Open set recognition using transduction
View Patent ↗An open set recognition system utilizing transductive inference including 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 encodes sample(s) and is derived using training samples. The feature extractor(s) generates signature(s) from sample(s) using the basis. The rejection threshold is created using a rejection threshold learning mechanism that calculates the rejection threshold using sample(s) by: swapping a sample identifier with other sample identifier(s); computing a credibility value for the swapped sample identifiers; deriving a peak-to-side ratio distribution using 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) using derived credibility values, derived peak-to-side ratios for the unknown sample and the rejection threshold.
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 feature extractor configured to generate at least one signature from at least one of said at least one sample using said basis;
d) 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;
e) 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;
f) a storage mechanism configured to store at least one of said at least one gallery sample; and
g) 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 said difficult to recognize samples are processed using data fusion techniques.
6. A system according to claim 4 , wherein said 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 at least some of said multitude of representative 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 , further including at least one quality checker configured to evaluate quality of at least one of said at least one sample using a calculated signal to noise ratio.
12. A system according to claim 1 , wherein at least one of said at least one sample is an image of a face.
13. 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.
14. A system according to claim 1 , wherein the pattern recognition system is used for outlier detection when engaged in clustering.
15. A system according to claim 1 , wherein the pattern recognition system is used for novelty detection.
16. A system according to claim 1 , wherein the pattern recognition system is used for change detection.
17. A pattern recognition method comprising using a system to perform 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)generating at least one signature from at least one of said at least one sample using said basis;
d) storing at least one of said at least one signature in a gallery, each of said at least one signature being a gallery sample;
e) 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
f) 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.
18. A method according to claim 17 , 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.
19. A method according to claim 17 , 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.
20. A method according to claim 17 , 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.
21. A method according to claim 17 , wherein at least one of said at least one sample is a biometric sample.