JP2011022916A - Device, method and system for estimation of individual attribute - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To estimate individual attributes, such as gender and age, of a target person with high accuracy on the basis of biological information. <P>SOLUTION: Individual attributes are classified into a plurality of categories, a criterion that discriminates which of two different categories they belong to is learned, and a discriminant is decided in advance. A feature value is extracted from acquired biological information and is input to the discriminant to discriminate which category it belongs to. The discrimination is performed with respect to the respective classified categories to estimate individual attributes on the basis of the respective discrimination results. In estimating a plurality of individual attributes, such as a gender and age, categories subdivided according to the plurality of individual attributes are created and the individual attributes are estimated stepwise. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a personal attribute estimation device, a personal attribute estimation method, and a personal attribute estimation system that acquire biometric information from a personal attribute estimation target person and estimate a personal attribute.

  2. Description of the Related Art Conventionally, personal attribute estimation technology for estimating gender and age from a customer's face image captured by a camera installed in a commercial facility has been widely used for marketing and advertising.

  For example, in Patent Document 1, an average feature value of a face for each gender and age is stored in advance, and a facial feature value extracted based on captured image data and a stored feature value are stored. A personal attribute estimation device that estimates gender and age with high similarity is disclosed.

JP 2003-242486 A

  However, since the above-described conventional personal attribute estimation device performs comparison based on the averaged feature value, there is a problem that erroneous estimation is likely to occur in the case of a feature value with little difference between categories.

  Specifically, when capturing an image, the feature amount extracted from the captured image data may fluctuate due to an unexpected reason such as a poor illumination intensity. In addition, there is a tendency that gender differences do not appear easily on the faces of young people who are growing up.

  For example, it is assumed that the feature value of a 5-year-old man and the feature value of a 10-year-old woman are very close values. In this case, when the feature value extracted from the image data of a 10-year-old woman has a value close to the feature value of a 5-year-old man, the conventional personal attribute estimation device uses a 5-year-old man. It is mistakenly estimated to be.

  From these facts, how to realize a personal attribute estimation device, a personal attribute estimation method, and a personal attribute estimation system that can improve the estimation accuracy of personal attributes has become a major issue. Such a problem is not limited to the exemplified gender and age, and is also a problem that applies when estimating individual attributes such as nationality and occupation.

  The present invention has been made to solve the above-described problems of the prior art, and provides a personal attribute estimation device, a personal attribute estimation method, and a personal attribute estimation system that can improve the estimation accuracy of personal attributes. With the goal.

  In order to solve the above-described problems and achieve the object, the present invention is a personal attribute estimation device that estimates the personal attributes of the target person from the biological information of the target person, wherein the biological information of the target person is the The living body acquired from the subject based on learned information generated in advance from a plurality of biological information whose personal attributes are known as criteria for determining which of two different categories related to the personal attribute A discriminating unit that discriminates which category of the two different categories the information belongs to, and a plurality of discrimination results discriminated by the discriminating unit for each combination of the two different categories. And a personal attribute estimating means for estimating a personal attribute. As described above, by identifying which of the two personal attributes belongs to the personal attribute, it is possible to improve the estimation accuracy of the personal attribute.

  In the invention, it is preferable that a discriminant for discriminating which category of the two different categories belongs to the learned information is determined in advance for each combination of the two different categories. The means is characterized in that a characteristic value of the biological information is input to the discriminant and a value calculated is output as the discrimination result. As described above, by identifying the personal attribute based on the identification value between the personal attributes calculated by the discriminant, there is an effect that the estimation accuracy of the personal attribute can be improved.

  Further, according to the present invention, the personal attribute estimation unit aggregates the discrimination results in a combination of the category and a plurality of other categories for each category, and determines the personal attribute of the target person based on the aggregated aggregation result. It is characterized by estimating. In this way, by identifying individual attributes based on the result of the aggregation of discrimination results, even if a discrepancy occurs in a discriminant result, discrimination using a plurality of discriminant results can be used. The influence of the result can be suppressed, and there is an effect that accurate estimation information can be obtained.

  Further, according to the present invention, the personal attribute estimation means estimates a first personal attribute based on the determination result, and based on only the combination of the personal attributes related to the estimated first personal attribute. It is characterized by estimating two personal attributes. In this way, by limiting the discriminant to estimate and use the first personal attribute in advance, the estimation range can be narrowed down efficiently, and accurate estimation information can be obtained.

  Further, the present invention is characterized in that the personal attribute includes a sex and an age group of the subject. In this way, by identifying the gender and age, there is an effect that accurate estimated information can be obtained by utilizing the estimated gender and age for marketing and advertising.

  According to the present invention, the personal attribute estimation device acquires a plurality of image data in advance, and obtains and stores an identification function between all categories based on the feature amount extracted from the acquired image data. Then, the personal attribute estimation device calculates the identification value for all categories by substituting the feature quantity extracted from the image data of the personal attribute estimation target person into the identification function. Furthermore, the personal attribute estimation device estimates gender and age, which are personal attributes, based on the calculated identification values between all categories. By doing in this way, there exists an effect that the estimation precision of a personal attribute can be improved.

