein said basis is pre-derived using said training samples.

10. A system according to claim 1, wherein at least one of said at least one capture device is one of the following:

a) a still image camera;

b) a video camera;

c) a micro array; and

d) a data acquisition instrument.

11. A system according to claim 1, wherein said quality checker calculates an intraocular distance and verifies said intraocular distance is within a predetermined range.

12. A system according to claim 1, wherein said quality checker calculates a signal to noise ratio.

13. A system according to claim 1, wherein said quality checker looks for the presence of facial landmarks.

14. A system according to claim 1, wherein at least one of said at least one sample is an image of a face.

15. A system according to claim 1, wherein said at least one sample is a multitude of samples and said one gallery sample is a multitude of gallery samples.

16. A system according to claim 1, wherein the pattern recognition system is used for video matching.

17. A system according to claim 1, wherein the pattern recognition system is used by a visual search engine.

18. A system according to claim 1, wherein the pattern recognition system is used for outlier detection when engaged in clustering.

19. A system according to claim 1, wherein the pattern recognition system is used for novelty detection.

20. A system according to claim 1, wherein the pattern recognition system is used for change detection.

23. A method according to claim 1, wherein said step of finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold includes calculating a confidence value for the closest of said at least one gallery sample.

24. A method according to claim 1, further including the step of selecting features for enhanced pattern recognition using strangeness and a p-value function where the stranger the feature values are the better the discrimination between patterns.

25. A method according to claim 1, wherein at least one of said at least one sample is a biometric sample.

21. A pattern recognition method comprising the steps of:

a) acquiring at least one sample acquired from at least one capture device, each of said at least one sample associated with a sample identifier;

b) encoding at least one of said at least one sample using a basis, said basis derived using a multitude of representative training samples;

c) evaluating the quality of at least one of said at least one sample;

d) generating at least one signature from at least one of said at least one sample using said basis;

e) storing at least one of said at least one signature in a gallery, each of said at least one signature being a gallery sample;

f) calculating a rejection threshold using at least one of said at least one sample by;

i) swapping one of said sample identifier with other possible said sample identifier;

ii) computing a credibility value (p) for each of the swapped sample identifiers;

iii) deriving a peak-to-side ratio (PSR) distribution using a multitude of said credibility value; and

iv) determining said rejection threshold using said peak-to-side ratio distribution; and

g) authenticating or rejecting as unknown the identity of at least one unknown sample, by;

i) deriving a set of credibility values by iteratively assign each of the gallery identifiers to the unknown sample and calculating a credibility value;

ii) deriving a peak-to-side ratio for said unknown sample using said set of credibility values;

iii) comparing said peak-to-side ratio for said unknown sample to said rejection threshold;

iv) rejecting said unknown sample as unknown if said peak-to-side ratio is less than or equal to said rejection threshold; and

v) finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold.

22. A method according to claim 21, wherein said step of finding the closest of said at least one gallery sample if said peak-to-side ratio is greater than said rejection threshold includes calculating a credibility value for the closest of said at least one gallery sample.