JP2005099903A - Update device for authenticating dictionary in biometrics authentication - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an update device for an authenticating dictionary in biometrics authentication, capable of updating the authenticating dictionary even in a device having limited accumulation capacity without requiring a biological image used for creating the authenticating dictionary before the update. <P>SOLUTION: The authenticating dictionary 1 before the update is registered with a characteristic amount extracted from the biological images 2-1 to 2-N by principal component analysis (S13). The characteristic amount is read from the authenticating dictionary 1, and a new characteristic amount is calculated by the principal component analysis by use of the read characteristic amount and a newly inputted biological image 2-(N+1) for updating the authenticating dictionary (S20). By the new characteristic amount, the authenticating dictionary 1 before the update is updated to the authenticating dictionary 1' after the update. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for updating a recognition dictionary in biometric authentication, and more particularly to an apparatus for updating a recognition dictionary in biometric authentication in which a feature amount with a relatively small amount of information extracted from a biological image with a large amount of information is registered. About.

  Biometrics authentication using biometric images such as faces and fingerprints has attracted attention for security management such as access to various systems and permission to enter the room. When biometrics authentication is performed, a problem of deterioration of recognition accuracy due to secular change becomes a problem. This is because the person's feature amount extracted from the biometric image is registered correctly in the biometrics recognition device as a recognition dictionary, but over time, the person is It is a problem that it is not correctly recognized as the person.

In order to solve such problems and prevent deterioration of recognition accuracy due to aging of the living body, a method of periodically updating the recognition dictionary is effective. Non-patent document 1 below describes a conventional technique that periodically updates a recognition dictionary to prevent a reduction in recognition accuracy due to aging.
Toshio Sato, Hiroshi Sukekawa, Kentaro Yokoi, Hiroyoshi Dobashi, Kei Ogata, Akio Okazaki: “Development of Face Matching Security System“ Face Pass ”Considering Standing Position Changes”, Journal of the Institute of Image Information and Television Engineers, 56, 7, pp.1111- 1117 (2002)

  However, although it has been described that updating of the recognition dictionary is effective against aging of a living body, how to efficiently update the recognition dictionary is not described.

  For example, in Non-Patent Document 1, the recognition dictionary is updated by adding a biometric image for recognition dictionary update newly input to a past biometric image. The biometric image used to create the dictionary for use is needed.

  FIG. 6 is a flowchart of a conventional recognition dictionary update process. The pre-update (new) recognition dictionary 1 uses a plurality of normalized biological images 2-1, 2-2,..., 2-N to register feature values obtained by principal component analysis. Created by. Further, the characteristics obtained by principal component analysis using the normalized biometric images 2-1, 2-2,..., 2-N and the newly input normalized biometric image 2- (N + 1) for update. By updating to a value, it is updated to the updated recognition dictionary 1 ′.

  As is clear from the above, in the prior art, in order to update the recognition dictionary, it is necessary to retain all past biometric images in the biometric authentication device. However, biometric authentication is performed and its use is permitted. In a device that performs biometric authentication on a mobile terminal such as a mobile phone terminal, since the storage capacity that can be used to hold a biometric image is limited, it is practically difficult to hold all past biometric images. There is a problem that the recognition dictionary cannot be updated.

  An object of the present invention is to solve the above-mentioned problems and does not require a biometric image used for creating a pre-update recognition dictionary, and the recognition dictionary even in an apparatus such as a mobile phone terminal having a limited storage capacity It is an object of the present invention to provide a recognition dictionary updating apparatus for biometric authentication that can update the above.

  In order to solve the above problems, the present invention provides a recognition dictionary update apparatus for biometric authentication in which feature values extracted from a biometric image are registered, and reads out feature values registered in the recognition dictionary. Means, a feature amount calculating means for calculating a new feature amount from the newly input biological image and the feature amount read by the reading means, and a new feature amount calculated by the feature amount calculating means. Registration means for registering in the recognition dictionary as an updated feature quantity.

