JP2007164498A - System and method for authentication, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an authentication technique enabling authentication with high accuracy by reducing the influence of the difference in the posture of a person, when authenticating the identity of the person. <P>SOLUTION: An authentication system acquires texture information (first texture information) and three-dimensional information in regard to the face of a person to be authenticated, and also texture information (second texture information) in regard to the face of a person to be compared. The first texture information is converted using an individual model (solid model) MDb. Then, by comparing the first texture information and the second texture information mutually from an identical viewpoint, the authentication system decides the identity between the person to be compared and the person to be authenticated. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an object authentication technique.

  In recent years, various services digitized by the development of network technology and the like have become widespread, and the need for non-face-to-face personal authentication technology that does not rely on people is increasing. Along with this, research on biometrics authentication technology (biometric authentication technology) for automatically identifying an individual based on a person's biometric features has been actively conducted. Face authentication technology, which is one of biometric authentication technologies, is a non-contact authentication method, and is expected to be applied in various fields such as security by a surveillance camera or image database search using a face as a key.

  As such an authentication technique, a registered image registered in advance and a photographed image obtained by photographing an authentication object are acquired only as plane information (“texture information” or “two-dimensional information”), respectively. It is determined whether or not the person in the registered image and the person in the captured image are the same person by comparing the plane information of the images (for example, see Non-Patent Document 1).

Dadet Pramady Hunt, Moto Kure and Masahiko Taniuchi, “Personal Identification from Input Face Images with Various Postures”, IEICE Transactions, D-II, August 1997, Vol. J80-D- II, No.8, pp.2232-2238

  However, the above-described comparison technique has a problem that high-accuracy authentication cannot be performed if the posture (orientation) of the person in both images to be compared is not the same.

  Accordingly, an object of the present invention is to provide an authentication technique capable of reducing the influence of a difference in the posture of a person and performing highly accurate authentication when authenticating the identity of the person.

  In order to solve the above-mentioned problem, the invention of claim 1 is an authentication system for performing an authentication operation on an authentication target person, the first acquisition means for acquiring texture information and three-dimensional information on the face of the authentication target person, A generating unit that generates a first three-dimensional model that is a three-dimensional model related to the face of the person to be authenticated based on the three-dimensional information, and a texture information that is related to the face of the authentication target person using the first three-dimensional model. Conversion means for converting 1 texture information; second acquisition means for acquiring second texture information that is texture information related to the face of the person to be compared; and the first texture information and the second texture information from the same viewpoint. And a determination unit that determines the identity of the comparison target person and the authentication target person by comparing.

  According to a second aspect of the present invention, in the authentication system according to the first aspect of the invention, the converting means includes each representative point of the first stereo model and each representative point of a standard model which is a standard stereo model of a human face. A plurality of representative points of the first three-dimensional model are projected on the same positions as the projection destination points of the corresponding representative points of the standard model, and the patches are configured by the plurality of representative points. The texture information corresponding to is assigned on the projection plane.

  According to a third aspect of the present invention, in the authentication system according to the second aspect of the invention, the conversion unit assigns the texture information of the first three-dimensional model to a predetermined projection plane arranged around the standard model. And

  According to a fourth aspect of the invention, in the authentication system according to the third aspect of the invention, the predetermined projection plane is a cylindrical plane.

  According to a fifth aspect of the present invention, in the authentication system according to the third aspect of the invention, the predetermined projection surface is a spherical surface.

  According to a sixth aspect of the present invention, in the authentication system according to the first or second aspect of the present invention, the conversion unit classifies the patches included in the first three-dimensional model into a plurality of groups, and each group after classification. Texture information corresponding to each patch in the group is assigned to the set projection plane.

  According to a seventh aspect of the present invention, in the authentication system according to any one of the first to sixth aspects of the present invention, the conversion means is configured such that the area of each patch on the first three-dimensional model is on the projection surface after the projection. Each patch of the first three-dimensional model is mapped onto the projection plane so as to have the same area as each patch.

  The invention according to claim 8 is the authentication system according to any one of claims 1 to 7, wherein the determination unit includes a feature amount extracted from the second texture information related to the face of the person to be compared, The identity between the comparison target person and the authentication target person is determined by comparing the feature amount extracted from the converted first texture information.

  The invention according to claim 9 is an authentication method for performing an authentication operation related to an authentication target person, and is a step of acquiring texture information and three-dimensional information related to the face of the authentication target person and a three-dimensional model related to the face of the authentication target person. A step of generating a first three-dimensional model based on the three-dimensional information, a step of converting first texture information that is texture information relating to the face of the person to be authenticated using the first three-dimensional model, and a comparison target The comparison target person and the authentication target person are identical by comparing the first texture information and the second texture information from the same viewpoint with the step of obtaining the second texture information that is the texture information related to the person's face And a step of determining the sex.

  The invention of claim 10 is a program for causing a computer to execute an authentication method for performing an authentication operation relating to an authentication target person, wherein the computer acquires texture information and three-dimensional information related to the face of the authentication target person. And a procedure for generating a first three-dimensional model that is a three-dimensional model related to the face of the person to be authenticated based on the three-dimensional information, and texture information about the face of the authentication target person using the first three-dimensional model. A procedure for converting certain first texture information, a procedure for obtaining second texture information that is texture information related to the face of the person to be compared, and the first texture information and the second texture information from the same viewpoint are compared. This is a program for executing the procedure for determining the identity of the comparison target person and the authentication target person. And wherein the door.

  According to the first to tenth aspects of the invention, the first texture information is converted using the first three-dimensional model, and by comparing the first texture information and the second texture information from the same viewpoint, Since the identity of the person to be authenticated and the person to be compared is determined, it is possible to reduce the influence based on the difference in the posture of the person and perform high-precision authentication.

  In particular, according to the invention described in claim 2, based on the correspondence between each representative point of the first stereo model and each representative point of the standard model, the plurality of representative points of the first stereo model are respectively the standard model. Since the corresponding representative point is projected at the same position as the projection destination point, and the texture information of the patch on the first three-dimensional model is allocated on a predetermined projection plane, the comparison process is performed after sufficiently preparing the comparison conditions. It is possible to perform authentication with higher accuracy.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. Hereinafter, an authentication system that performs an authentication operation using a face image will be described.

<A. Overview>
FIG. 1 is a conceptual diagram showing an outline of a face authentication system 1 according to this embodiment.

  As will be described later, the face authentication system 1 uses multiple cameras CA1 and CA2 (see FIG. 2) at the time of authentication to multi-view the face of a person HMb (a person who claims to be a person HMa) as an authentication target person ( Here, a plurality of (here, two) images (stereo images) G1 and G2 are captured. Specifically, an image G1 is captured by the camera CA1, and an image G2 is captured by the camera CA2. Then, the face authentication system 1 performs an authentication operation by comparing (or collating) information obtained from the two images G1 and G2 with information based on a face image registered in advance for the person HMa. It is determined whether or not the person HMb who is the person is the same person as the person HMa who is the person to be compared (registrant).

  Information obtained from the images G1 and G2 relating to the person HMb is broadly classified into shape information (information relating to the position and the like of each part) and texture information (surface information obtained as a set of surface luminance at each point and the like).

