JP2000222572A - Gender determination method - Google Patents

Abstract

(57) [Summary] [Problem] To determine the gender of a man and a woman at a high discrimination rate regardless of whether a learning face image and a face image to be identified are shot under the same environment or under different environments. A face image to be identified is converted into a principal component by a principal component vector, and the converted principal component is input to a calculated linear discriminant analysis formula to determine the gender of the gender. That is, the gender unknown face image x is converted into p principal component values by the principal component vector, and k principal components are selected to create a k-dimensional vector z. Then, the k-dimensional vector z is substituted into the linear primary discriminant analysis equation for which the parameters have been determined in the learning phase to obtain w.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for determining gender based on a face image.

[0002]

2. Description of the Related Art As a method for determining gender, for example, the literature “Medical Imaging Technology” Vol. 12 No. 6 710-715
In “Determination of Gender from 8 × 6 Ultra-Low-Resolution Images Using Neural Network”, the face
A method is described in which a small image of 8 or 8 × 6 is set, and each luminance value of the small image is input to a neural network to determine gender. In this method, a learning face image is first mosaiced, input to a neural network to perform network learning, and a face image to be identified is similarly mosaicized and input to the neural network to determine gender. is there. On the other hand, discriminant analysis in pattern recognition is usually performed using a feature amount obtained as a result of performing image processing on all luminance values of a face image.

[0003]

However, in the former determination method, since the average value of the luminance of the image is obtained for each block and the image size is reduced, luminance information that may be important for identification is also obtained. There is a problem that the gender of a man and a woman cannot be identified with high accuracy because the data is averaged for each block.

When the latter method of performing discriminant analysis by applying image processing to all luminance values of a face image is applied to the determination of gender, a set of face images photographed in a certain environment is used for learning. When trying to determine the gender of a face image captured in another environment based on the information, there is a problem that the discrimination rate decreases due to the influence of the environment. That is, it is good if the learning face image and the face image to be identified are images captured in the same environment, but there is a problem that gender of the gender cannot be accurately distinguished when captured in different environments.

[0005] Therefore, the inventions described in the claims claim that the gender of a man and a woman can be determined not only when the learning face image and the face image to be identified are photographed under the same environment but also under different environments. Provided is a gender determination method capable of determining with high identification rate and high accuracy.

[0006]

According to the first aspect of the present invention,
Principal component analysis is performed on a set of male and female learning face images to obtain a principal component vector, each learning face image is converted into a principal component by the principal component vector, and each of the converted learning face images is converted to a principal component. The parameters of the linear discriminant analysis for discriminating between men and women are determined using the principal components. Subsequently, the face image to be identified is converted into the principal components by the principal component vectors, and the linear discriminant analysis is performed using the converted principal components. To determine the gender of the gender.

According to a second aspect of the invention, a principal component analysis is performed on a set of male and female learning face images to obtain a principal component vector, and each learning face image is converted into a principal component by the principal component vector. Using the converted principal components of each of the learning face images, parameters for nonlinear discriminant analysis for discriminating between men and women are determined, and then the face image to be identified is converted into the principal components by the principal component vector. The gender of a man and a woman is determined by performing a nonlinear discriminant analysis using the converted principal components.

According to a third aspect of the present invention, in the gender determination method according to the first or second aspect, the determination of the parameters of the discriminant analysis and the discriminant analysis of the face image to be identified are performed using the upper principal components or the selected principal components. To do.

[0009]

Embodiments of the present invention will be described with reference to the drawings. (First Embodiment) FIG. 1 is a block diagram showing the configuration of a gender determination apparatus for males and females. Reference numeral 1 denotes an image provided with a CCD camera and the like, which captures a face image of a person including a background and inputs the information as image information. Input means. Reference numeral 2 denotes a face detection unit that extracts only a face image from the image information captured by the image input unit 1. Reference numeral 3 denotes a gender determining unit which functions as a main part, determines the gender of the male and female from the face image extracted by the face detecting unit 2, and outputs the result.

In order for the gender judging means 3 to judge gender, a learning phase for learning a face image is required in advance.
This learning phase is performed based on the flowchart shown in FIG. That is, in the learning phase S1, the principal component analysis of n learning face images is performed. Subsequently, in the learning phase S2, the n learning face images are converted into the principal components by the principal component vector. In the learning phase S3, the parameters of the linear discriminant analysis based on the principal components of the n learning face images are calculated.