  Further, according to the present invention, the feature amount corresponds to an eigenvalue matrix in which eigenvalues obtained by performing principal component analysis on an average biometric image, an average biometric image, and a covariance matrix of each biometric image are arranged in a diagonal component, and corresponding And a diagonalization matrix in which eigenvectors are arranged as column vectors, and the feature amount calculating means calculates a new average biological image from the newly input biological image and the average biological image read by the reading means. New biometric image calculation means, newly input biometric image, average biometric image calculated by the average biometric image calculation means, and eigenvalue matrix and diagonalization matrix read by the reading means. A covariance matrix calculating means for calculating a dispersion matrix and a principal component analysis of the covariance matrix calculated by the covariance matrix calculating means to calculate a new eigenvalue matrix and a diagonalization matrix Matrix / diagonalization matrix calculation means, wherein the registration means calculates a new average biological image, a new eigenvalue matrix, and a diagonalization calculated by the average biological image calculation means and the eigenvalue matrix / diagonalization matrix calculation means. A matrix is registered in the recognition dictionary.

  In the present invention, a new feature amount is calculated from the newly input biometric image and the feature amount registered in the recognition dictionary before update, and this is registered in the recognition dictionary as the updated feature amount. Therefore, the biometric image used for creating the recognition dictionary before update is not required, and the recognition dictionary can be updated even in a device such as a portable terminal having a limited storage capacity.

  In other words, it is not necessary to store and hold the biometric image used to create the pre-update recognition dictionary for future recognition dictionary updates, so even biometric authentication devices with limited storage capacity can be recognized. The dictionaries can be updated. Specifically, it can be a biometric authentication device that can efficiently update a recognition dictionary on a mobile terminal such as a mobile phone terminal having a limited storage capacity.

  The present invention will be described below with reference to the drawings. In the following, an example will be described in which biometric authentication is performed by calculating a feature amount from a biological image by principal component analysis and using a face image as the biological image. In the following description, although described as a process, the apparatus can be configured as a device by realizing the function of the process with software or hardware.

  FIG. 1 is a flowchart showing a basic process for creating a new recognition dictionary and a personal authentication process. In the new recognition dictionary creation process, N face images (N is an integer equal to or greater than 1) are acquired (S10), and face regions in these face images are detected (S11). Next, an image of the face area is extracted, normalized so that the image has a standardized luminance and size, and a plurality of normalized face images 2-1, 2-2,. N is obtained (S12). Next, a feature base is obtained by principal component analysis to calculate a feature amount (S13), and a new recognition dictionary 1 is created by registering the feature amount.

  In the personal authentication process, first, a face image is acquired as a biometric image to be authenticated (S10), and a face region in the face image is detected (S11). Next, an image of the face area is extracted and normalized so that the image has a standardized luminance and size, thereby obtaining a normalized face image (S12).

In the feature amount calculation / identification process (S14), the degree of difference or similarity is obtained from the feature amount of the normalized face image and the feature amount previously registered in the recognition dictionary 1, and the personal authentication is determined based on the magnitude of the value. I do.
FIG. 2 is a flow diagram illustrating the recognition dictionary update process in one embodiment of the invention. As described above, the recognition dictionary 1 before update is created in the past using a plurality of normalized face images 2-1, 2-2,..., 2-N.

  When the past recognition dictionary 1 is to be updated to a new recognition dictionary 1 ′ using the new face image 2- (N + 1), the new face image 2- (N + 1) is input and the recognition face image 2- The feature amount registered in the dictionary 1 is read out. In the feature amount calculation process (S20), a new feature amount updated by principal component analysis is calculated from the newly input face image 2- (N + 1) and the feature amount read from the recognition dictionary 1, The past feature quantity is updated to a new feature quantity and updated to the updated recognition dictionary 1 ′.

  The face image authentication using the principal component analysis is specifically shown below using mathematical expressions, and the newly input face image 2- (N + 1) and the pre-update feature amount read from the recognition dictionary 1 The reason why the updated feature value can be derived from the above will be described.