FIG. 1 is a diagram showing an outline of a personal attribute estimation method according to the present invention. FIG. 2 is a block diagram illustrating the configuration of the personal attribute estimation apparatus according to the present embodiment. FIG. 3 is a diagram illustrating an example of the identification value. FIG. 4 is a diagram illustrating an example of a sex degree calculation processing procedure. FIG. 5 is a diagram illustrating an example of a detailed gender degree calculation processing procedure. FIG. 6 is a diagram illustrating an example of an age calculation processing procedure. FIG. 7 is a flowchart showing an outline of a processing procedure executed by the personal attribute estimation device. FIG. 8 is a flowchart showing the processing procedure of the identification value calculation processing. FIG. 9 is a flowchart showing a processing procedure of the gender degree calculation processing. FIG. 10 is a flowchart showing the processing procedure of the detailed gender degree calculation processing. FIG. 11 is a flowchart showing the processing procedure of the age calculation processing procedure. FIG. 12 is a flowchart showing the processing procedure of the detailed age calculation processing.

  Exemplary embodiments of a personal attribute estimation method according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, the outline of the personal attribute estimation method according to the present invention will be described with reference to FIG. 1, and then an embodiment of the personal attribute estimation apparatus to which the personal attribute estimation method according to the present invention is applied will be described with reference to FIGS. This will be described using.

  FIG. 1 is a diagram showing an outline of a personal attribute estimation method according to the present invention. 1A shows a personal attribute estimation method according to the prior art, and FIG. 1B shows a personal attribute estimation method according to the present invention.

  As shown in FIG. 1A, in the personal attribute estimation method according to the prior art, the personal attribute is estimated based on the averaged feature amount calculated in advance for each gender and age. For example, in the case shown in the figure, if the characteristic amount acquired from the individual attribute estimation target person (hereinafter simply referred to as “estimation target person”) is “C”, the personal attribute is “male” and “age” It is estimated that “20 to 39 years old”.

  However, the personal attribute estimation method according to the conventional technique has a problem that erroneous estimation is likely to occur when the difference in feature amount calculated in advance for each category is small between the categories. For example, as shown in the figure, it is assumed that the feature value C of a man in his 20s is close to the feature value C ′ of a woman in his 10s. Here, when a face of a woman in her 10s is imaged, the feature amount extracted from the imaged image data may fluctuate due to an unexpected reason such as poor illumination.

  When the feature value extracted from the image data of a 10-year-old woman becomes a value close to the feature value C of a 20-year-old man, Mistakenly presumed to be a teenager man.

  Therefore, in the personal attribute estimation method according to the present invention, as shown in FIG. 1B, it is determined which personal attribute is probable for each combination of personal attributes shown in a matrix, and each determination result. The final personal attributes were estimated by comprehensively judging

  Specifically, in the personal attribute estimation method according to the present invention, for each cell of the matrix shown in the figure, the personal attribute on the vertical axis or the personal attribute on the horizontal axis is quantified. A determination formula to be output (hereinafter referred to as “discriminant function”) is generated by prior learning.

  In the personal attribute estimation method according to the present invention, a value calculated by substituting the feature quantity extracted from the image data of the estimation target person into the identification function of each cell (hereinafter referred to as “identification value”). To estimate the personal attributes of the person to be estimated.

  Here, FIG. 1B shows a value (identification value) calculated by substituting the feature quantity of the person to be estimated into the identification function obtained in advance. For example, the identification value of the cell in which “female: 10” on the horizontal axis and “male: 20 on” on the vertical axis in FIG. In this case, it indicates that the estimation target person is identified as belonging to “male: 20 to”, either “female: 10” or “male: 20 to”.

  In addition, the identification value of the cell where “female: 10” on the horizontal axis and “male: 40 on” on the vertical axis in FIG. In this case, if the estimation target person is “female: 10” or “male: 40-”, it indicates that the estimation target person has been identified as belonging to “female: 10”.

  That is, if the identification value of a cell where the horizontal axis category and the vertical axis category intersect is a positive number, it belongs to the horizontal axis category, and if it is a negative number, it belongs to the vertical axis category. Identified. When the absolute value of the identification value is large, it indicates that the probability (accuracy) belonging to the corresponding category is large. When the identification value is 0, it cannot be identified which category it belongs to.

  As described above, when the identification value is calculated by the identification function of each cell, the personal attribute estimation method according to the present invention performs a process of adding the identification values in the horizontal axis direction or the vertical axis direction. Such processing is performed in order to globally evaluate the identification value calculated finely for each cell.

  For example, when looking at “female: 20-” on the horizontal axis in the vertical axis direction, all numerical values are positive. This indicates that the category “female: 20-” is dominant over all other categories. Further, when “male: 60˜” on the horizontal axis is viewed in the vertical axis direction, all numerical values are negative. This indicates that the category of “male: 60-” is inferior to all other categories.