  From the images G1 and G2 related to the person HMb, texture information TB is obtained, and three-dimensional information SB is obtained as shape information. More specifically, first, texture information TB of the face of the person HMb is obtained from at least one of the image G1 and the image G2. Further, three-dimensional information SB including a set of three-dimensional positions of each point of the face of the person HMb and the like is obtained from the two images G1 and G2 viewed in stereo.

  For the person (comparison target) HMa, for example, the same process as the process for the person (authentication target) HMb is performed in advance. Specifically, the photographing operation of the comparison target person HMa is performed in advance, and the texture information TA and the three-dimensional information SA can be obtained from the obtained plurality of images.

  Basically, the texture information TA and the texture information TB obtained in this way are compared. In this embodiment, the three-dimensional information SA and the three-dimensional information SB are also compared. The comparison of the three-dimensional information SA and SB (in other words, “the comparison of shape information”) includes the comparison of the three-dimensional shape information and the comparison of the two-dimensional shape information.

  However, since the texture information TA, TB and the shape information SA, SB obtained in this way have an enormous amount of information, a feature selection operation (in other words, an information compression operation (described later)) is performed. It is preferable to carry out the comparison operation after a while. Also in this embodiment, an information compression operation is performed. According to this, it is possible to reduce the amount of information and perform a more efficient collation operation.

  Further, in this embodiment, the face authentication system 1 performs a process for making it less likely to be affected by a change in the posture of the person in the comparison between the texture information TA and the texture information TB. Specifically, a conversion process for projecting the texture on the three-dimensional model onto a common projection surface (two-dimensional image) is performed. Thereafter, the texture information TA after the conversion process and the texture information TB after the conversion process are compared.

  According to this face authentication system 1, since the identity of the comparison target person and the authentication target person is determined by comparing the texture information TA and the texture information TB from the same viewpoint, It is possible to reduce the influence based on it and to perform highly accurate authentication.

<B. Detailed configuration>
FIG. 2 is a configuration diagram showing the face authentication system 1 according to the embodiment of the present invention. As shown in FIG. 2, the face authentication system 1 includes a controller 10 and two image capturing cameras (hereinafter also simply referred to as “cameras”) CA1 and CA2. The camera CA1 and the camera CA2 are arranged so that the face of the person to be photographed (the person to be registered HMa or the person to be authenticated HMb) can be photographed from different positions. When the face image of the person to be photographed is photographed by the camera CA1 and the camera CA2, the appearance information of the person to be photographed, that is, the two face images G1 and G2 are transmitted to the controller 10 via the communication line. Note that the communication method of image data between each camera and the controller 10 is not limited to a wired method, and may be a wireless method.

  FIG. 3 is a diagram showing a configuration outline of the controller 10. As shown in FIG. 3, the controller 10 includes a CPU 2, a storage unit 3, a media drive 4, a display unit 5 such as a liquid crystal display, an input unit 6 such as a keyboard 6a and a mouse 6b that is a pointing device, It is composed of a general computer such as a personal computer provided with a communication unit 7 such as a network card. The storage unit 3 includes a plurality of storage media, specifically, a hard disk drive (HDD) 3a and a RAM (semiconductor memory) 3b capable of processing at higher speed than the HDD 3a. The media drive 4 can read information recorded in a portable recording medium 8 such as a CD-ROM, a DVD (Digital Versatile Disk), a flexible disk, or a memory card. Note that the information supplied to the controller 10 is not limited to the case of being supplied via the recording medium 8, and may be supplied via a network such as a LAN and the Internet.

  Next, the function of the controller 10 will be described with reference to FIGS.

  FIG. 4 is a block diagram illustrating various functional configurations of the controller 10, and FIG. 5 is a block diagram illustrating a detailed functional configuration of the personal authentication unit 14.

  4 and FIG. 5 conceptually shows functions realized by executing a predetermined software program (hereinafter also simply referred to as “program”) using various hardware such as a CPU in the controller 10. It is shown.

  As shown in FIG. 4, the controller 10 includes an image input unit 11, a face area search unit 12, a face part detection unit 13, a personal authentication unit 14, and an output unit 15.

  The image input unit 11 has a function of inputting two images for authentication photographed by the cameras CA1 and CA2 to the controller 10.

  The face area search unit 12 has a function of specifying a face area from the input face image.

  The face part detection unit 13 has a function of detecting the position of a characteristic part (for example, eyes, eyebrows, nose, mouth) of the face from the specified face region.

  The personal authentication unit 14 is configured mainly for face authentication, and has a function of authenticating each individual with a face image. Details of the personal authentication unit 14 will be described below.

  The output unit 15 has a function of outputting the authentication result obtained by the personal authentication unit.

Next, a detailed configuration of the personal authentication unit 14 will be described with reference to FIG.
As shown in FIG. 5, the personal authentication unit 14 includes a three-dimensional reconstruction unit 21, an optimization unit 22, a correction unit 23, a feature extraction unit 24, an information compression unit 25, and a comparison unit 26.

  The three-dimensional reconstruction unit 21 has a function of calculating the three-dimensional coordinates of each part from the coordinates of the characteristic part of the face obtained from the input image. This three-dimensional coordinate calculation function is realized using camera information stored in the camera parameter storage unit 27.

  The optimization unit 22 uses the calculated three-dimensional coordinates to determine a unique three-dimensional model (also referred to as “standard three-dimensional model” or “standard model”) of the face stored in the three-dimensional model database 28. It has a function of generating an “individual model” related to a person.

  The correction unit 23 has a function of correcting the generated individual model, and a function (conversion function) for re-attaching the texture pasted on the individual model (three-dimensional model) on the two-dimensional image (projection plane). have.

  By each of these processing units 21, 22, and 23, information regarding the authentication target person HMb and the registrant HMa is standardized (normalized) and converted into a state that can be easily compared with each other. In addition, the individual model created by the function of each processing unit is formed as including both three-dimensional information and two-dimensional information related to a person's face. “3D information” is information related to a three-dimensional configuration such as a 3D coordinate value of a representative point, and “2D information” is surface information (texture information) and / or planar position information of a representative point. This is information related to a planar configuration such as (two-dimensional shape information). The individual model is also expressed as including both shape information (three-dimensional shape information and two-dimensional shape information) and texture information related to a person's face.

  The feature extraction unit 24 has a feature extraction function for extracting three-dimensional information and two-dimensional information from the individual models created in the processing units 21, 22, and 23.

  The information compression unit 25 converts the three-dimensional information and the two-dimensional information used for face authentication by converting the three-dimensional information and the two-dimensional information extracted by the feature extraction unit 24 into appropriate face feature amounts for face authentication, respectively. Each has a function of compressing information. This information compression function is realized by using information stored in the basis vector database 29.

  The comparison unit 26 calculates the similarity between the facial feature amount of the registrant (comparison subject) HMa registered in the person database 30 in advance and the facial feature amount of the authentication subject person HMb obtained by each of the functional units. , Has a function to perform face authentication.