The learning phase process will be described in more detail. First, in the learning phase S1, a learning face image including n men and women is prepared, and principal component analysis is performed on the n face images. The size of each of the n learning face images is defined as length × width = d.

Assuming that the image x is a d-dimensional vector x = [x ₁ , x ₂ ,..., X _d ] ^T (1), a covariance matrix S is calculated from a plurality of images x of the learning sample images. , S = E ｛(x−m) (x−m) ^T ｝ (2) However, an autocorrelation matrix that does not subtract an average value may be used instead of the covariance matrix.

M = E {x} (3) is an average vector of x. Then, taking p number of eigenvectors of _{Su j = λ j u j (} 4) S from the largest of the corresponding eigenvalues. Usually p is the total number of samples n or the dimension d of the original image
Is smaller than the smaller one of the two. This is the principal component vector extracted from the learning face image.

[0014] _{{u 1, u 2, ·····} , u p} (5) Subsequently, the learning phase S2, each learning face image _{x y = [u 1, u} 2, ···· _·, u ^p] T x ^{p-dimensional} vector composed of p pieces of the main component values by _{(6) y = [y 1} , y 2, ·····, converted to y ^{p] T} (7). That is, the n learning face images are
According to the equation (6), it is converted into the p-dimensional vector of the equation (7). The principal component values are arranged in descending order, and are respectively referred to as a first principal component, a second principal component,....

For each of the n learning face images, k of the p principal components are input to the discriminant analysis described below. That is, a k-dimensional vector z = [z ₁ , z ₂ ,..., Z _k ] ^T (8) using k principal components is used as an input.

The k principal components may be selected from the first principal component to the k-th principal component as the top k principal components among the p principal components. ,
Some may be arbitrarily selected, such as a first principal component, a third principal component, and a fifth principal component. Also, at the time of the principal component analysis in the learning phase S1, the top k principal component vectors are calculated in advance without calculating the p principal component vectors, and the learning face S2 is used in this learning phase S2. The image may be converted into a k-dimensional vector composed of k principal component values. Thus, utilizing the k-dimensional vector z ₁ to z _n of n sheets constituted by the k principal components obtained in the parameter calculation for the discriminant analysis in the next step.

[0017] In the learning phase S3, first of all, out of the n pieces of learning sample face images, including men and women, the number of face image male n _1, the number of face image women and n _2. Learning phase S2
Let β _{1 be} the average vector of the k-dimensional vector z of the n ₁ male face images and β _{2 be} the average vector of z of the n ₂ female face images.

Further, the covariance matrix of z of n ₁ male face images is S ₁ , and the covariance matrix of z of n ₂ female face images is S ₂
And the pooled covariance matrix is S = 1 / (n ₁ + n ₂ -2) × {(n ₁ -1) S ₁ + (n ₂ -1) S ₂ } (9), and w = ｛ A linear discriminant of z−1 / 2 (β ₁ + β ₂ )｝ ^{TS −1} (β ₁ −β ₂ ) (10) is obtained. Thus, the parameter calculation of the linear primary discriminant analysis is completed.

The gender determination means 3 executes an identification phase based on the flow chart shown in FIG. 3 using the linear discriminant obtained in the learning phase described above. That is, in the identification phase S11, the face image to be identified is converted into the principal component by the principal component vector. Subsequently, in the identification phase S12, the converted principal components are input to the calculated linear discriminant analysis formula,
In the discrimination phase S13, the gender of the gender is determined.

More specifically, the processing in the identification phase will be described. First, in the identification phase S11, the face image x of unknown gender extracted by the face detection means 2 is expressed by the main equation (5) as in the learning phase. The component vectors are converted into p principal component values, and k principal components are selected to create a k-dimensional vector z. If the top k principal component vectors have been calculated in advance by equation (5), the face image is converted into k principal component values to create a k-dimensional vector z.

Subsequently, in the discrimination phase S12, the obtained k-dimensional vector z is converted to a linear primary discriminant analysis equation whose parameters are determined in the learning phase, that is, z in equation (10).
To obtain w. Then, in the identification phase S13, if the value of w is positive, it is determined that it is male, and if it is negative, it is determined that it is female.
In this case, it is assumed that the prior probabilities of the men and women are equal.

The above is a linear primary discriminant analysis in a multivariate analysis, and Fisher's Linear Discriminant Analyst used for pattern recognition and the like.
s) can also be used. This is a method of finding a base vector that reduces intra-group variance and increases inter-group variance. Also, a one-layer perceptron for obtaining a hyperplane separating the two groups is included because it is substantially equivalent to linear discriminant analysis.