  Thus, the personal attribute estimation method according to the present invention estimates the most dominant category among all categories as the personal attribute of the estimation target person.

  Note that the personal attribute estimation method according to the present invention first estimates whether the estimation target person is “male” or “female”, limits the discrimination function to be used, and then estimates the estimation target based on the estimated gender accuracy. The process of estimating the “age group” of the person is performed, and details of this point will be described later.

  Below, the Example about the personal attribute estimation apparatus to which the personal attribute estimation method concerning this invention is applied is described in detail. FIG. 2 is a block diagram illustrating the configuration of the personal attribute estimation device 10 according to the present embodiment.

  The personal attribute estimation device 10 includes an imaging unit 11, a result output unit 12, a storage unit 13, and a control unit 14. The control unit 14 further includes an identification function creation unit 14a, a feature amount extraction unit 14b, an identification value calculation unit 14c, and a personal attribute determination unit 14d. The storage unit 13 includes learning information 13a, The discriminating function 13b is stored.

  The imaging unit 11 includes an imaging device such as a CCD (Charge Coupled Devices) camera. For example, if the imaging unit 11 is configured with a visible camera, the visible light reflected by the authentication object is detected and captured as a visible image. Then, the imaging unit 11 outputs the captured visible image to the feature amount extraction unit 14b of the control unit 14.

  The result output unit 12 is an output device such as a display device that displays the determination result determined by the personal attribute determination unit 14d, and includes, for example, a liquid crystal display or a speaker.

  The storage unit 13 is a storage unit configured by a storage device such as a nonvolatile memory or a hard disk drive, and stores learning information 13a and an identification function 13b. The learning information 13a is information in which feature amounts of image data whose personal attributes are known are collected in advance.

  The discriminant function 13b is a function created by the discriminant function creating unit 14a, and is a linear discriminant in which each coefficient is determined by learning using the learning information 13a. In this embodiment, the case where a linear discriminant is used as the discriminant function 13b will be described. However, a nonlinear discriminant may be used.

  The discriminant function creating unit 14a is a processing unit that creates a discriminant function representing a boundary for discriminating two categories based on the learning information 13a and stores the discriminating function in the storage unit 13 as the discriminant function 13b.

  Here, the discrimination function creating unit 14a included in the personal attribute estimation device 10 creates the discrimination function based on the learning information 13a stored in the storage unit 13, but the discrimination function creating unit 14a and the learning information 13a may not be present in the personal attribute estimation device 10. The personal attribute estimation device 10 may acquire an identification function previously created by a device other than the personal attribute estimation device 10 and store it in the storage unit 13 as the identification function 13b.

  The feature amount extraction unit 14b compares the image data picked up by the image pickup unit 11 with the average face image, detects the positions of eyes, nose, mouth, etc., and extracts the feature amounts corresponding to each position. , A processing unit that performs processing to be passed to the identification value calculation unit 14c.

  The identification value calculation unit 14c calculates the identification value by substituting the feature amount received from the feature amount extraction unit 14b into the identification function 13b, and performs processing for passing the calculated identification value to the personal attribute determination unit 14d Part.

  The personal attribute determination unit 14d is a processing unit that performs processing for determining a personal attribute based on the identification value received from the identification value calculation unit 14c. In addition, a present Example demonstrates the case where the sex and age of an estimation object person are determined.

  Next, the identification analysis performed by the personal attribute estimation device 10 will be described with reference to FIG. FIG. 3 is a diagram illustrating an example of the identification value. Here, the discriminant analysis is a technique for obtaining a boundary (hereinafter referred to as “discrimination function”) for identifying which category belongs to between two different categories.

  Below, the boundary which identifies a different category is made into the space which can be expressed with a linear equation, and an identification function is illustrated using the function which can be expressed with a linear equation called a linear discriminant function. Note that boundaries that identify different categories may be expressed using a non-linear function.

  First, the captured image data is compared with an average face image, the positions of eyes, nose, mouth, and the like are detected, and feature amounts corresponding to the positions are extracted. Here, feature values at 1024 locations (32 × 32 = 1024), which are 32 horizontal and 32 vertical lattice points, are extracted. The extracted feature quantity is represented by a 1024-dimensional real vector X (hereinafter referred to as “feature vector”), and identification analysis is performed in a space converted into a 1024-dimensional multidimensional space.

この場合、特徴ベクトルＸは式（１−１）のように表現することができる。ここで、「Ｍ」は、１０２４次元の空間であるので、式（１−２）のようにＭ＝１０２４とする。

In this case, the feature vector X can be expressed as in Expression (1-1). Here, since “M” is a 1024-dimensional space, M = 1024 as in Expression (1-2).

Subsequently, the discriminant function gij (x) obtained based on a plurality of image data acquired in advance is as shown in Expression (2).

