JP5287333B2 - Age estimation device - Google Patents

Description

  The present invention relates to an age estimation device for estimating age from a face image.

  A technique for estimating a person attribute such as age from a face image is known (see Patent Documents 1 and 2). However, when age estimation is performed by image processing, it is impossible to reach 100% estimation accuracy. Therefore, when the user uses the estimation result, it is not known how reliable the result is.

JP 2003-242486 A JP 2007-128262 A

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a technique for improving the convenience of the user who uses the age estimation result.

  In order to achieve the above object, the present invention adopts the following configuration.

The age estimation device according to the present invention includes a means for detecting a face from an image, a means for estimating the age of the detected face, a means for calculating an allowable range representing an error in the estimated estimated age , The estimated age and the allowable range
X ± Y (X: estimated age [age], ± Y: tolerance [age])
Output means for outputting in the format.

  The allowable range is preferably determined based on the reliability of the estimation process.

  The allowable range is preferably determined based on the estimated age.

  It is preferable that the estimated age and the allowable range are displayed in the vicinity of the face in the image.

  The present invention may be regarded as an age estimation device having at least a part of the above means, an age estimation method including at least a part of the above processing, or a program for realizing such a method and a program thereof. It can also be understood as a recorded recording medium. Each of the above means and processes can be combined with each other as much as possible to constitute the present invention.

  ADVANTAGE OF THE INVENTION According to this invention, the convenience of the user using an age estimation result can be improved.

It is a block diagram which shows the structure of the age estimation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the flow of a process of an age estimation apparatus. It is a figure explaining the age estimation method. It is a figure explaining an example of the reliability calculation method. It is a figure explaining an example of the reliability calculation method. It is a modification of the flowchart of FIG. It is a flowchart of an allowable range calculation process. It is an example of the tolerance | permissible_range data of a distribution chart form. It is an example of the tolerance | permissible_range data of a distribution chart form. It is an example of the tolerance | permissible_range data of a table format. It is a flowchart of a result integration process. It is a result display example. It is a result display example. 2 is an example of an estimated age distribution of Example 1. FIG. It is a figure explaining the reliability calculation process of Example 1. FIG. It is a figure explaining the reliability calculation process of Example 4. FIG. It is a figure explaining the reliability calculation process of Example 5. FIG. It is a figure explaining the reliability calculation process of Example 6. FIG. It is a figure explaining the reliability calculation process of Example 7. FIG. FIG. 10 is a diagram illustrating a reliability calculation process according to an eighth embodiment. It is a figure explaining the reliability correction process of Example 9. FIG.

  Exemplary embodiments of the present invention will be described in detail below with reference to the drawings.

<Device configuration>
FIG. 1 is a block diagram showing a configuration of an age estimating apparatus according to an embodiment of the present invention. For example, (1) Pachinko halls, game centers, movie theaters, devices for checking the age of persons in situations where age restrictions are imposed, such as the purchase of cigarettes (2) In a convenience store or a store, a device for collecting the age of customers as marketing information can be considered.

  The age estimation device includes an input unit 1, a face detection unit 2, a feature amount extraction unit 3, an attribute estimation unit 4, a reliability calculation unit 5, a reliability storage unit 6, an allowable range calculation unit 7, an allowable range storage unit 8, The result integration unit 9 and the output unit 10 are provided. These functions are realized by a computer executing a program and appropriately controlling hardware resources such as a memory, a camera, an input device, and a display device. The age estimation device may be configured as a single device (for example, a combination of a personal computer and a camera), or may be a built-in module of a device such as a vending machine, a mobile phone, a digital camera, or a printer. .

  The input unit 1 has a function of taking an image into the age estimation device. As the input unit 1, a camera including an optical system and an image sensor (CCD, CMOS sensor, etc.) can be preferably applied. However, an image may be captured from a network or an external storage medium. In that case, an interface with a network or an external storage medium corresponds to the input unit 1. The captured image is stored in the memory.

  The output unit 10 is a function for displaying the age estimation result. As the output unit 10, a display device such as a liquid crystal display or an EL display can be preferably applied.

  The reliability storage unit 6 and the allowable range storage unit 8 are configured from a nonvolatile storage medium such as a memory or a hard disk. The reliability storage unit 6 stores setting values and tables used for reliability calculation processing, and the allowable range storage unit 8 stores setting values and tables used for allowable range calculation processing. Yes.

<Processing flow>
FIG. 2 is a flowchart showing the flow of processing of the age estimation device. The processing described below is basically executed by a computer (program).

  (Step 1: Image Acquisition) First, an image to be processed is acquired from the input unit 1.