<C. Operation>
<C1. Outline of operation>
Hereinafter, the face authentication operation of the controller 10 described above will be described in more detail. Specifically, as described above, the person to be authenticated HMb is a registered person using information on the face image of the person to be authenticated HMb photographed by the cameras CA1 and CA2 and the face image of the person HMa registered in advance. The case of authenticating whether the person is the same as HMa (when performing face authentication) will be described. Here, as an example, three-dimensional information using three-dimensional shape information measured according to the principle of triangulation using images from the cameras CA1 and CA2 is used, and texture (luminance) information is used as two-dimensional information.

  FIG. 6 is a flowchart showing the overall operation (authentication operation) of the controller 10, and FIG. 7 is a flowchart showing the registration operation (step SP30) performed prior to the processing of FIG. Further, FIG. 8 is a diagram exemplifying feature points of characteristic parts in the face image, and FIG. 9 shows how three-dimensional coordinates are calculated from the feature points in the two-dimensional image using the principle of triangulation. It is a schematic diagram shown. A point Q20 in the images G1 and G2 in FIG. 9 corresponds to the right end of the mouth in FIG.

  As illustrated in FIG. 6, the controller 10 acquires a facial feature amount related to the authentication target person based on an image obtained by photographing the face of the authentication target person in the processes from step SP1 to step SP8. Thereafter, the controller 10 further performs steps SP9 and SP10 to realize face authentication. In practice, the processing shown in FIG. 7 (steps SP31 to SP39) is performed prior to the processing shown in steps SP1 to SP10 in FIG. 6, but the processing shown in FIG.

<C2. Collecting information on HMb subject to authentication>
<Individual model creation processing>
First, in step SP1, a face image of a predetermined person (person to be authenticated) HMb photographed by the cameras CA1 and CA2 is input to the controller 10 via a communication line. The cameras CA1 and CA2 that capture a face image are each configured by a general imaging device that can capture a two-dimensional image. Further, camera parameters Bi (i = 1 ·· N) indicating the positions and orientations of the respective cameras CAi are known and stored in advance in the camera parameter storage unit 27 (FIG. 5). Here, N indicates the number of cameras. In the present embodiment, the case of N = 2 is illustrated, but N ≧ 3 may be used (three or more cameras may be used). The camera parameter Bi will be described later.

  Next, in step SP2, a region where a face exists is detected in each of the two images G1 and G2 input from the cameras CA1 and CA2. As the face area detection technique, for example, a technique of detecting a face area from each of the two images G1 and G2 by template matching using a standard face image prepared in advance can be employed.

  Next, in step SP3, the position of the characteristic part of the face is detected from the face area image detected in step SP2. For example, eyes, eyebrows, nose or mouth may be considered as characteristic parts of the face, and in step SP3, the coordinates of the characteristic points Q1 to Q23 of the respective parts as shown in FIG. 8 are calculated. . The feature part can be detected by, for example, template matching performed using a standard template of the feature part. The calculated feature point coordinates are expressed as coordinates on the images G1 and G2 input from the camera. For example, with respect to the feature point Q20 corresponding to the right end of the mouth in FIG. 8, coordinate values in each of the two images G1 and G2 are obtained as shown in FIG. Specifically, the coordinates (x1, y1) of the feature point Q20 on the image G1 are calculated using the upper left end point of the image G1 as the origin O. Similarly, in the image G2, the coordinates (x2, y2) of the feature point Q20 on the image G2 are calculated.

  In addition, the luminance value of each pixel in a region having each feature point as a vertex in the input image is acquired as information (hereinafter also referred to as “texture information”) of the region. The texture information in each region is pasted on the individual model in step SP5 and the like described later. In the present embodiment, since two images are input, the average luminance value in the corresponding pixel in the corresponding region of each image is used as the texture information of the region.

In the next step SP4, a three-dimensional reconstruction process is performed. Specifically, the two-dimensional coordinates Ui ^(j) in each image Gi (i = 1,..., N) of each feature point Qj detected in step SP3 and the camera parameters of the camera that captured each image Gi. Based on Bi, the three-dimensional coordinates M ^(j) (j = 1... M ^{) of} each feature point Qj are calculated. In short, the three-dimensional position of each feature point is calculated based on the principle of triangulation. Note that m indicates the number of feature points.

Hereinafter, the calculation of the three-dimensional coordinate M ^(j) will be specifically described.

The relationship between the three-dimensional coordinates M ^(j) of each feature point Qj, the two-dimensional coordinates Ui ^{(j) of} each feature point Qj, and the camera parameter Bi is expressed as in Expression (1).

  Note that μi is a parameter that indicates a change in the scale. The camera parameter matrix Bi is a value unique to each camera obtained by photographing an object with known three-dimensional coordinates in advance, and is represented by a 3 × 4 projection matrix.

For example, as a specific example of calculating the three-dimensional coordinates using the above equation (1), consider the case of calculating the three-dimensional coordinates M feature points Q20 ⁽²⁰⁾ with reference to FIG. Expression (2) shows the relationship between the coordinates (x1, y1) of the feature point Q20 on the image G1 and the three-dimensional coordinates (x, y, z) when the feature point Q20 is represented in a three-dimensional space. Similarly, Expression (3) represents the relationship between the coordinates (x2, y2) of the feature point Q20 on the image G2 and the three-dimensional coordinates (x, y, z) when the feature point Q20 is represented in a three-dimensional space. Show.

The unknowns in the above formulas (2) and (3) are a total of five of the two parameters μ1, μ2 and the three component values x, y, z of the three-dimensional coordinate M ⁽²⁰⁾ . On the other hand, since the number of equations included in the equations (2) and (3) is 6, each unknown, that is, the three-dimensional coordinates (x, y, z) of the feature point Q20 can be calculated. Similarly, three-dimensional coordinates M ^(j) for all feature points Qj can be acquired.

In the next step SP5, model fitting is performed. This “model fitting” is performed by deforming a “standard model of a face”, which is a general (standard) model of a face prepared in advance, using information related to the authentication target person. This is a process for generating an “individual model” in which input information related to a face is reflected. Specifically, a process of changing the three-dimensional information of the standard model using the calculated three-dimensional coordinates M ^(j) , and a process of changing the two-dimensional information of the standard model using the texture information (the texture concerned) Information is pasted on the surface of the individual model).

  FIG. 10 shows a standard model of a three-dimensional face.

  As shown in FIG. 10, the standard model of a face is composed of vertex data and polygon data, and is stored in the storage unit 3 or the like as a three-dimensional model database 28 (FIG. 5). The vertex data is a set of coordinates of the vertices (hereinafter also referred to as “standard control points”) COj of the characteristic parts in the standard model, and has a one-to-one correspondence with the three-dimensional coordinates of the characteristic points Qj calculated in step SP4. is doing. The polygon data is obtained by dividing the surface of the standard model into minute polygons (for example, triangles) and expressing the polygons as numerical data. FIG. 10 illustrates the case where the vertexes of each polygon are also configured by intermediate points other than the standard control point COj, and the coordinates of the intermediate point are appropriately complemented using the coordinate values of the standard control point COj. Can be obtained by:

  Here, the model fitting which comprises an individual model from a standard model is explained in full detail.