Further, a discriminant linear quadratic discriminant analysis, w = 1/2 {( z-β 1) T S 1 -1 (z-β 1) - (z-β 2) T S 2 - ^{By 1} (z-β ₂ ) + log | S ₂ | −log | S ₁ |｝ (11), the gender can be similarly determined by the sign of w. This secondary discriminant analysis is a statistical function for a normal distribution pattern in pattern recognition, g _i (z) = 1/2 {(z-β _i ) ^T S _i ⁻¹ (z-β _i ) -1/2 Log | S _i | + log P (w _i ) -n / 2 · log 2π｝ (12) Usually, the classification is made into a class in which g _i (z) is large. The above equation (11) is obtained by taking the difference from the above equation (12) with i = 1,2.

FIG. 4 shows the determination of gender by the combination of the above-described principal component analysis and discriminant analysis. That is, the learning sample face image on the d-dimensional space is dropped on the k-dimensional subspace by principal component analysis, and a hyperplane M for separating male and female is put in the k-dimensional subspace. In the case of linear primary discriminant analysis (including a single-layer perceptron), a hyperplane is obtained, whereas in the case of linear quadratic discriminant analysis, a hypersurface is obtained.

In this way, the gender determination processing is performed by the gender determination means 3. In this case, the gender determination is performed after selecting only the main components or only some of the main components by principal component analysis. In particular, the gender determination ability is improved particularly in an image captured in a different environment from the learning face image.

That is, FIG. 5 is a graph showing the relationship between the number of principal components and the discrimination rate, and the graph g1 shows the photographing environment A in which both the photographing of the face image during learning and the photographing of the tested face image are the same. Graph g2 shows a case where shooting of a face image during learning and shooting of a tested face image are performed in different shooting environments such as shooting environment A during learning and shooting environment B during testing. I have.

From these graphs, it can be seen that the discrimination rate increases as the number of principal components increases up to the point where there is a face image photographed in any of the photographing environments A and B. But,
After a certain point, the recognition rate decreases. In the case of a face image shot in the same shooting environment A as the face image at the time of learning, the drop is not significant, but in the case of a face image shot in shooting environment B, the drop is significant.

From these results, it is clear that the principal component analysis and the k-dimensional vector are subjected to the discriminant analysis rather than the direct discriminant analysis of the d-dimensional vector without the principal component analysis, even if the photographing environment changes. It can be seen that the gender determination ability can be kept high. Therefore, it can be concluded that adopting a method of performing discriminant analysis after performing principal component analysis is effective for gender determination in various environments. In particular, it is effective for determining the gender of a face image captured in an environment different from the learning face image.

Next, comparison between gender judgment by discriminant analysis and another gender judgment will be described. FIG. 6 is a graph showing the relationship between the number of principal components and the discrimination rate of gender judgment by each of linear primary discriminant analysis, linear quadratic discriminant analysis, nearest neighbor determination rule, K-nearest neighbor determination rule, and CLAFIC method. g11 shows the result of linear primary discriminant analysis, graph g12 shows the result of linear secondary discriminant analysis, graph g13 shows the result of nearest neighbor decision rule, graph g14 shows the result of K-neighbor decision rule, Graph g15 shows the result of the CLAFIC method.

From this result, it can be seen that the discrimination rate by male and female by the linear discriminant analysis is high. The CLAFIC method is a subspace method in which eigenvectors obtained by principal component analysis at the origin are used for similarity calculation. In all other identification methods, after a normal principal component analysis, the principal component data is used for learning data or the like to make a distinction. Therefore, although the method is slightly different, the number of principal components can be similarly plotted on the horizontal axis, so they are shown on the same graph. As described above, by performing the discriminant analysis after the principal component analysis, it is possible to enhance the ability to determine the gender of males and females, and it is found that the gender determination is extremely effective. (Second Embodiment) In this embodiment, a gender determination means 3 uses a determination method based on non-linear discriminant analysis. Note that the configurations of the image input unit 1 and the face image detection unit 2 are the same as those of the first embodiment described above.
Performing the learning phase in advance is also the same as in the first embodiment.

This nonlinear discriminant analysis utilizes a multilayer neural network. That is, a non-linear discrimination plane is inserted between a male k-dimensional vector z and a female k-dimensional vector z by a multilayer neural network. Usually, a back propagation method (error back propagation learning method) is used for learning of the multilayer neural network.

