US7027620B2
Expired / Lapsed
US
⏱️ 0.0 years remaining
Method of recognizing partially occluded and/or imprecisely localized faces
Assignee: Sony Corp
•
Filing: 2001-06-07
•
Granted: 2006-04-11
•
Expires: 2021-06-07
Abstract
Abstract A method of recognizing a test face given an imprecise localization and/or a partially occluded face. Using an eigenspace representation of the learning image, a subspace is computed that represents localization errors within the eigenspace. This is computed by calculating an actual localization error of the localization of the learning image. Accounting for the calculated actual localization error, all possible morphed faces are projected onto the eigenspace. For all possible faces, the subspace is modeled where all of the possible localizations of the possible faces lie using a Gaussian distribution. To accommodate for possible partial occlusions, for all possible faces, the possible faces are divided into n different local parts. Each of the n local parts are modeled using a Gaussian distribution. A global probability of the test face is computed by adding all local probabilities defined by the Gaussian distribution. The test face is recognized based on the computed subspace and/or the accounted for possible partial occlusions.
Inventors
- 👤 Aleix M. Martinez
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 10:44
CPC Classifications
G; G06; G06V; G06V40/00; G06V40/10; G06V40/16; G06V40/172
Claims Analysis
Total Claims: 10
Claim 1
Independent Claim
115 words
A method of recognizing an imprecisely localized test face, comprising: a) during a learning stage, using a feature space representation of a learning image for each of plurality of subjects, computing a subspace that represents localization errors within the feature space by 1) calculating an actual localization error of the localization of the learning images; 2) accounting for the calculated actual localization error, projecting all possible morphed faces onto the feature space; and 3) for all possible faces, modeling the subspace where all of the possible localizations of the possible faces lie using a Gaussian distribution; and b) during an identification stage, recognizing a test face based on the computed subspace from the learning stage.
Claim 2
Independent Claim
115 words
A method of recognizing an imprecisely localized test face, comprising: a) during a learning stage, using a feature space representation of a learning image for each of plurality of subjects, computing a subspace that represents localization errors within the feature space by 1) calculating an actual localization error of the localization of the learning images; 2) accounting for the calculated actual localization error, projecting all possible morphed faces onto the feature space; and 3) for all possible faces, modeling the subspace where all of the possible localizations of the possible faces lie using a Gaussian distribution; and b) during an identification stage, recognizing a test face based on the computed subspace from the learning stage.
Claim 3
Independent Claim
72 words
A method of recognizing a partially occluded test face, comprising: for all possible faces, dividing the possible face into n different local parts; modeling each of the n local parts using a Gaussian distribution corresponding to a predetermined subset that represents localization error; computing a global probability of the test face by adding all local probabilities defined by the Gaussian distribution; and recognizing the test face based on the computed global probability.
Claim 4
Independent Claim
72 words
A method of recognizing a partially occluded test face, comprising: for all possible faces, dividing the possible face into n different local parts; modeling each of the n local parts using a Gaussian distribution corresponding to a predetermined subset that represents localization error; computing a global probability of the test face by adding all local probabilities defined by the Gaussian distribution; and recognizing the test face based on the computed global probability.
Claim 5
Dependent Claim
65 words
The method of claim 1 , further comprising the steps of: prior to computing the subspace, computing the localization of the learning image by: pre-localizing the learning image; omitting unwanted segments; localizing both eyes and the mouth of the learning image; rotating the face until the learning image is frontal; identifying boundaries of the learning image; and morphing the learning image to a standard face.
Claim 6
Dependent Claim
48 words
The method of claim 1 , wherein the step of calculating the actual localization error comprises: manually localizing a number of pertinent features of all possible faces; and computing the actual localization error as a difference between the localization of the learning image and the manually localized features.
Claim 7
Dependent Claim
113 words
The method of claim 2 , wherein the modeling step further comprises the steps of: generating a local eigenspace representation by: applying a localization stage to each of the possible faces; generating a learning set; finding a mean vector of each element of the learning set; computing covariance matrices for each element of the learning set; obtaining the local eigenspaces by taking the eigenvectors associated with the largest eigenvalues of the covariance matrices; and generating a Gaussian model by: constructing all eigenspaces by finding eigenvectors associated to the highest eigenvalues of each covariance matrix; and searching for subspaces that account for the localization error by projecting all learning instances onto the constructed eigenspaces.
Claim 8
Dependent Claim
73 words
The method of claim 2 , wherein the step of computing the global probability of the test face further comprises the steps of: localizing the test face; morphing the test face to a final standard face; projecting each of the n local parts onto computed eigenspaces, wherein a probability of a given match is directly associated to a Mahalanobis distance; adding all local Mahalanobis distances; and searching for minima from the added distances.
Claim 9
Independent Claim
175 words
A method of recognizing a test face, the method comprising: a) using a feature space representation of a learning image for each of a plurality of subjects, computing a subspace that represents localization errors within the feature space by 1) calculating an actual localization error of the localization of the learning image; 2) accounting for the calculated actual localization error, projecting all possible morphed faces onto the feature space; and 3) for all possible faces, modeling the subspace where all of the possible localizations of the possible faces lie using a Gaussian distribution; b) accounting for possible partial occlusions of the test face by: 1) for all possible faces, dividing the possible face into n different local parts; 2) modeling each of the n local parts using a Gaussian distribution corresponding to a predetermined subset that represents localization error; 3) computing a global probability of the test face by adding all local probabilities defined by the Gaussian distribution; and recognizing the test face based on the computed subspace and/or the accounted for possible partial occlusions.
Claim 10
Independent Claim
175 words
A method of recognizing a test face, the method comprising: a) using a feature space representation of a learning image for each of a plurality of subjects, computing a subspace that represents localization errors within the feature space by 1) calculating an actual localization error of the localization of the learning image; 2) accounting for the calculated actual localization error, projecting all possible morphed faces onto the feature space; and 3) for all possible faces, modeling the subspace where all of the possible localizations of the possible faces lie using a Gaussian distribution; b) accounting for possible partial occlusions of the test face by: 1) for all possible faces, dividing the possible face into n different local parts; 2) modeling each of the n local parts using a Gaussian distribution corresponding to a predetermined subset that represents localization error; 3) computing a global probability of the test face by adding all local probabilities defined by the Gaussian distribution; and recognizing the test face based on the computed subspace and/or the accounted for possible partial occlusions.
References Cited (Backward Citations)
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2009-04-10 | 2009-10-10 | 2010-04-11 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2013-04-10 | 2013-10-10 | 2014-04-11 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2017-04-10 | 2017-10-10 | 2018-04-11 | $7700.00 | Window Closed / Lapsed |