First, the vertex (standard control point COj) of each feature part of the standard model is moved to each feature point calculated in step SP4. Specifically, the three-dimensional coordinate value of each feature point Qj is substituted as the three-dimensional coordinate value of the corresponding standard control point COj, and the moved standard control point (hereinafter also referred to as “individual control point”) Cj obtain. As a result, the standard model can be transformed into an individual model represented by the three-dimensional coordinates M ^(j) . Note that the coordinates of intermediate points other than the individual control points Cj in the individual model can be obtained by an appropriate interpolation method using the coordinate values of the individual control points Cj.

  In addition, the scale, inclination, and position of the individual model based on the standard model used in step SP6 described later can be obtained from the amount of movement of each vertex due to this deformation (movement). Specifically, the position change of the individual model with respect to the standard model can be obtained from the deviation amount between the predetermined reference position in the standard model and the corresponding reference position in the individual model after deformation. In addition, a change in the inclination of the individual model with respect to the standard model is determined by the amount of deviation between the reference vector connecting the two predetermined points in the standard model and the reference vector connecting the corresponding two points in the individual model after deformation. And a change in scale can be determined. For example, the position of the individual model is obtained by comparing the coordinates of the midpoint QM of the right-eye feature point Q1 and the left-eye feature point Q2 with the coordinates of the point corresponding to the midpoint QM in the standard model. Furthermore, the scale and inclination of the individual model can be calculated by comparing the midpoint QM with other feature points.

  The following equation (4) represents a conversion parameter (vector) vt that represents the correspondence between the standard model and the individual model. As shown in the equation (4), the conversion parameter (vector) vt includes a scale conversion index sz of both, a conversion parameter (tx, ty, tz) indicating translational displacement in the orthogonal three-axis directions, and a rotational displacement (inclination). ) Representing a conversion parameter (φ, θ, ψ).

As described above, the process of changing the three-dimensional information of the standard model using the three-dimensional coordinates M ^(j) regarding the person to be authenticated is performed.

  Thereafter, processing for changing the two-dimensional information of the standard model using the texture information is also performed. Specifically, the texture information of each region in the input images G1 and G2 is pasted (mapped) to the corresponding region (polygon) on the three-dimensional individual model. Each region (polygon) to which texture information is pasted on a three-dimensional model (individual model or the like) is also referred to as a “patch”.

  The model fitting process (step SP5) is performed as described above. Thereby, the information related to the authentication target person HMb is generated as an “individual model” (MDb) including both the three-dimensional information and the two-dimensional information related to the authentication target person HMb.

<Conversion processing (outline)>
In the next step SP6, the individual model is corrected based on the standard model. In this step, position (alignment) correction relating to three-dimensional information and conversion processing for projecting texture information on the three-dimensional model onto a two-dimensional plane are executed.

  The former alignment (face orientation) correction is a correction process for the position and orientation of the three-dimensional information. The alignment correction is performed based on the scale, inclination, and position of the individual model when the standard model is used as a reference. More specifically, by converting the coordinates of the individual model using the conversion parameter vt (see Expression (4)) indicating the relationship between the standard model and the individual model when the standard model is used as a reference, A three-dimensional face model having the same posture can be created. That is, this alignment correction can properly normalize the three-dimensional information related to the authentication subject.

  The latter conversion process is a process for projecting texture information on a three-dimensional model (individual model) onto a two-dimensional plane. As will be described later, an image (projected image) projected on a two-dimensional plane by this conversion processing is obtained for both the authentication target person and the comparison target person, and the two projected images are compared (specifically, both projections An authentication operation is performed (by comparing feature quantities extracted from the images).

<Move process>
This conversion process involves a process of moving (warping) a feature point (PB or the like) on an individual model (for example, MDb) to a corresponding point (PT) on the projection plane PL (see FIG. 11). As a result, the feature points in the three-dimensional space are converted into feature points on the two-dimensional plane.

  Here, the case where the point of movement of the corresponding feature point in the standard model (“target point”) is used as the point of movement of the feature point in the individual model (also referred to as “target point”) is illustrated.

  For example, the feature points PB0 to PB5 (see FIG. 11) on the individual model (three-dimensional model) MDb are moved (also referred to as projections) to points PT0 to PT5 (described below) on the projection plane PL, respectively. The Specifically, the feature point PB0 is projected onto the corresponding point PT0, the feature point PB1 is projected onto the corresponding point PT1, and the feature point PB2 is projected onto the corresponding point PT2. Similarly, the other feature points PB3, PB4, PB5,... Are projected onto the feature points PT3, PT4, PT5,.

  Here, the points PT0 to PT5 are points on the projection plane PL on which the feature points PS0 to PS5 on the “standard model” MDs are projected by the method MT1 or the method MT2 (described below) (destination points or thru points). Also referred to as “target point”). Specifically, the point PT0 is the target point of the feature point PS0, the point PT1 is the target point of the feature point PS1, and the point PT2 is the target point of the feature point PS2. The same applies to the other points PT3, PT4, PT5,.

  By moving the feature points PB0 to PB6,... On the individual model MDb to such target points PT0 to PT6,..., The feature points on the individual model MDb are moved on the standard model MDs. A two-dimensional image associated with the same position as the feature point is generated.

  Specifically, first, target points PT0 to PT5,... When projecting the characteristic points PS1 to PS5,... Of the standard model MDs on the projection plane PL are determined by the method MT1 or the method MT2. deep. Next, the correspondence relationship between the feature points PB0 to PB5,... Of the individual model MDb and the feature points PS0 to PS5,. Then, based on this correspondence relationship, the feature points PB0 to PB5,... Of the individual model MDb are converted into the projection points PT0 to PT5,... Of the corresponding feature points PS0 to PS5,. Project to the same position. For example, since the feature point PS1 of the standard model MDs corresponds to the feature point PB1 of the individual model MDb, the target point PT1 related to the feature point PS1 of the standard model MDs is determined as the target point related to the feature point PB1 of the individual model MDb. . The same applies to other feature points.

Here, each method MT1, MT2 is respectively
(1) Method MT1: A method of setting a common projection plane (and a common coordinate system) for all patches and projecting the vertices of each patch onto the common projection plane,
(2) Method MT2: A method of classifying all patches into several groups and projecting the vertices of each patch belonging to the group onto individual projection planes set for each group after classification.
It is.

  Hereinafter, these two types of methods MT1 and MT2 will be described in order.

  First, (1) method MT1 will be described.

  As the common projection plane and the common coordinate system in the technique MT1, for example, a cylindrical plane PL11 (see FIG. 12) and a cylindrical coordinate system can be employed, respectively. Specifically, as shown in FIG. 12, the vertices of the patches of the standard model MDs are projected onto the cylindrical surface PL11 arranged around the standard model MDs, and the projected vertices are used as the above “target points”. Can be determined. For example, the feature points PS1 to PS5 of the standard model MDs are respectively projected onto the projection points PT0 to PT5 (see FIG. 11) on the projection plane PL (PL11). At this time, projection points corresponding to the feature points PS1 to PS5 are determined as target points (corresponding target points) PT0 to PT5.