Gij (x) in Expression (2) is an identification function for identifying category i and category j. Wij is a weighting vector learned in advance to identify category i and category j, and can be expressed as in equation (3). Category i and category j correspond to the horizontal axis or vertical axis of the matrix shown in FIG. Wij is created by the discriminant function creation unit 14a based on the learning information 13a.

  In this way, the feature vector X (Equation (1)) of the person attribute estimation target extracted from the captured image data is substituted into the discriminant function gij (x) (Equation (2)) obtained in advance. The identification values calculated for all categories are represented by a matrix as shown in FIG.

  If the discrimination value calculated by substituting the feature vector X into the discrimination function gij (x) is greater than 0, the feature vector X belongs to the category i, and if it is less than 0, the feature vector X belongs to the category i. Identify. In addition, the greater the absolute value of the identification value, the higher the probability of belonging to the identified category.

  For example, as shown in the figure, since the identification value at which the horizontal axis category B intersects with the vertical axis category A is “20”, the personal attribute estimation device 10 has an estimation target more than category A. This indicates that the possibility of belonging to category B is high.

  Further, as shown in the figure, since the identification value at which the horizontal axis category B and the vertical axis category C intersect is “−50”, the personal attribute estimation device 10 has the estimation target person from category B. Represents that the possibility of belonging to category C is high.

  On the other hand, as shown in the figure, the identification value at which the horizontal axis category B and the vertical axis category D intersect is “0”. This indicates that it was not possible to identify which of the two.

  As shown in the figure, for example, the value obtained by reversing the sign of the identification value where the horizontal axis category D and the vertical axis category C intersect is the horizontal axis category C and the vertical axis category D. It is an identification value that intersects.

  That is, the identification values are the same although the signs are opposite to each other symmetrically on the left and right and up and down across a diagonal line connecting cells intersecting vertically and horizontally between the same categories. For this reason, when calculating by substituting the feature vector X (formula (1)) for the discriminant function gij (x) (formula (2)), either one may be calculated with a diagonal line in between. For example, it is only necessary to calculate the six identification values shown in the upper right of the figure.

  In this way, the discrimination function is obtained in advance for all categories by discrimination analysis, and the discrimination value is calculated by substituting the feature vector of the person attribute estimation target person extracted from the captured image data.

  Next, details of the personal attribute estimation process performed by the personal attribute estimation device 10 will be described with reference to FIGS. First, the details of the gender degree calculation processing procedure performed by the personal attribute estimation device 10 will be described with reference to FIG. FIG. 4 is a diagram illustrating an example of a sex degree calculation processing procedure.

  As shown in the matrix of FIG. 5A, the identification value calculation unit 14c uses the identification function 13b obtained in advance for the sex and age-specific categories in the personal attribute estimation target extracted from the captured image data. The identification value is calculated by substituting the person's feature vector.

  Here, in the category of age in the table (A), “0, 10, 20, 40, 60,” is described, but the age group “0-9 years, 10-19 years, 20-39 years old, 40-59 years old, over 60 years old ".

  The corresponding age group is “10-year-old width, 10-year-old width, 20-year-old width, 20-year-old width, 60-year-old and older”, but it is not always necessary to categorize the categories with an equal width. The width may be in increments of 1 year.

  The personal attribute estimation device 10 adds the calculated identification values vertically, and obtains the total value of the identification values for each category on the horizontal axis (see (B) in the figure). Then, the added value is smoothed and a histogram is created for each gender (see (C) in the figure).

  Then, the personal attribute estimation device 10 sets the maximum frequency of the histogram created for each gender (hereinafter referred to as “age sex histogram”) as “gender degree” and the age at which the maximum frequency is “gender degree calculation age”.

  Specifically, as shown in FIG. 5C, when the maximum frequency of the male sex-sex histogram is “76” and the maximum frequency is “26.8 years old”, the male sex level ( Hereinafter, “male degree” is assumed to be “76”, and the male degree calculation age is “26.8”.

  On the other hand, when the maximum frequency of the age-gender histogram of the woman is “138” and the age at which the maximum frequency is “40.1 years”, the female gender (hereinafter referred to as “female”) is set to “138”. ”, The female age calculation age is“ 40.1 ”.

  Thereafter, when the difference between the male degree and the female degree exceeds a predetermined first threshold, the personal attribute determination unit 14d determines the gender having a high gender degree as the estimated gender. For example, when the difference 62 between the male degree 76 and the female degree 138 exceeds a predetermined first threshold, the estimated sex is determined to be female because the female degree is higher than the male degree. When the difference 62 between the male degree and the female degree does not exceed the predetermined first threshold value, the process proceeds to “detailed gender degree calculation process” described later. Here, the maximum frequency of the histogram is the gender degree, but the maximum sum of the identification values may be the gender degree.

  Here, it is difficult for the feature amount difference between categories to appear, that is, when the determination is made only with a value whose identification value is close to 0, the feature amount extracted from the image data of the individual attribute estimation target person is shaken. It is easy to be mistakenly estimated.

