Pattern recognition apparatus and method

• US 6,466,685 B1
• Filed: 07/13/1999
• Issued: 10/15/2002
• Est. Priority Date: 07/14/1998
• Status: Expired due to Term

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Maeda, Kenichi, Suzuki, Kaoru, Yamaguchi, Osamu, Fukui, Kazuhiro
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): Kassa, Yosef
• Application Number:
  US09/352,839
• Time in Patent Office: 1,190 Days
• Field of Search: 382/115, 382/117, 382/118, 382/155, 382/156, 382/158, 382/159, 382/160
• US Class Current: 382/115
• CPC Class Codes: G06V 40/172 Classification, e.g. identification

Abstract

An image pattern of an object is inputted. An input subspace calculation section calculates an input subspace from the image pattern. A dictionary subspace calculation section calculates a dictionary subspace from a learning pattern of each object. A constraint subspace calculation means calculates a constraint subspace from a plurality of input subspaces previously calculated according to constraints to suppress unnecessary patterns. A projection section projects the input subspace and the dictionary subspace onto the constraint subspace. A recognition section recognizes the object by comparing the projected input subspace with the projected dictionary subspace.

Claims

1. A pattern recognition apparatus, comprising:

image input means for inputting a plurality of face images of each of a plurality of people under a condition of similar illumination, each face image of the same person being input at a different time;

input subspace calculation means for respectively calculating at least one subspace of each person from the plurality of face images of each person;

difference subspace calculation means for respectively calculating a difference between two subspaces of two peolpe, and for calculating a common susbspace of all differences of each two people as a difference subspace;

self variation subspace calculation means for respectively calculating a difference between two subspaces of the same person, and for calculating a common subspace of all differences of each person as a self variation subspace;

constraint subspace calculation means for calculating a constraint subspace by subtracting the self variation subspace from the difference subspace;

dictionary subspace memory for storing a plurality of dictionary subspaces each representing a face pattern of a registered person;

projection means for projecting an input subspace and each dictionary subspace onto the constraint subspace, when said image input means inputs a face image of a person in question and said input subspace calculation means calculates the input subspace from the face image; and

recognition means for recognizing the person in question by comparing the projected input subspace with each projected dictionary subspace.

2. The pattern recognition apparatus according to claim 1,

wherein said recognition means respectively calculates a canonical angle between the projected input subspace and each projected dictionary subspace, and respectively calculates a similarity degree between the projected input subspace and each projected dictionary subspace according to the canonical angle.

3. The pattern recognition apparatus according to claim 2,

wherein said recognition means decides the person in question as the registered person of the dictionary pattern of the highest similarity degree.

4. The pattern recognition apparatus according to claim 1,

wherein said difference subspace calculation means extracts a principal component subspace from the difference subspace.

5. The pattern recognition apparatus according to claim 4,

wherein said self variation subspace calculation means extracts a principal component subspace from all subspaces of each person as the self variation subspace.

6. The pattern recognition apparatus according to claim 5,

wherein said constraint subspace calculation means calculates an orthogonal component subspace for the self variation subspace, and calculates the constraint subspace as a sum of the difference subspace and the orthogonal component subspace.

7. The pattern recognition apparatus according to claim 4,

wherein said difference subspace calculation means respectively calculates a difference vector between each two subspaces from all subspaces, normalizes each difference vector, and generates a projection matrix from each normalized difference vector.

8. The pattern recognition apparatus according to claim 7,

wherein said difference subspace calculation means calculates a generation matrix as an average of each projection matrix, calculates eigenvalues larger than a threshold from the generation matrix, and regards an eigenvector corresponding to the eigenvalues as a base vector of the difference subspace.

9. The pattern recognition apparatus according to claim 1,

wherein said input means previously inputs a plurality of face images of each registered person under the condition of similar illumination, and

wherein said dictionary subspace calculation means respectively calculates the dictionary subspace from the plurality of face images of each registered person.

10. The pattern recognition apparatus according to claim 1,

further comprising feature vector extraction means for reducing an influence of illumination for the image pattern, extracting a feature vector from the image pattern from which the influence of illumination is reduced, and supplying the feature vector to said input subspace calculation means.