  Note that the projection processing for projecting the feature points of the individual model to the points on the projection plane PL (PL11) determined by the method MT1 projects the vertices of each patch of the individual model onto the projection plane arranged around the individual model. It corresponds to the processing. In addition, here, the cylindrical plane PL11 and the cylindrical plane coordinate system are illustrated as the common projection plane (and the common coordinate system) in the technique MT1, but the present invention is not limited thereto. For example, the common projection plane and the common plane are common. As the coordinate system, a spherical PL 12 (see FIG. 13), a spherical coordinate system, or the like may be employed.

  Next, (2) Method MT2 will be described.

  As described above, the technique MT2 is a technique in which all patches are classified into several groups, and the vertices of the patches belonging to the group are projected onto individual projection planes set for each group after classification. .

  FIG. 14 is a diagram illustrating a classification example of each patch. In FIG. 14, the plurality of patches forming the surface of the standard model MDs are a group GR1 including a patch near the right eye, a group GR2 including a patch near the left eye, a group GR3 including a patch near the nose, and a vicinity of the mouth. A state of being classified into the group GR4 and the like including the patches of is shown. For example, the patches of the standard model MDs can be classified into such groups GR (GR1 to GR4, etc.).

  In the method MT2, an individual projection plane is set for each group.

  FIG. 15 is a diagram showing how the characteristic points PS0 to PS5 of the standard model MDs are projected onto the projection points PT0 to PT5 on the projection plane PL15, respectively. In FIG. 15, the feature points PS0 to PS5 of the standard model MDs are respectively projected onto the projection points PT0 to PT5 on the projection plane PL15, and the projection points corresponding to the feature points PS0 to PS5 are target points ( Corresponding destination point: determined as PT0 to PT5. FIG. 15 also shows how each patch belonging to a certain group of standard models MDs is projected onto a certain projection plane PL15. Specifically, a state is shown in which patches C1 to C5 composed of feature points PS0 to PS5 on the standard model MDs are projected onto the projection plane PL15.

  The projection plane PL15 is a plane whose vector is a vector (also referred to as a synthesized vector) obtained by synthesizing (outward) normal vectors of the patches C1 to C5 near the point of interest (here, the feature point PS0). It is determined as This combined vector is also expressed as an average vector of (outward) normal vectors of the plurality of patches C1 to C5. According to this, an image in which the patch near the point of interest is viewed from directly above (more precisely, the direction of the combined vector is the line-of-sight direction) is generated as a projection image.

  The same processing as described above is performed for patches belonging to another group in the standard model. Thereby, the corresponding point obtained by projecting each feature point belonging to each of the plurality of groups in the standard model onto the projection plane unique to the belonging group is determined as the corresponding target point.

  As described above, the target point related to the feature point of the standard model is determined. As described above, the target point related to the feature point of the standard model is determined as the target point related to the feature point of the individual model.

<Paste processing>
In this conversion processing, each feature point of the “individual model” is moved to each “target point” on the projection plane by the warp processing as described above, and a patch (triangle) surrounded by the feature points on the individual model. Area) is assigned to a patch (triangular area) on the projection plane PL. That is, the process of pasting the texture information of each patch on the individual model onto the projection plane PL is also performed.

  Specifically, the pixel in each patch on the individual model before conversion (also referred to as “pre-conversion patch”) corresponds to the corresponding position in each patch on the projection plane after conversion (also referred to as “post-conversion patch”). Consider a transformation that warps (moves) In this conversion, for example, the two-dimensional position (x, y) in the pre-conversion patch surrounded by three vertices is converted into coordinate values (u, v). When the third row of the matrix H is (0, 0, 1), the transformation represented by the matrix H is an affine transformation.

When the pixel value of the pixel at each position (u, v) after conversion is actually obtained, first, H ^-1 that is an inverse matrix of this matrix H is obtained, and each (u, v) in the post-conversion patch is obtained. The corresponding position (x, y) is obtained by equation (6).

  Then, the pixel value of the pixel at the position (x, y) before conversion obtained by the equation (6) or the pixel value to be interpolated (such as linear interpolation) by the pixel values of a plurality of pixels existing in the vicinity of the position. The pixel value of the pixel at the position (u, v) in the post-conversion patch is determined.

  Thus, the pixel value of each pixel in the post-conversion patch, that is, the texture information projected on the two-dimensional plane is acquired.

  Note that, when the method MT2 is used, the same processing is performed on a plurality of groups, thereby generating a projection image on the projection plane PL set for each group. In this case, each projection image has an individual coordinate system.

<Fine adjustment in conversion processing>
Basically, conversion processing can be performed by moving each feature point on the standard model MDs to a destination point (target point) obtained by the methods MT1, MT2, etc. as described above. Is possible.

  However, in the above conversion processing, the area of each patch on the standard model MDs and conversion are converted because the normal line of each patch having a three-dimensional shape and the normal line of the projection plane PL are different from each other. The area of each patch on the projection plane PL projected by the processing may be different (see FIG. 15). When comparing the texture information of patches, it is preferable that the area of the patch before conversion on the standard model MDs and the area of the patch after conversion on the projection plane PL are the same. Further, it is preferable that the area ratio before and after the conversion of the patch does not differ depending on the patch.

  Therefore, here, for each patch, a conversion process is performed in which the area of the pre-conversion patch and the area of the post-conversion patch are made closer. In other words, the point of each destination is determined so that the area of each patch on the three-dimensional model and the area of each patch on the projection surface projected by the conversion process are closer to each other. And According to this, the patch on each standard model is converted into a state having approximately the same area as the surface area on the three-dimensional model. Therefore, the comparison conditions in the texture comparison of the individual models can be further adjusted.

  Below, the case where the technique which concerns on this modification is applied to method MT2 is demonstrated.

  Specifically, the evaluation function FD of Expression (7) is calculated while moving the patch vertices PT1 to PT5, and the moved patch vertices PT1 to PT5 that minimize the evaluation function FD are obtained. As a minimization method (optimization method), an appropriate method such as a steepest descent method may be used.

  Where Sm (p) represents the area of the p-th patch on the standard model, Sr (p) represents the area of the p-th patch on the projection plane, Np represents the number of vertices, and Θ (q) Represents an angle formed by a coordinate axis AXq (described later) and a coordinate axis CXq on the projection plane, and β is an appropriate constant.

  By minimizing the first term on the right side of Equation (7), a conversion process is performed to bring the area of the pre-conversion patch close to the area of the post-conversion patch. In addition, the second term on the right side of Equation (7) is a term having an effect of constraining the directionality of area optimization.

  FIGS. 16, 17, and 18 are diagrams for illustrating how texture information on a model is projected on a two-dimensional plane.

  FIG. 16 is a diagram illustrating coordinate axes set at the center of interest of the standard model. When the right eye of the standard model is the attention point, the coordinate axis is set along the surface shape of the standard model around the attention point. The coordinate axes AX1 to AX4 are shown. Here, although the center part of the right eye is shown as the attention point, the point Q7 in FIG.

  The coordinate axis AX1 is a line of intersection between the plane PL1 and the standard model patch (specifically, the right point patch C2 of the target point (see FIG. 15)), and the coordinate axis AX3 is the plane PL1 and the standard model patch (detailed attention). This is a line of intersection with the left patch C5) of the point. The plane PL1 is a vertical plane that passes through both the attention point and the center point of the standard model MDs.