  However, since the personal attribute estimation device 10 performs the determination based on the total value of the identification values between all categories, even if the feature amount extracted from the image data fluctuates, The estimation accuracy can be improved.

  Next, details of the processing procedure of the “detailed gender degree calculation process” will be described with reference to FIG. FIG. 5 is a diagram illustrating an example of a detailed gender degree calculation processing procedure.

  As shown in FIG. 4A, among the identification values calculated for all categories by sex and age by the above-described method, the horizontal axis category is male and the vertical axis category is female. Is used to perform detailed gender degree calculation processing.

  When the estimated gender is not determined, the personal attribute determination unit 14d is surrounded by squares from the category corresponding to the male degree calculated age to the category corresponding to the female degree calculated age in both the vertical and horizontal directions. This area is the detailed gender degree calculation area.

  Specifically, the male degree calculated age 26.8 corresponds to the category “20˜”, and the female degree calculated age 40.1 corresponds to the category “40˜”. Accordingly, an area (in this case, a total of 4 squares) surrounded by the categories “20˜” to “40˜” on both the horizontal axis and the vertical axis is set as the detailed gender degree calculation area (see (B) in the figure). .

  Then, the personal attribute determination unit 14d calculates the average value of the identification values in the detailed gender degree calculation area to obtain the detailed gender degree. Here, the average value is calculated as −18.75 ((−6−33−33) ÷ 4 = −18.75), and −18.75 is set as the detail gender degree.

  Thereafter, the personal attribute determination unit 14d calculates the male / female sex ratio from the detailed sex grade as follows. If the detailed gender degree is a negative number, the probability of being a woman is high, and if the absolute value of the detailed gender degree is larger than a predetermined second threshold, the estimated gender is determined to be female. On the other hand, if the detail gender degree is a positive number, the probability of being male is high, and if the absolute value of the detail gender degree is greater than a predetermined second threshold, the estimated gender is determined to be male. .

  For example, the case where the detailed gender degree is -18.75 and the predetermined second threshold value is 20 is illustrated (see (C) in the figure). In this case, since the detailed gender degree is between −20 and 20, the estimated gender has not been determined yet.

  In this case, the ratio between the distance from the detailed gender degree to the negative predetermined second threshold −20 and the distance from the detailed gender degree to the positive predetermined second threshold 20 is referred to as “male / female sex ratio”. calculate. Then, the calculated “male / female ratio” is used as “weight (R)” when performing “age calculation processing” described later. Here, the male / female ratio is male: female = 3%: 97%, and the weight (R) is 0.03.

  Here, the detailed gender degree calculation area is an area surrounded by a range from a male degree calculation age to a female degree calculation age. However, an area including all adjacent 8 cells (9 cells in total) centering on a cell where the category corresponding to the maleity calculation age and the category corresponding to the femaleity calculation age intersect is used as the detailed gender degree calculation area. Also good.

  Here, among the identification values calculated for all categories, the detailed gender degree is calculated using an area where the horizontal axis category is male and the vertical axis category is female.

  However, an area where the horizontal axis category is female and the vertical axis category is male may be used. In this case, since the positive / negative of the identification value is reversed, the male / female sex ratio is also opposite to that described above.

  Next, the details of the above-described “age calculation process” will be described with reference to FIG. FIG. 6 is a diagram illustrating an example of an age calculation processing procedure.

  As shown in FIG. 5A, among the identification values calculated for all categories by sex and age by the method described above, the horizontal axis category is male and the vertical axis category is also male, The age calculation process is performed using an area where the horizontal axis category is female and the vertical axis category is also female.

  The personal attribute determination unit 14d calculates the identification values calculated for the area where the horizontal axis category is male and the vertical axis category is male, and the horizontal axis category is female and the vertical axis category is also female. By adding vertically, the total value of the identification values for each age category is obtained for each gender (see (B) in the figure).

  Then, the personal attribute determination unit 14d smoothes the added value to create a gender-specific histogram (see (C) in the figure), and a histogram created for each gender (hereinafter referred to as “age histogram”). The age at which the maximum frequency of is male and female age.

  Specifically, as shown in FIG. 5C, when the age that is the maximum frequency of the age histogram of men is 28.4 years old, the man age is set to “28.4”. On the other hand, when the age that is the maximum frequency in the age histogram of the woman is 41.4 years old, the woman age is set to “41.4”.

  Here, when the estimated gender is confirmed, the male age is used as the estimated age if the estimated gender is male, and the female age is used as it is if the estimated gender is female. On the other hand, when the estimated gender is not fixed, the male / female degree ratio (see FIG. 5C) is converted into a weight to calculate the estimated age.

  Specifically, the estimated age is an age obtained by adding a numerical value obtained by multiplying a male age by a male ratio of male to female ratio to a female age by a female ratio of male to female ratio. For example, if the male / female ratio is male: female = 3%: 97%, the estimated age is calculated based on the age histogram shown in FIG. 4C (41.01 (28.4 × 3% + 41). .4 × 97%). At this time, the final gender determination is a gender having a high male / female sex ratio or is indefinite.