First, the process for creating the new recognition dictionary 1 will be described. FIG. 3 collectively shows this process. If N normalized face images 2-1, 2-2,..., 2-N used to create the new recognition dictionary 1 are x _i (i = 1, 2,..., N), The average face image m and the covariance matrix Σ can be defined by the following equations (1) and (2), respectively.

The covariance matrix Σ is subjected to principal component analysis as shown in the following equation (3), and d (where d <N) eigenvalues λ _i (i = 1, 2,..., D) are in descending order. The eigenvalue matrix Λ shown in the following equation (4) arranged in the diagonal component and d eigenvectors (feature bases) u _i (i = 1, 2,..., D) corresponding to them are column vectors. The diagonal matrix A shown in the following equation (5) is obtained.

The average face image m, the eigenvalue matrix Λ, and the diagonalization matrix A obtained as described above are registered in the new recognition dictionary 1 as the feature values of the biological image. Here, the diagonalization matrix A is created by arranging d feature bases u _i as column vectors.

Next, a process for actually performing personal authentication will be described. FIG. 4 summarizes this process. In the process of performing personal authentication, for example, using the average face image m registered in the recognition dictionary 1 and the feature base u _i , the difference between the normalized face image x input by the following equation (6) The degree D is obtained, and personal authentication is performed based on the degree of difference D. In general, there are many definitions of the degree of difference, and the following formula (6) is merely an example. As the feature base u _i , one stored as a column vector in the diagonalization matrix A registered in the recognition dictionary 1 is used.

  The degree of similarity S expressed by the following equation (7) is obtained, and personal authentication can also be performed based on the value of the degree of similarity S. There are many definitions of similarity, and equation (7) is just one example.

  In the above example, the degree of difference D (≧ 0) indicates that the closer the value is to 0, the closer the input normalized image x is to the person registered in the recognition dictionary 1, and conversely, the degree of similarity ( 0 ≦ S ≦ 1) indicates that the larger the value is, the closer the input normalized image x is to the person registered in the recognition dictionary 1.

Next, the normalized face image x _i (i = 1, 2, _i) used to create the recognition dictionary 1 in the past with respect to the face image authentication apparatus that performs the recognition dictionary creation and personal authentication described above. .., N), the process of updating the recognition dictionary 1 and updating it to a new recognition dictionary 1 ′ will be described. FIG. 5 summarizes this process.

If the above formulas (1) to (3) are used as they are, in order to calculate the average face image m and the d feature bases u _i of the diagonalization matrix A that need to be registered in the recognition dictionary 1, N normalized face images x _i (i = 1, 2,..., N) are required. Therefore, when a new adds the input N + 1 th normalized face image x _{N + 1} would like to update the dictionary 1 'for the new recognition, N + 1 th normalized face image x _{N + 1} as well, for recognition in the past Using all of the N normalized face images x _i (i = 1, 2,..., N) used to create the dictionary 1, the covariance matrix is re-established from a total of N + 1 normalized face images. Σ ′ is obtained.

However, when the N + 1-th normalized face image xN _{+ 1} is newly input and the recognition dictionary is updated, it is assumed that the average face image before update is m and the average face image after update is m ′. The following equation (8) holds.

That is, the updated average face image m ′ does not use the N normalized face images x _i (i = 1, 2,..., N) that were used to create the recognition dictionary 1 in the past. In addition, it can be derived from the newly input N + 1-th normalized face image xN _{+ 1} and the average face image m before update.

  Similarly, when the covariance matrix before updating the recognition dictionary is Σ and the updated covariance matrix is Σ ′, the following equation (9) is established.

  Note that the following formulas (10) and (11) hold in the derivation process of the formula (9) for the following reason.

(M′−m) ^T in the Σ operation is a constant row vector regardless of the change of i, and can be handled like a constant. Therefore, if (m′−m) ^T = p ^T is set, Equation (12) holds.

