Feature extraction system and face image recognition system

• US 5,901,244 A
• Filed: 12/12/1996
• Issued: 05/04/1999
• Est. Priority Date: 06/18/1996
• Status: Expired due to Term

• Current Assignee: Panasonic Intellectual Property Corporation of America (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Souma, Masaki, Nagao, Kenji
• Primary Examiner(s): Tran, Phuoc
• Application Number:
  US08/764,546
• Time in Patent Office: 873 Days
• Field of Search: 382/118, 382/160, 382/190, 382/201, 382/224, 382/225, 382/228, 382/276
• US Class Current: 382/190
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 40/16 Human faces, e.g. facial parts, sketches or expressions

Abstract

A feature extraction system for statistically analyzing a set of samples of feature vectors to calculate a feature being an index for a pattern identification, which is capable of identifying confusing data with a high robustness. In this system, a storage section stores a feature vector inputted through an input section and a neighborhood vector selection section selects a specific feature vector from the feature vectors existing in the storage section. The specific feature is a neighborhood vector close in distance to the feature vector stored in the storage section. Further, the system is equipped with a feature vector space production section for outputting a partial vector space. The partial vector space is made to maximize the local scattering property of the feature vector when the feature vector is orthogonally projected to that space.

Claims

1. A feature extraction system for statistically analyzing a set of samples of feature vectors to calculate an amount indicative of a feature which plays an important role when pattern-identification is performed, said system comprising:

feature vector input means for inputting a feature vector;

storage means for storing said feature vector inputted through said input means;

neighborhood vector selection means for selecting a neighborhood vector, being a feature vector close to said feature vector inputted and stored in said storage means, from feature vectors existing in said storage means; and

feature vector space production means for outputting a partial vector space in which feature vectors locally most disperse when said feature vectors are orthogonally projected thereinto,

wherein, when said feature vectors stored in said storage means are taken to be V_i (1≦

i≦

N) and the dimension of these feature vectors is taken as n, said neighbor vector selection means selects M (≦

N) feature vectors close to these feature vectors through the use of a distance d(x, y), where x and y represent feature vectors, determined according to the following equation (1);

##EQU16## and said feature vector space production means obtains eigenvectors by solving an eigenvalue problem of a local covariance matrix given from the following equation (2);

##EQU17## where V_j.sup.(i) denotes a neighborhood vector to V_i and n(i) depicts the total number of possible V_j.sup.(i) which varies at every i, moreover, said feature vector space production means selects m (≦

N, n) eigenvectors from the obtained eigenvectors in the order of decreasing corresponding eigenvalue to output them as bases of a partial vector space which maximizes a local scattering property.

2. A feature extraction system as defined in claim 1, wherein said neighborhood vector selection means selects said M feature vectors close to the respective feature vectors through the use of a correlation value cor(x, y), where x and y denote feature vectors, determined according to the following equation (3) in place of the above-mentioned equation (1) ##EQU18## where x=(x_i) and y=(y_i).

3. A feature extraction system as defined in claim 1, wherein, in place of calculating said distance according to the above-mentioned equation (1), said neighborhood vector selection means first solves an eigenvalue problem of a covariance matrix obtained according to the following equation (4):

##EQU19## and selects one eigenvector ei (1≦

i≦

1) corresponding to a large eigenvalue, and subsequently said neighborhood vector selection means calculates a distance k1(x, y), where x and y express feature vectors, according to the following equation (5);

##EQU20##

4. A feature extraction system as defined in claim 1, wherein, in place of calculating said distance according to the above-mentioned formula (1), said neighborhood vector selection means solves an eigenvalue problem of a covariance matrix obtained according to the following equation (4):

and selects one eigenvector ei (1≦

i≦

1) corresponding to a large eigenvalue, and subsequently said neighborhood vector selection means calculates a distance math(x, y), where x and y express feature vectors, according to the following equation (6);

##EQU21## where λ

_i represents an eigenvalue corresponding to an eigenvector e_i, and σ

² designates a bias.

5. A feature extraction system as defined in claim 1, wherein said feature vector space production means outputs, in addition to said m eigenvectors, eigenvalues corresponding to said m eigenvectors.