  So far, the details of the personal attribute estimation processing procedure performed by the personal attribute estimation device 10 have been described using a specific matrix of identification values. Next, the personal attribute estimation processing procedure performed by the personal attribute estimation device 10 will be described. Will be described with reference to the flowcharts shown in FIGS.

  First, an outline of a processing procedure performed by the personal attribute estimation device 10 will be described with reference to FIG. FIG. 7 is a flowchart showing an outline of a processing procedure executed by the personal attribute estimation device 10.

  It should be noted that the discrimination function creating unit 14a obtains the discrimination function 13b for all categories based on the learning information 13a that is information obtained by collecting the feature amounts of image data whose personal attributes are known in advance.

  Then, as shown in the figure, the identification value calculation unit 14c calculates the identification value by inputting the feature amount extracted by the feature amount extraction unit 14b to the identification function 13b (step S101), and the personal attribute determination unit 14d calculates gender degree based on the identification value calculated in step S101 (step S102), and calculates age based on the identification value calculated in step S101 and the gender degree calculated in step S102 (step S103). ), The personal attribute estimation process is terminated.

  Next, the details of the “identification value calculation” procedure performed in step S101 shown in FIG. 7 will be described with reference to FIG. FIG. 8 is a flowchart showing the processing procedure of the identification value calculation processing.

  As shown in the figure, the imaging unit 11 captures the face of the individual attribute estimation target person and acquires image data (step S201), and the feature amount extraction unit 14b is based on the image data acquired in step S201. The feature amount is extracted (step S202).

  Then, the identification value calculation unit 14c calculates the identification value for all categories by substituting the feature amount extracted in step S202 into the identification function 13b obtained in advance (step S203), and creates a matrix of identification values. Then, the identification value calculation process ends.

  Next, details of the “gender degree calculation” procedure performed in step S102 shown in FIG. 7 will be described with reference to FIG. FIG. 9 is a flowchart showing a processing procedure of the gender degree calculation processing.

  As shown in the figure, the personal attribute determination unit 14d adds the identification values vertically in order to obtain the total value of the identification values for each category on the horizontal axis of the matrix of identification values (step S301).

  Then, the personal attribute determination unit 14d smoothes the total value in which the horizontal axis category is male among the values added in step S301, and creates an age-sex histogram (male) (step S302). The maximum frequency of the created histogram is assumed to be masculine (step S303).

  On the other hand, the personal attribute determination unit 14d smoothes the total value in which the horizontal axis category is female among the values added in step S301, and creates an age-sex histogram (female) (step S304). The maximum frequency of the created histogram is set to be female (step S305).

  Thereafter, the personal attribute determination unit 14d determines whether or not the difference between the male degree and the female degree exceeds a predetermined first threshold (step S306), and the difference between the male degree and the female degree is a predetermined first value. When it is determined that the threshold value has been exceeded (step S306: Yes), it is determined whether or not the male degree is greater than the female degree (step S307).

  If it is determined that the male degree is greater than the female degree (step S307: Yes), the personal attribute determination unit 14d determines that the estimated gender is male (step S308), and ends the gender degree calculation process.

  Further, when it is determined that the male degree is smaller than the female degree (step S307: No), the personal attribute determining unit 14d determines that the estimated gender is female (step S309), and ends the gender degree calculating process.

  On the other hand, when it is determined that the difference between the male degree and the female degree does not exceed the predetermined first threshold value (step S306: No), the personal attribute determination unit 14d performs a detailed gender degree calculation process (step S310). ), The gender degree calculation process is terminated. Details of step S310 will be described later with reference to FIG.

  Next, a detailed processing procedure of the detailed gender degree calculation processing shown in FIG. 9 will be described with reference to FIG. FIG. 10 is a flowchart showing the processing procedure of the detailed gender degree calculation processing.

  As shown in the figure, the personal attribute determination unit 14d sets the age that is the maximum frequency of the age-sex histogram (male) created in step S302 shown in FIG. 9 as a male degree calculation age (step S401).

  Also, the personal attribute determination unit 14d sets the age that is the maximum frequency in the age-sex histogram (female) created in step S304 shown in FIG. 9 as the female degree calculated age (step S402).

  Then, the personal attribute determination unit 14d compares the male degree calculated age calculated in the gender degree calculating process with the female degree calculated age, and determines whether the male degree calculated age is greater than the female degree calculated age (step S403). ).

  When it is determined that the male degree calculated age is greater than the female degree calculated age (step S403: Yes), the personal attribute determining unit 14d sets the minimum range value as the female degree calculated age (step S404), and sets the maximum range value as the male degree. The calculated age is set (step S405).

  On the other hand, if the individual attribute determination unit 14d determines that the male degree calculated age is smaller than the female degree calculated age (step S403: No), the minimum range value is set as the male degree calculated age (step S406), and the maximum range value is set. It is set as a female degree calculation age (step S407).