  The coordinate axis AX2 is an intersection line between the plane PL2 and the standard model patch (specifically, the upper patch C1 of the target point), and the coordinate axis AX4 is the plane PL2 and the standard model patch (specifically, below the target point). This is the intersection line with the side patch C3). The plane PL2 is a horizontal plane passing through the attention point.

  FIG. 17 is a diagram showing the relationship between the coordinate axes CX1 to CX4 of the coordinate system on the projection plane PL15 and the coordinate axes AX1 to AX4 before the optimization process. FIG. 18 is a diagram illustrating a state after the optimization process. In FIG. 15, coordinate axes AX1 to AX4 before the optimization process are shown on the projection plane PL15.

  Here, the direction of the composite vector regarding the target group is different from the direction AR1, more specifically, the standard model is slightly inclined from the upper right side with respect to the direction AR1 along the intersection line of both planes PL1 and PL2 in FIG. A situation is assumed in which the vicinity of the target group is viewed. In this case, as shown in FIG. 17, on the projection plane PL15, the coordinate axes AX1 to AX4 on the polygon surface are in a state of being shifted from the coordinate axes CX1 to CX4, respectively.

  According to the second term on the right side of Equation (7), this deviation can be suppressed, and the horizontal and vertical lines on the three-dimensional model can be more accurately aligned. In addition, the area can be made identical under such restrictions.

  Further, if the projection state of each patch is changed to the state as shown in FIG. 18, the areas of the patches C1 to C5 on the projection plane PL are respectively the areas of the patches C1 to C5 on the standard model MDs. Almost equal.

  As described above, in step SP6, the three-dimensional position information related to the authentication target person HMb is generated as the three-dimensional position information of each feature point in the individual model MDb, and the texture information related to the authentication target person HMb is as described above. It is generated as a projection image on the projection plane.

<Extraction processing, etc.>
In the next step SP7 (FIG. 6), three-dimensional shape information (three-dimensional information) and texture information (two-dimensional information) are extracted as information representing the characteristics of the person to be authenticated HMb.

As the three-dimensional information, three-dimensional coordinate vectors of m individual control points Cj in the individual model MDb are extracted. Specifically, as shown in Expression (8), a vector h whose elements are three-dimensional coordinates (Xj, Yj, Zj) of m individual control points Cj (j = 1,..., M). ^S is extracted as three-dimensional information (three-dimensional shape information).

  As the two-dimensional information, texture (luminance) information (hereinafter referred to as “local 2”) included in a characteristic part of the face that is important information for personal authentication, that is, a patch or a group of patches (local region) in the vicinity of an individual control point. Also referred to as “dimensional information”). Here, information after the conversion process (step SP6) (that is, information based on the projection image) is used as the texture information (local two-dimensional information).

  The local two-dimensional information is configured as luminance information of each pixel included in each local area, for example. Each local area corresponds to a group area composed of a plurality of patches or an area composed of a single patch.

The local two-dimensional information h ^(k) (k = 1,..., L; L is the number of local regions) is represented by n as the number of pixels in the local region and BR1 as the luminance value of each pixel. , BRn, it is expressed in a vector format as shown in Equation (9). In addition, information ^obtained by collecting local two-dimensional information h ^(k) for L local regions is also expressed as comprehensive two-dimensional information.

  As described above, in step SP7, the three-dimensional shape information (three-dimensional information) SB (FIG. 1) and the texture information (two-dimensional information) TB are extracted as information representing the characteristics of the individual model.

  The extracted information is used for an authentication operation (steps SP9 and SP10) described later. In the authentication operation, the authentication operation may be performed using the information obtained by Expression (9) as it is. However, in this case, when the number of pixels in the local area is large, the amount of calculation in the authentication operation is very large. Will become bigger. Therefore, in this embodiment, the authentication operation is performed by reducing the amount of calculation and performing the authentication operation efficiently, and further compressing the information obtained by Expression (9) and using the compressed information. And

  For this reason, in the next step SP8, the information compression processing described below for converting the information extracted in step SP7 into a state suitable for authentication is performed.

The information compression processing is performed using the same method for each of the three-dimensional shape information h ^S and each local two-dimensional information h ^(k) . Here, for the local two-dimensional information h ^(k) , A case where information compression processing is performed will be described in detail.

The local two-dimensional information h ^(k) is calculated in advance by KL expansion of the average information (vector) have ^{(k) of the} local area acquired in advance from a plurality of sample face images and the plurality of sample face images. Using the matrix P ^(k) (described below) expressed by the set of eigenvectors of the local region, it can be expressed in a basis-decomposed format as shown in Equation (10). As a result, the local two-dimensional face information amount (vector) c ^(k) is acquired as compressed information about the local two-dimensional information h ^(k) .

As described above, the matrix P ^(k) in the equation (10 ⁾ is calculated from a plurality of sample face images. Specifically, the matrix P ^(k) is obtained as a set of several eigenvectors (base vectors) having large eigenvalues among a plurality of eigenvectors obtained by KL expansion of a plurality of sample face images. These basis vectors are stored in the basis vector database 29. By expressing the face image using an eigenvector indicating a larger feature of the face image as a base vector, it is possible to efficiently express the feature of the face image.

For example, let us consider a case where local two-dimensional information h ^(GR) of a local region of a certain group GR is expressed in a basis decomposed form. Assuming that the set P of eigenvectors of the local region is expressed as P = (P1, P2, P3) by three eigenvectors P1, P2, and P3, the local two-dimensional information h ^(GR) The average information have ^(GR) and the set of eigenvectors P1, P2, and P3 are used to express the equation (11). The average information have ^(GR) is a vector obtained by averaging a plurality of local two-dimensional information (vectors) for various sample face images for each corresponding element. In addition, what is necessary is just to use the some standard face image which has moderate dispersion | variation for a some sample face image.

Further, the above equation (11) indicates that the original local two-dimensional information can be reproduced by the face information amount c ^(GR) = (c1, c2, c3) ^T. That is, it can be said that the face information amount c ^(GR) is information obtained by compressing the local two-dimensional information h ^(GR) of the local region including the group GR.

The local two-dimensional face information amount c ^(GR) acquired as described above may be used for the authentication operation as it is, but in this embodiment, further information compression is performed. Specifically, a process of converting the feature space represented by the local two-dimensional face information amount c ^(GR) into a partial space that increases separation between individuals is further performed. More specifically, a transformation matrix A that reduces the local two-dimensional face information amount c ^(GR) of the vector size f to the local two-dimensional feature amount d ^(GR) of the vector size g as expressed in Expression (12) is considered. . Thereby, the feature space represented by the local two-dimensional face information amount c ^(GR) can be converted into the partial space represented by the local two-dimensional feature amount d ^(GR) , and the difference in information between individuals is remarkable. become.

  Here, the transformation matrix A is a matrix having a size of f × g. Using a multiple discriminant analysis (MDA) method, the transformation matrix A is determined by selecting g principal components having a large ratio (F ratio) between intra-class variance and inter-class variance from the feature space. Can do.

In addition, by executing the same processing as the information compression processing performed on the local two-dimensional information h ^(GR) described above for all other local regions, the local two-dimensional face feature amount d ^(k) for each local region is also ^{obtained. )} can be acquired. Further, by applying the same method to the three-dimensional shape information h ^S , the three-dimensional face feature value d ^S can be acquired.