Therefore, in learning by the multilayer neural network, a k-dimensional vector composed of principal components is used. Therefore, as shown in FIG. 7, there are k inputs (k + 1 if a constant is input) and an output Is 2 for male and female. However, it is also possible to set the output to one of male and female by setting a threshold value.

The learning phase is performed based on the flowchart shown in FIG. 8, that is, in the learning phase S21, the first phase described above is performed.
Similarly to the learning phase S1 of the embodiment, principal component analysis is performed on n face images. In the subsequent learning face S22, similarly to the learning phase S2 of the first embodiment described above, the n learning face images are converted into k-dimensional vectors by principal component analysis.

In the subsequent learning phase S23, each k-dimensional vector is input to the neural network, and if the output is male, a teacher signal of "1,0" is output, and if it is female, a teacher signal of "0,1" is given. The weight of the input signal to each neuron is modified so as to reduce the error. Normally, the back propagation method is used for the calculation for this correction.

Then, when actually determining the gender of a face image whose gender is unknown, an identification phase based on the flowchart shown in FIG. 9 is executed. That is, in the identification phase S31, the face image to be identified is converted into the principal component by the principal component vector and k
Find the dimension vector z. Subsequently, in the identification phase S32, the converted principal components are input to the calculated linear discriminant analysis equation. That is, the k-dimensional vector z is input to a k-input multilayer neural network. Subsequently, in the discrimination phase S33, the gender of the gender is determined. That is, if the output is "1" for the male output and "0" for the female output, the output is male, and if the male output is "0" and the female output is "1", the output is female. The output takes a value in the range of 0 to 1, and the gender is determined based on which of the male output and the female output is closer to 1.

As described above, even when nonlinear discriminant analysis is used, gender is determined by a combination of principal component analysis and discriminant analysis, which is shown in FIG. 4 as in the first embodiment. However, in the non-linear discriminant analysis, the surface separating men and women in the k-dimensional subspace is a hypersurface, and the surface is a more complicated surface than in the case of the linear quadratic discriminant analysis.

In addition, even when this nonlinear discriminant analysis is used, gender is determined by a combination of principal component analysis and discriminant analysis. Is very effective for gender determination.

[0038]

According to the invention described in each of the claims, not only when the learning face image and the face image to be identified are photographed under the same environment but also under different environments, the Gender can be determined with a high identification rate and with high accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a flowchart showing processing in a learning phase according to the embodiment;

FIG. 3 is a flowchart showing processing in an identification phase according to the embodiment;

FIG. 4 is an exemplary view for explaining a gender determination method in the embodiment.

FIG. 5 is a graph showing the relationship between the number of principal components and the discrimination rate according to a difference in a shooting environment in the embodiment.

FIG. 6 is a graph showing the relationship between the number of principal components and the discrimination rate when the determination method of the embodiment is compared with another determination method.

FIG. 7 is a diagram showing a multilayer neural network used in a second embodiment of the present invention.

FIG. 8 is a flowchart showing processing in a learning phase according to the embodiment;

FIG. 9 is a flowchart showing processing in an identification phase according to the embodiment;

[Explanation of symbols]

 2: Face detection means 3: Gender determination means

Claims (3)

[Claims] 1. A principal component analysis is performed on a set of male and female learning face images to obtain a principal component vector, each learning face image is converted into a principal component by the principal component vector, and each of the converted learning face images is converted. The parameters of linear discriminant analysis for discriminating between men and women are determined by using the respective principal components of the face image. Subsequently, the face image to be identified is converted into the principal component by the principal component vector, and this converted A gender determination method characterized by determining gender of male and female by performing linear discriminant analysis using principal components. 2. A principal component analysis is performed on a set of male and female learning face images to obtain a principal component vector, each learning face image is converted into a principal component by the principal component vector, and each of the converted learning face images is converted. Determine the parameters of nonlinear discriminant analysis for discriminating between men and women using each principal component of the face image,
Subsequently, a gender determination method is characterized in that a face image to be identified is converted into a principal component by the principal component vector, and the converted principal component is subjected to non-linear discriminant analysis to determine gender of the gender. 3. The gender determination method according to claim 1 or 2, wherein the determination of the parameters of the discriminant analysis and the discriminant analysis of the face image to be identified are performed using the upper principal component or the selected principal component.