  Then, the personal attribute determination unit 14d calculates the average value of the identification values in the detailed gender degree calculation area from the category corresponding to the minimum range value to the category corresponding to the maximum range value, and sets the detailed gender level (step S408). Then, the sex degree calculation process is terminated.

  Next, the details of the “age calculation” procedure performed in step S103 shown in FIG. 7 will be described with reference to FIG. FIG. 11 is a flowchart illustrating a processing procedure of age calculation processing.

  As shown in the figure, the personal attribute determination unit 14d uses the identification value matrix where the horizontal axis is the male category and the vertical axis is the male category, and the horizontal axis is the female category and the vertical axis is the female category. Are added vertically (step S501).

  Then, the personal attribute determination unit 14d smoothes the total value in which the horizontal axis category is male among the values added in step S501 to create an age histogram (male) (step S502), and creates it in step S502. The age that is the maximum frequency of the histogram is set as male age (step S503).

  On the other hand, the personal attribute determination unit 14d smoothes the total value in which the horizontal axis category is female among the values added in step S501, creates an age histogram (female) (step S504), and creates it in step S504. The age corresponding to the maximum frequency of the histogram is set as the female age (step S505).

  Thereafter, the personal attribute determination unit 14d determines whether or not the estimated gender is already determined (step S506). If it is determined that the estimated gender is determined (step S506: Yes), the estimated gender is male. It is determined whether or not (step S507).

  Then, when it is determined that the estimated gender is male (step S507: Yes), the personal attribute determination unit 14d sets the estimated age as the male age calculated in step S503 (step S508), and ends the age calculation process.

  On the other hand, if it is determined that the estimated gender is female (step S507: No), the personal attribute determination unit 14d sets the estimated age as the female age calculated in step S505 (step S509) and ends the age calculation process.

  In addition, when it is determined that the estimated gender is not fixed (step S506: No), the personal attribute determination unit 14d performs a detailed age calculation process (step S510), and ends the age calculation process. Details of step S510 will be described later with reference to FIG.

  Next, a detailed processing procedure of the detailed age calculation processing shown in FIG. 11 will be described with reference to FIG. FIG. 12 is a flowchart showing the processing procedure of the detailed age calculation processing.

  As shown in the figure, the personal attribute determination unit 14d determines whether or not the detailed gender degree is larger than a predetermined second positive threshold (step S601), and the predetermined second gender degree is positive. When it is determined that it is smaller than the threshold value (step S601: No), it is determined whether the detailed gender degree is a predetermined second small negative value (step S602).

  When it is determined that the detail gender degree is greater than the predetermined negative second threshold (step S602: No), the personal attribute determination unit 14d converts the male ratio of the detailed gender degree into the weight (R) (step S602). S603).

  Then, the personal attribute determination unit 14d adds the numerical value obtained by multiplying the male age calculated in step S503 by the weight (R) and the numerical value obtained by multiplying the female age calculated in step S505 by (1-weight (R)). The detailed age calculation process is terminated with the value as the estimated age (step S604).

  Further, when it is determined that the detailed gender degree is smaller than the predetermined negative second threshold value (step S602: Yes), the personal attribute determination unit 14d sets the female age calculated in step S505 as the estimated age (step S605). The detailed age calculation process ends.

  On the other hand, if it is determined that the detailed gender degree is greater than the predetermined positive second threshold value (step S601: Yes), the personal attribute determination unit 14d sets the male age calculated in step S503 as the estimated age (step S606). The detailed age calculation process ends.

  As described above, in this embodiment, the personal attribute estimation device 10 acquires a plurality of image data in advance, and obtains an identification function 13b between all categories based on the feature amount extracted from the acquired image data. Keep it.

  And the personal attribute estimation apparatus 10 calculates an identification value between all categories by substituting the feature-value extracted from the image data of the personal attribute estimation object person for the identification function 13b. Furthermore, the personal attribute estimation device 10 estimates gender and age, which are personal attributes, based on the calculated identification values between all categories. By doing in this way, the estimation accuracy of a personal attribute can be improved.

  The personal attribute estimation device described in the claims is the personal attribute estimation device 10, the learned information storage means is the identification function 13b, the discrimination means is the identification value calculation section 14c, and the personal attribute estimation means is the personal attribute determination section 14d. It is equivalent.

  In the embodiment described above, the discriminant function creating unit 13a calculates the discriminant value by obtaining the discriminant function 13b based on the feature amount, substituting the feature amount extracted by the feature amount extracting unit 14b into the discriminant function 13b. However, the present invention is not limited to this.

  For example, the personal attribute estimation device 10 obtains an identification function 13b based on biological information such as image data, substitutes the biological information such as image data captured by the imaging unit 11 into the identification function 13b, and calculates an identification value. You may do it.

  In the above-described embodiment, the personal attribute estimation target person's face is imaged and the personal attribute is estimated. However, other physical features such as the eye retina capillary pattern and iris pattern The estimation may be performed based on a gray value or the like, or a voice receiving device may be provided to estimate the personal attribute based on the received voiceprint.