11. The pattern recognition apparatus according to claim 10,

wherein said feature vector extraction means reduces the influence of illumination for a predetermined area of the image pattern, extracts first feature vector from the predetermined area of the image pattern and second feature vector from other area of the image pattern, and supplies the first feature vector and the second feature vector to said input subspace calculation means.

12. The pattern recognition apparatus according to claim 10,

wherein said feature vector extraction means reduces the influence of illumination by high pass filtering.

13. The pattern recognition apparatus according to claim 10,

wherein said feature vector extraction means reduces the influence of illumination by a Fourier transform process to determine spectral intensity distribution of spatial frequency component.

14. The pattern recognition apparatus according to claim 10,

wherein said feature vector extraction means reduces the influence of illumination by expanding a range of low intensity areas and by relatively compressing a range of high intensity areas.

15. The pattern recognition apparatus according to claim 12,

wherein said feature vector extraction means eliminates small edges from the image pattern after the high pass filtering.

16. The pattern recognition apparatus according to claim 13,

wherein said feature vector extraction means eliminates low frequency components from the image pattern after the Fourier transform process.

17. The pattern recognition apparatus according to claim 12,

wherein the high pass filtering comprises differentiation processing.

18. The pattern recognition apparatus according to claim 17,

wherein the differentiation processing comprises embossing processing to derivate edges along a predetermined direction on the image pattern.

19. A pattern recognition method, comprising:

inputting a plurality of face images of each of a plurality of people under a condition of similar illumination, each face image of the same person being input at a different time;

respectively calculating at least one subspace of each person from the plurality of face images of each person;

respectively calculating a difference between two subspaces of two people;

calculating a common subspace of all differences of each two people as a difference subspace;

respectively calculating a difference between two subspaces of the same person;

calculating a common subspace of all differences of each person as a self variation subspace;

calculating a constraint subspace by subtracting the self variation subspace from the difference subspace;

storing a plurality of dictionary subspaces each representing a face pattern of a registered person;

projecting an input subspace and each dictionary subspace onto the constraint subspace, when a face image of a person in question is input and the input subspace is calculated from the face image; and

recognizing the person in question by comparing the projected input subspace with each projected dictionary subspace.

20. A computer readable memory containing computer readable instructions, comprising:

instruction means for causing a computer to input a plurality of face images of each of a plurality of people under a condition of similar illumination, each face image of the same person being input at a different time;

instruction means for causing a computer to respectively calculate at least one subspace of each person from the plurality of face images of each person;

instruction means for causing a computer to respectively calculate a difference between two subspaces of two people;

instruction means for causing a computer to calculate a common subspace of all differences of each two people as a difference subspace;

instruction means for causing a computer to respectively calculate a difference between two subspaces of the same person;

instruction means for causing a computer to calculate a common subspace of all differences of each person as a self variation subspace;

instruction means for causing a computer to calculate a constraint subspace by subtracting the self variation subspace from the difference subspace;

instruction means for causing a computer to store a plurality of dictionary subspaces each representing a face pattern of a registered person;

instruction means for causing a computer to project an input subspace and each dictionary subspace onto the constraint subspace, when a face image of a person in question is input and the input subspace is calculated from the face image; and

instruction means for causing a computer to recognize the person in question by comparing the projected input subspace with each projected dictionary subspace.

21. A pattern recognition apparatus, comprising:

an image input unit configured to input a plurality of images of each of a plurality of people under a condition of similar illumination, each face image of the same person being input at a different time;

an input subspace calculation unit configured to respectively calculate at least one subspace of each person from the plurality of face images of each person;

a difference subspace calculation unit configured to respectively calculate a difference between two subspaces of two people, and to calculate a common subspace of all differences of each two people as a difference subspace;

a self variation subspace calculation unit configured to respectively calculate a difference between two subspaces of the same person, and to calculate a common subspace of all differences of each person as a self variation subspace;

a constraint subspace calculation unit configured to calculate a constraint subspace by subtracting the self variation subspace from the difference subspace;

a dictionary subspace memory configured to store a plurality of dictionary subspaces each representing a face pattern of a registered person;

a projection unit configured to project an input subspace and each dictionary subspace onto the constraint subspace, when said image input unit inputs a face image of a person in question and said input subspace calculation unit calculates the input subspace from the face image; and

a recognition unit configured to recognize the person in question by comparing the projected input subspace with each projected dictionary subspace.