  Similarly, (m′−m) in the Σ operation is a constant row vector regardless of the change of i, and can be handled like a constant, so that the following expression (13) holds.

  As is clear from the above results, the covariance matrix Σ ′ after the recognition dictionary is updated is obtained from the covariance matrix Σ before the update and the average face image m ′ after the recognition dictionary updated obtained by the above equation (8). Can be derived.

Further, from the above equation (3), the covariance matrix Σ before updating the recognition dictionary is obtained by converting the feature basis u _i (i = 1, 2,..., N) into a column vector as shown in the following equation (14). Can be derived from the diagonalization matrix A and the eigenvalue matrix Λ.

Therefore, the following formula (15) is finally established from the above formulas (9) and (14). In other words, the covariance matrix Σ ′ after the recognition dictionary is updated is the recognition matrix A, the eigenvalue matrix Λ, the newly input N + 1th normalized face image x _{N + 1,} and the recognition obtained by the above equation (8). It can be derived from the average face image m ′ after the dictionary update.

  As a result, the diagonalization matrix A ′ and the eigenvalue matrix Λ ′ that need to be registered after registering the recognition dictionary are obtained by replacing the covariance matrix Σ ′ after the recognition dictionary update obtained by the above equation (15) with the following (16): It can be derived by principal component analysis with an equation.

  Based on the above results, when the recognition dictionary is newly updated, it is not necessary to store and hold all the normalized face images used when the recognition dictionary was created in the past. It can be seen that the updated average face image m ′, eigenvalue matrix Λ ′, and diagonalization matrix A ′ can be derived using m, eigenvalue matrix Λ, and diagonalization matrix A.

  As mentioned above, although embodiment was described to this invention, this invention is not limited to the said embodiment. For example, in addition to a face image, other biological images such as fingerprints, irises, and retinas can be used as the biological image.

It is a flowchart which shows the basic dictionary creation process for new recognition, and a personal authentication process. It is a flowchart which shows the dictionary update process for recognition in one Embodiment of this invention. It is a flowchart which shows the creation process of the dictionary for new recognition. It is a flowchart which shows the process of performing personal authentication. It is a flowchart of the process of the dictionary update for recognition in one Embodiment of this invention. It is a flowchart of the process of the dictionary for the conventional recognition.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Pre-update recognition dictionary, 1 '... Post-update recognition dictionary, 2-1 to 2- (N + 1) ... Normalized biological (face) image

Claims (2)

In a recognition dictionary update device in biometrics authentication in which feature values extracted from biological images are registered,
Reading means for reading out the feature quantity registered in the recognition dictionary;
Feature amount calculating means for calculating a new feature amount from the newly input biological image and the feature amount read by the reading means;
An apparatus for updating a recognition dictionary in biometric authentication, comprising: a registration unit that registers a new feature amount calculated by the feature amount calculation unit in the recognition dictionary as an updated feature amount. The feature amount includes an average biometric image, an eigenvalue matrix in which eigenvalues obtained by performing principal component analysis on the average biometric image and a covariance matrix of each biometric image are arranged in a diagonal component, and eigenvectors corresponding thereto as column vectors. A diagonalization matrix arranged,
The feature amount calculating means includes:
Average biological image calculation means for calculating a new average biological image from the newly input biological image and the average biological image read by the reading means;
A covariance matrix that calculates a new covariance matrix from the newly input biological image, the average biological image calculated by the average biological image calculation means, and the eigenvalue matrix and the diagonalization matrix read by the reading means A calculation means;
A principal component analysis of the covariance matrix calculated by the covariance matrix calculation means, and an eigenvalue matrix / diagonalization matrix calculation means for calculating a new eigenvalue matrix and diagonalization matrix,
The registration means registers the new average biological image, the new eigenvalue matrix and the diagonalization matrix calculated by the average biological image calculation means and the eigenvalue matrix / diagonalization matrix calculation means in the recognition dictionary. The device for updating a recognition dictionary in biometric authentication according to claim 1.

Images