  Furthermore, personal attributes may be estimated based on various biological information such as behavioral characteristics such as blinking and lip movement. Thereby, it can classify | categorize into the category which can be identified with biometric information other than a face, and can estimate a personal attribute.

  Specifically, based on the pattern of capillaries in the retina of the eye and the shade value of the iris pattern, etc., the category is classified as a disease or not, and further, it is classified into a category such as what disease, Personal attributes can be estimated.

  In the above-described embodiment, the gender and the age are estimated as the personal attributes, but the present invention is not limited to this. The present invention may also be applied to a personal attribute estimation device that classifies into a relatively small number of multiple categories.

  For example, you may classify | categorize according to race or career years. Specifically, the category may be classified into two categories, Japanese and foreigners, or the category is classified into employees in companies such as newcomers, mid-level employees, and veterans, and personal attributes that are classified categories May be estimated. Further, when the personal attributes such as race, gender, and age are in three or more levels, the personal attributes can be estimated by similarly classifying the categories.

  In the above-described embodiment, a smoothed histogram is created in order to calculate a continuous value called age. However, when dealing with non-continuous categories such as the above-described race, the total value of the respective identification values. Should be used as is. In addition, the method for aggregating each identification value gives not only the total value of the identification values but also points such as “+1”, “0”, “−1” depending on whether the identification values are positive or negative, and sums them. May be.

  Further, gender may be estimated for individuals of a predetermined limited age group. Specifically, the whole body of the individual attribute estimation target person in their 20s may be imaged, and gender estimation may be performed based on feature quantities such as clothes and hairstyle.

  In this way, it is possible to obtain accurate estimated information and to derive an accurate analysis result from personal information estimated when the personal attribute estimation device is used for marketing, advertisement, or the like.

  As described above, the personal attribute estimation device according to the present invention is useful when it is desired to improve the estimation accuracy of the personal attribute when estimating the personal attribute based on the biological information acquired by the personal attribute estimation target person. Even when the feature amount extracted from the acquired biometric information fluctuates, it is suitable for the case where it is desired to improve the estimation accuracy of the personal attribute.

DESCRIPTION OF SYMBOLS 10 Personal attribute estimation apparatus 11 Imaging part 12 Result output part 13 Storage part 13a Learning information 13b Discrimination function 14 Control part 14a Discrimination function creation part 14b Feature-value extraction part 14c Discrimination value calculation part 14d Personal attribute determination part

Claims (7)

A personal attribute estimation device for estimating personal attributes of the subject from the biological information of the subject,
A criterion for determining which of the two different categories related to the personal attribute the biometric information of the subject belongs is based on learned information generated in advance from a plurality of the biometric information whose personal attributes are known Determining means for determining which of the two different categories the biometric information acquired from the subject belongs;
A personal attribute estimation unit for estimating the personal attribute of the target person based on a plurality of discrimination results discriminated by the discrimination unit for each combination of the two different categories. . The learned information is
A discriminant for determining which of the two different categories belongs to a predetermined formula for each combination of the two different categories,
The discrimination means includes
The personal attribute estimation apparatus according to claim 1, wherein a value calculated by inputting a feature amount of the biological information to the discriminant is output as the discrimination result. The personal attribute estimating means includes
3. The discrimination attribute in a combination of the category and a plurality of other categories is aggregated for each category, and the personal attribute of the target person is estimated based on the aggregated aggregation result. The personal attribute estimation device described. The personal attribute estimating means includes
The first personal attribute is estimated based on the discrimination result, and the second personal attribute is estimated based only on the combination of the personal attributes related to the estimated first personal attribute. Item 4. The personal attribute estimation device according to Item 1, 2 or 3. The personal attribute is
The personal attribute estimation device according to any one of claims 1 to 4, further comprising a gender and an age group of the subject. A personal attribute estimation method applied to a personal attribute estimation device that estimates a personal attribute of the subject from biological information of the subject,
Which of the two different categories the biometric information acquired from the subject belongs to is determined from among the two different categories related to the personal attribute based on a plurality of the biometric information whose personal attributes are known A discrimination step in which the personal attribute estimation device discriminates based on learned information generated in advance as a criterion for discriminating which category;
A personal attribute estimation step in which the personal attribute estimation device estimates the personal attribute of the subject based on a plurality of discrimination results discriminated by the discrimination step for each combination of the two different categories. And personal attribute estimation method. A personal attribute estimation system for estimating a personal attribute of the subject from biological information of the subject,
A criterion for determining which of the two different categories related to the personal attribute the biometric information of the subject belongs is based on learned information generated in advance from a plurality of the biometric information whose personal attributes are known Determining means for determining which of the two different categories the biometric information acquired from the subject belongs;
A personal attribute estimation system comprising: personal attribute estimation means for estimating the personal attribute of the subject based on a plurality of discrimination results discriminated by the discrimination means for each combination of the two different categories. .

Images