The face feature amount d obtained by combining the three-dimensional face feature amount d ^S and the local two-dimensional face feature amount d ^(k) acquired through step SP8 can be expressed as a formula (13) in a vector format. .

  In the steps SP1 to SP8 described above, the facial feature amount d (Bd) of the subject person is acquired from the face images G1 and G2 of the subject person HMb to be inputted. The face feature amount d (Bd) of the person to be authenticated HMb is used in the processing of steps SP9 and SP10 together with the feature amount d (Ad) related to the registrant HMa's face image described below.

<C3. Collecting information on registrant HMa>
Next, with reference to FIG. 7, the operation (step SP30) for obtaining the feature quantity d (Ad) related to the face image at the time of registration will be described. The process of step SP30 includes the processes of steps SP31 to SP39. Each process of steps SP31 to SP38 is the same as the corresponding process of steps SP1 to SP8.

  First, information regarding the registrant (comparator) HMa is generated as an “individual model” MDa including both the three-dimensional information and the two-dimensional information regarding the registrant HMa through the processing of steps SP31 to SP35.

  In step SP36, the same process as the process in step SP6 is performed. The position (alignment) correction related to the three-dimensional information is performed, and the texture information pasted on the patch on the individual model is converted into a state projected onto a predetermined projection plane.

  In the next step SP37 (FIG. 6), three-dimensional shape information and texture information are extracted as information representing the characteristics of the registrant HMa. In step SP37, the same processing as in step SP7 is performed. Here, information on a state projected onto a predetermined projection plane is used as the texture information.

  Further, in step SP38, information compression processing for converting the information extracted in step SP37 into a state suitable for authentication is performed. In step SP38, the same processing as in step SP8 is performed.

As a result, a facial feature value d (Ad) obtained by combining the three-dimensional facial feature value d ^S and the local two-dimensional facial feature value d ^(k) obtained in step SP38 is obtained.

  In the next step SP39, the acquired face feature amount d (Ad) is registered and stored in the person database 30 in the controller 10 as information on the registrant HMa.

  As described above, the process of step SP30 (FIG. 7) is performed.

<C4. Matching process between registrant HMa and authentication target HMb>
Referring to FIG. 6 again, the processing after step SP9 will be described.

  In steps SP9 and SP10, face authentication is performed using the two face feature values Ad and Bd described above.

Specifically, the total similarity Re, which is the similarity between the authentication target person (authentication target object) and the comparison target person (comparison target object), is calculated (step SP9), and then authentication based on the total similarity degree Re A comparison operation or the like (step SP10) between the subject and the comparison subject is performed. Overall similarity Re is, the 3-dimensional similarity Re ^S calculated from 3-dimensional face feature amount d ^S, a local two-dimensional similarity calculated from the local 2-dimensional face feature amount d ^(k) Re ^(k) In addition, appropriate weighting factors that define the weights of the three-dimensional similarity Re ^S and the local two-dimensional similarity Re ^(k) (hereinafter also simply referred to as “weighting factors”) WT, WS (see Expression (14)) Is calculated using

In step SP9, the face feature amount (comparison feature amount) d (Ad) of the comparison target registered in advance in the person database 30 and the face feature amount d of the authentication target calculated through steps SP1 to SP8. Similarity with (Bd) is evaluated. Specifically, it is similar between the registered facial feature quantities (comparison feature quantities) (Re ^SM and Re ^{(k) M} ) and the authentication target person's facial feature quantities (Re ^SI and Re ^{(k) I} ). The degree calculation is executed, and the three-dimensional similarity Re ^S and the local two-dimensional similarity Re ^(k) are calculated.

Now, the three-dimensional similarity Re ^S between the authentication subject and the comparison subject is obtained by obtaining the Euclidean distance Re ^S between the corresponding vectors as shown in the equation (15).

Further, the local two-dimensional similarity Re ^(k) is obtained by obtaining the Euclidean distance Re ^(k) for each vector component of the feature quantities in the corresponding local regions as shown in the equation (16). .

Then, as shown in Expression (17), the three-dimensional similarity Re ^S and the local two-dimensional similarity Re ^(k) are synthesized using predetermined weighting factors WT and WS, and the person to be authenticated ( The overall similarity Re, which is the similarity between the authentication object) and the comparison person (comparison object), can be acquired.

  Next, in step SP10, authentication determination is performed based on the total similarity Re. Specifically, the degree of similarity Re between the facial feature quantity of the authentication target person HMb and the facial feature quantity of the specific registrant (comparison target person) HMa is compared with a certain threshold TH1, thereby comparing with the authentication target person HMb. Identity with the subject person HMa is determined. Specifically, when the similarity Re is smaller than a certain threshold TH1, it is determined that the person to be authenticated is the same person as the person to be compared.

  As described above, according to the authentication system of the above embodiment, each individual model is based on the correspondence between each representative point (feature point, etc.) of the individual model (MDb, etc.) and each representative point of the standard model MDs. The representative point is projected at the same position as the projection destination point of the corresponding representative point (corresponding feature point etc.) of the standard model (see FIG. 11 etc.). Then, the patches on the individual models MDa and MDb are similarly allocated on the common projection plane. This corresponds to the texture information regarding the individual model MDa and the texture information regarding the individual model MDb being compared in a state viewed from the same viewpoint. Therefore, in determining the identity between the authentication target person and the comparison target person, it is possible to reduce the influence based on the difference in the posture of the person and perform high-accuracy authentication.

  Further, in a state (standard state) in which the individual models MDa and MDb are projected on a common projection plane in accordance with the projection position of the standard model, the texture information extracted from the individual model related to the authentication target person HMb and the comparison target person HMa Is compared with the texture information extracted from the individual models. In this case, the corresponding patches of the individual models MDa and MDb are projected on a common projection plane so that the positions thereof are the same. Therefore, since it is possible to perform the comparison process after sufficiently setting the comparison conditions, more accurate authentication can be performed.

  Furthermore, when both the individual models MDa and MDb are projected onto a common projection surface in accordance with the projection position of the standard model, the corresponding projection planes of the individual models MDa and MDb have the same shape so as to have the same shape. Will be projected. Therefore, the texture information extracted from the individual model related to the authentication target person HMb is compared with the texture information extracted from the individual model related to the comparison target person HMa in a standardized state with respect to the shape of the corresponding patch. Accurate texture comparison can be performed.

<D. Modification>
In the above-described embodiment, “feature points” are illustrated as representative points of the three-dimensional model. However, the present invention is not limited thereto, and points other than the feature points may be adopted as representative points of the three-dimensional model.

In the above embodiment, as the local two-dimensional information h ^(k) , the luminance values of the respective pixels are enumerated as shown in Expression (9), but the present invention is not limited to this. For example, the feature amount of each feature point may be calculated using a Gabor filter. Specifically, for each of a plurality of feature points included in the group, a feature amount using a two-dimensional Gabor filter as shown in Expression (18) can be calculated. Here, σ represents the spread of the Gaussian function, ε represents the scale parameter, and τ represents the direction parameter.