  (Step 2: Face Detection Processing) The face detection unit 2 detects a face from an image by image processing. Any technology of existing face detection processing may be applied to the face detection processing by the face detection unit 2. For example, a method for detecting a face by template matching using a reference template corresponding to the outline of the entire face, a method for detecting a face by template matching based on a facial organ (eyes, nose, ears, etc.), chroma key processing Detecting vertices such as the head, detecting a face based on the vertices, detecting a region close to the skin color, detecting that region as a face, and learning with a teacher signal using a neural network And a method for detecting a face-like area as a face.

  (Step 3: Face selection) When a plurality of faces are detected from the image, the attribute estimation and allowable range calculation described below is executed for each face.

  (Step 4: Feature Quantity Extraction) The feature quantity extraction unit 3 extracts a feature quantity from the face detected by the face detection unit 2. For example, feature points can be detected by edge detection, pattern matching, and the like, and gray values near the feature points, their periodicity / directivity, positional relationships of feature points, and the like can be extracted as feature amounts. The number (dimensions) of feature quantities can be arbitrarily set according to the expected estimation accuracy. In general, tens to hundreds of feature quantities are extracted.

  (Step 5: Attribute Estimation Processing) The attribute estimation unit 4 estimates the age of the face (person) based on the feature amount extracted from the face image.

  FIG. 3 is a diagram for explaining an example of the age estimation method. Discriminators are provided for each age group (class) in their 20s, 30s, 40s, and 50s. Each classifier is obtained by learning with a large number of sample face images, and has a distribution of feature amounts indicated by faces of each age class. In this embodiment, a normal distribution is used as the distribution of the feature amount, the average value is the median value of the age group (for example, 25 years of age in the 20s), and the variance value is an empirically determined value. As shown in FIG. 3, it can be seen that the distribution of the feature amount is different for each age class.

  When the feature quantity obtained in step 4 is input to the discriminator as an evaluation sample, the probability that the evaluation sample belongs to each age class is obtained. The probability that an evaluation sample belongs to an age class increases as the evaluation sample approaches the center of the feature quantity distribution of the class (the closer the distance between the evaluation sample and the class center). Note that the attribute estimation unit 4 may calculate the probability from the distance between the evaluation sample and the class center, or use the classifier that can output the class membership probability to directly determine the output of the classifier. It may be used as

  Then, the attribute estimation unit 4 determines the estimated age as shown in the equation of FIG. 3 from the probability of belonging to each class of the evaluation sample and the class-centered age of each class. In the example of FIG. 3, the age of the evaluation sample is estimated to be 42 years old.

  (Step 6: Attribute Estimation Reliability Calculation) The reliability calculation unit 5 calculates the reliability of the estimation process.

4 and 5 are diagrams for explaining an example of the reliability calculation method. (1) The reliability calculation unit 5 obtains a class having the maximum probability from among age classes. In the example of FIG. 4, the class in the 40s shows the maximum probability (40%). (2) Next, the reliability calculation unit 5 obtains the class having the higher probability among the classes adjacent to the class having the maximum probability. In the example of FIG. 4, a class of 30's (probability: 30%) is selected. (3) The reliability calculation unit 5 maps the total value of the probabilities of the two classes selected in (1) and (2) to the probability-reliability table shown in FIG. 5, and calculates the reliability. This probability-reliability table is set so that the higher the probability, the higher the reliability. Further, the lower limit value (40% in the example of FIG. 5) is set to an empirical value that the two selected classes may be at least as reliable as this. The reliability takes a value of 0 to 1000, and the higher the value, the higher the reliability. In the examples of FIGS. 4 and 5, a reliability of 500 is obtained for a total probability of 70%.

  The age estimation method and the reliability calculation method in FIGS. 3 to 5 are examples, and other methods may be used. For example, the class membership probability may be calculated by any method. It is also preferable to calculate the attribution probability using the variance value of the normal distribution. The conversion from probability to reliability may use a non-linear table instead of the linear table as shown in FIG. Further, the lower limit value of the reliability is not 40%, but may be other values, or the lower limit value may be dynamically determined. Further, although the reliability is calculated from the total probability of the two classes having a high probability, the reliability may be calculated based on the probability of the class having the highest probability (that is, the maximum probability).

  Furthermore, as shown in FIG. 6, it is also preferable to calculate not only the reliability of attribute estimation but also the reliability of feature quantities. The reliability of the feature amount represents the reliability of the feature amount value itself extracted from the face image. The reliability of the feature quantity itself can greatly affect the reliability of the estimation result. As a factor that affects the reliability of the feature amount, for example, there is a face direction in the image. The reliability of the face facing front is the highest, and the reliability decreases as the face turns sideways or up or down. Another factor is the shape of the facial organ. For example, if you close your eyes or laugh, your confidence level is low. Another factor is lighting conditions. If a part of the face is shaded or whitened, the reliability is low. Regarding the factors that may affect the reliability of these feature quantities, the presence or absence and the degree of the factors may be determined by image processing, and the reliability of the feature quantities may be determined according to the determination result. As a specific method, for example, the method disclosed in JP-A-2007-128262 can be preferably used.