6. A feature extraction system as defined in claim 1, wherein said feature vector space production means outputs, in addition to said m eigenvectors, dispersion values σ

(ei) determined according to the following equation (7) corresponding to these eigenvectors ei;

##EQU22##

7. A face image recognition system comprising:

model face image input means for receiving a model face image;

first feature extraction means for extracting a feature from said model face image;

first model feature vector holding means for holding a result of the feature extraction obtained when said first feature extraction means extracts said feature from said face image inputted through said model face image input means;

first and second model limitation means for taking a set of models having feature vectors satisfying a given condition on the variation from a feature vector of each of the model face images;

model eigenspace holding means for calculating the greatest to Nth eigenvalues of eigenvalues of a scattering property evaluation matrix defined by the following equation (8) in relation to said set of model faces limited by said first model limitation means and eigenvectors corresponding to the greatest to Nth eigenvalues and further for holding these calculation results at every model, ##EQU23## where M_i represents the i^th model face vector, M_i^j designates the j^th model face vector belonging to the neighborhood of said model M_i, N_M denotes the total number of model face images, N_Mi depicts the number of models belonging to the neighborhood of said model M_i, and T signifies the transposition of a matrix (vector);

object face image input means for receiving an object face image under recognition;

second feature extraction means for extracting a feature from said face image;

first object image feature vector holding means for holding a feature vector extracted by said second feature extraction means from said object face image inputted through said object face image input means;

second model feature vector holding means for holding a feature vector extracted by said first feature extraction means from each of the model face images inputted through said model face image input means;

first model selection means for finding a model feature vector of the model feature vectors held in said second model feature vector holding means, said model feature vector to be found showing a minimum variation with respect to the object image feature vector; and

second model selection means for selecting a model face vector from among the model face vectors M selected by said first model selection means and limited by said second model limitation means, said model face vector to be selected being a vector by which variation of a point m obtainable in a manner of being projected onto a space defined by N eigenvectors Ej (j=1 . . . N) held in said model eigenspace holding means according to the following equation (9) with respect to a point i of the object face vector I projected to the same space according to the same equation (9) stands at a minimum,

space="preserve" listing-type="equation">m=(m.sub.1, m.sub.2, . . . ,m.sub.N)

space="preserve" listing-type="equation">i=(i.sub.1, i.sub.2, . . . ,i.sub.N)

space="preserve" listing-type="equation">m.sub.k =(E.sub.k, M)

space="preserve" listing-type="equation">i.sub.l =(E.sub.l, I) (9)

where m_k and i_l respectively denote the k^th and l^th coordinate values on the projection of said model face vector M and an object face vector I to an eigenspaces, and (a,b) depicts the inner product of vectors a and b.

8. A face image recognition system as defined in claim 7, wherein said second model selection means selects a model face vector, of the model face vectors restricted by said second model limitation means, said model face vector selected showing the fact that the variation of a point m obtained in the manner of being projected onto a space made by N eigenvectors Ej (j=1 . . . N) held in said model eigenspace holding means according to said equation (9) with respect to a point i of an object face vector I projected onto the same eigenspace according to the same equation (9), which variation is calculated in accordance with the following equation (10), stands at a minimum, ##EQU24## where σ

_l² signifies the dispersion of a model face vector in the direction of a vector E_l.

9. A face image recognition system as defined in claim 7, wherein said second model selection means selects a model face vector, of the model face vectors restricted by said second model limitation means, said model face vector selected showing the fact that the variation of a point m obtained in the manner of being projected onto a space made by N eigenvectors Ej (j=1 . . . N) held in said model eigenspace holding means according to said equation (9) with respect to a point i of an object face vector I projected onto the same eigenspace according to the same equation (9), which variation is calculated in accordance with the following equation (11), stands at a minimum, ##EQU25##10.

10. A face image recognition system as defined in claim 7, wherein said second model selection means selects a model face vector, of the model face vectors restricted by said second model limitation means, said model face vector selected showing the fact that the variation of a point m obtained in the manner of being projected onto a space made by N eigenvectors Ej (j=1 . . . N) held in said model eigenspace holding means according to said equation (9) with respect to a point i of an object face vector I projected onto the same eigenspace according to the same equation (9), which variation is calculated in accordance with the following equation (12), comes to a minimum, where (m_lj -i_lj)² signifies the j^th value of (m_l -i_l)² (l=1, . . . , N) when counting from the minimum value.

11. A face image recognition system as defined in claim 7, wherein said second model selection means selects a model face vector, of the model face vectors restricted by said second model limitation means, said model face vector selected showing the fact that the variation of a point m obtained in the manner of being projected onto a space made by N eigenvectors Ej (j=1 . . . N) held in said model eigenspace holding means according to said equation (9) with respect to a point i of an object face vector I projected onto the same eigenspace according to the same equation (9), which variation is calculated in accordance with the following equation (13), comes to a minimum, ##EQU26## where (m_lj -i_lj)² /σ

_lj² signifies the j^th value of (m_l -i_l)² (l=1, . . . , N) when counting from the minimum value, and σ

_l² signifies the dispersion of a model face vector in the direction of a vector.