Specifically, at each feature point, four values of the direction parameter τ (for example, τ = 0, π / 4, π / 2, 3π / 4 (rad)) and three values of the scale parameter ε ( For example, while changing ε = π, π / 2, π / 4 (rad)), the Gabor filter of Expression (18) is applied to obtain a total of 12 (4 × 3) feature values. A vector in which a total of 12 values related to each feature point in the group are arranged for a plurality of feature points can be obtained as each local two-dimensional information h ^(k) . At this time, the number of dimensions of each local two-dimensional information h ^(k) is a value obtained by multiplying the number of feature points included in the group by “12”.

  According to the Gabor filter, it is possible to efficiently extract local shading information (such as a contour line in a specific direction) of an image.

  Moreover, in the said embodiment, as shown in Formula (17), the identity of a person to be authenticated and a registrant is determined using not only texture information but also shape information. However, the present invention is not limited to this. The identity of the person to be authenticated and the registrant may be determined using only the texture information.

  In the above embodiment, the case where it is determined whether or not the input face (face of the person to be authenticated) is a specific registrant has been described. However, the present invention is not limited to this. You may make it apply said idea to the face identification which determines who is among the registrants. For example, the similarity between each face feature quantity of a plurality of registered persons and the face feature quantity of the person to be authenticated is calculated, and the identity between each person to be compared (registrant) and the person to be authenticated is calculated. What is necessary is just to judge. Further, when it is sufficient to narrow down to a specific number of persons, a specific number of comparison target persons may be selected in descending order of identity among a plurality of comparison target persons.

  In the above embodiment, the three-dimensional information of the face of the person to be authenticated is acquired using a plurality of images input from a plurality of cameras. However, the present invention is not limited to this. Specifically, the reflected light of the laser irradiated by the laser beam emitting unit L1 is measured by the camera LCA using a three-dimensional shape measuring instrument configured by the laser beam emitting unit L1 and the camera LCA as shown in FIG. By doing so, the three-dimensional information of the face of the person to be authenticated may be acquired. That is, three-dimensional information may be acquired by a light cutting method. However, according to the method of acquiring three-dimensional shape information using an input device including two cameras as in the above-described embodiment, the three-dimensional shape is relatively simple compared to an input device using laser light. Shape information can be acquired.

It is a conceptual diagram which shows the outline | summary of a face authentication system. It is a block diagram which shows a face authentication system. It is a figure which shows the structure outline | summary of a controller. It is a block diagram which shows the various function structures of a controller. It is a block diagram which shows the detailed functional structure of a personal authentication part. It is a flowchart which shows authentication operation | movement. It is a flowchart which shows registration operation | movement. It is a figure which shows the feature point of the characteristic site | part in a face image. It is a schematic diagram which calculates a three-dimensional coordinate from the feature point in a two-dimensional image. It is a figure which shows the standard model of a three-dimensional face. It is a figure which shows the point of the movement destination of each feature point in an individual model, and the point of movement of the corresponding feature point in a standard model. It is a figure which shows the cylindrical surface arrange | positioned around a standard model. It is a figure which shows the spherical surface arrange | positioned around a standard model. It is a figure showing an example of classification of each patch. It is a figure which shows a mode that each patch which belongs to a certain group of a standard model is projected on a projection surface. It is a figure which shows the coordinate axis set along the surface shape of a standard model. It is a figure which shows the relationship between the coordinate axis set in the attention point of a standard model, and the coordinate axis of the coordinate system in a projection plane. It is a figure which shows the state after an optimization process. It is a figure which shows the three-dimensional shape measuring device using the optical cutting method.

Explanation of symbols

1 Face recognition system (authentication system)
10 Controllers AX1 to AX4, CX1 to CX4 Coordinate axes Ad, Bd Facial features C1 to C5 Patch CA1, CA2 Camera HMa Comparison target HMb Authentication target PB1 to PB5 Feature points (on individual models) PL, PL11, PL12, PL15 Projection plane PS0 to PS5 Feature points (on standard model) PT0 to PT5 Target point

Claims (10)

An authentication system for performing an authentication operation on an authentication target person,
First acquisition means for acquiring texture information and three-dimensional information related to the face of the person to be authenticated;
Generating means for generating a first three-dimensional model that is a three-dimensional model related to the face of the person to be authenticated based on the three-dimensional information;
Conversion means for converting first texture information, which is texture information related to the face of the person to be authenticated, using the first three-dimensional model;
Second acquisition means for acquiring second texture information that is texture information relating to the face of the person to be compared;
A determination unit that determines the identity of the comparison target person and the authentication target person by comparing the first texture information and the second texture information from the same viewpoint;
An authentication system comprising: The authentication system according to claim 1,
The converting means includes a plurality of representative points of the first stereo model based on a correspondence relationship between each representative point of the first stereo model and each representative point of a standard model that is a standard stereo model of a human face. Each of which is projected onto the same position as the projection destination point of the corresponding representative point of the standard model, and texture information corresponding to a patch composed of the plurality of representative points is allocated on the projection plane. . The authentication system according to claim 2,
The authentication system, wherein the conversion means assigns texture information of the first three-dimensional model to a predetermined projection plane arranged around the standard model. The authentication system according to claim 3,
The authentication system according to claim 1, wherein the predetermined projection surface is a cylindrical surface. The authentication system according to claim 3,
The authentication system, wherein the predetermined projection surface is a spherical surface. In the authentication system according to claim 1 or 2,
The converting means classifies patches included in the first three-dimensional model into a plurality of groups, and assigns texture information corresponding to each patch in the group to a projection plane set for each group after classification. Authentication system. The authentication system according to any one of claims 1 to 6,
The converting means applies each patch of the first three-dimensional model to the projection plane so that the area of each patch on the first three-dimensional model is the same as the area of each patch on the projection plane after projection. An authentication system characterized by mapping. In the authentication system according to any one of claims 1 to 7,
The determination means compares the feature amount extracted from the second texture information related to the face of the comparison target person with the feature amount extracted from the converted first texture information, and An authentication system for determining identity with an authentication target person. An authentication method for performing an authentication operation on an authentication target person,
Obtaining texture information and three-dimensional information related to the face of the person to be authenticated;
Generating a first three-dimensional model that is a three-dimensional model related to the face of the person to be authenticated based on the three-dimensional information;
Converting the first texture information, which is texture information about the face of the person to be authenticated, using the first three-dimensional model;
Obtaining second texture information that is texture information related to the face of the person to be compared;
Determining the identity of the comparison target person and the authentication target person by comparing the first texture information and the second texture information from the same viewpoint;
An authentication method comprising: A program for causing a computer to execute an authentication method for performing an authentication operation related to an authentication target person.
A procedure for acquiring texture information and three-dimensional information related to the face of the person to be authenticated;
Generating a first three-dimensional model that is a three-dimensional model related to the face of the person to be authenticated based on the three-dimensional information;
A step of converting first texture information, which is texture information related to the face of the person to be authenticated, using the first three-dimensional model;
A procedure for obtaining second texture information that is texture information relating to the face of the person to be compared;
A procedure for determining the identity of the comparison target person and the authentication target person by comparing the first texture information and the second texture information from the same viewpoint;
A program for running