  (Step 7: Allowable Range Calculation) The allowable range calculation unit 7 calculates an allowable range representing an error of the estimated age based on the reliability.

  FIG. 7 is a flowchart of the allowable range calculation process. The allowable range calculation unit 7 acquires the attribute estimation reliability result and the feature amount reliability result (the feature amount reliability is used only in the flow of FIG. 6), and the allowable range is stored from the allowable range storage unit 8. Read data. Then, an allowable range of the estimated age is calculated with reference to the allowable range data.

  FIG. 8 shows an example of allowable range data in a distribution diagram format. The relationship between the reliability and the allowable range (upper limit value and lower limit value) is defined. When the reliability is the highest (1000), the allowable range is ± 0 years old, and the allowable range increases as the reliability decreases. For example, when the reliability is 700, the allowable range is ± 3 years old.

  FIG. 9 shows a modification of the allowable range data in the distribution chart format. In FIG. 9, the distribution varies depending on whether the estimated age is less than 20 years old (the distribution range of 20 years old and over has a wider allowable range). This is based on the fact that the error range when humans estimate age differs depending on the age group. That is, for example, when determining the age of a 3-year-old infant, the error has an error width of about ± 2 years old at most. On the other hand, when determining the age of a 70-year-old elderly person, the age of judging from 60 to 80 years old is large depending on the person. Such an empirical rule of the person's judgment result is applied to the machine estimation result, and an allowable range is set according to the estimated age of the person to be estimated.

  FIG. 10 is an example of allowable range data in a table format. In the table of FIG. 10, the correspondence relationship between the estimated age and the allowable range (upper limit, lower limit) is defined, and the allowable range is determined based on the estimated age. For example, when the estimated age is 15 years, the allowable range is ± 4 years.

  (Step 8: Result Integration Process) The result integration unit 9 integrates the result (estimated age) obtained by attribute estimation and the allowable range calculation result, and calculates the final attribute estimation result (FIG. 11).

  (Step 9: Attribute estimation result output) The output unit 10 outputs the final attribute estimation result to the display device. FIG. 12 shows a result display example. Here, an allowable range (± 3 years old) is displayed together with the estimated age (25 years old) in the vicinity of the face in the image. FIG. 13 shows another result display example. In this example, the tolerance is shown as a percentage rather than an age.

<Summary>
According to the above configuration, the allowable range (± X years old or Y%) is displayed together with the estimated age. By doing in this way, the estimation result user can be given a room for studying the usage method. By displaying the allowable range (reliability of the estimated age), it is intuitively possible to know how much the result can be trusted, so that, for example, selection of an optimal camera installation position is facilitated. Further, an output method that is easy to understand for human beings of XX years ± YY is particularly useful when the user uses the estimation result. For example, when 27 years ± 5 years old is output, there is room for the user to select whether to enter the 25-year-old category or the 30-year-old category.

  That is, the advantages of the present embodiment can be summarized as follows: (1) By displaying the allowable range, it can be intuitively understood how much the result can be trusted. By applying the ambiguity that a person has when estimating the age of the person to the result output of the machine, a result display method that makes it easier for the person to understand is obtained. (2) In the embodiment described with reference to FIG. 6, since the output result of the reliability changes depending on the difficulty of the environment for estimating the age (the state of the face in the image), the allowable range is more preferable. Can be calculated. This is because there are a simple situation and a difficult situation (for example, variations in illumination and face orientation).

  The allowable range described above can be referred to as reliability information indicating the reliability of the estimated age itself. It can also be paraphrased as information indicating the degree of estimation error or information indicating the uncertainty (or certainty) of the estimated age. As a method for expressing the reliability of the estimated age, “42 years ± 5 years old”, “40 years old to 45 years old”, “42 years old ± 10%”, “42 years old (85% reliability)”, etc. Various expression methods may be adopted.

<Example>
Hereinafter, a specific example of a method for calculating the reliability (allowable range) of the estimated age will be described.

Example 1
The first embodiment is a method that uses a human estimation result.

  First, learning data is prepared. As learning data, data of samples (persons) of various ages, preferably data covering the entire age range to be estimated by this apparatus is prepared. Typically, image data of the face, the whole body, or the upper body is used, but not only image data but also other data that can affect age estimation such as audio data can be used. Also, a plurality of images with different angles, facial expressions, hairstyles, clothes, shooting conditions, etc. may be prepared for one sample.

Show these learning data to the subject and estimate the age of each sample. At this time, in order to eliminate the bias by the subjects, experiments with a plurality of subjects with different genders and ages are conducted.
It is recommended to perform age estimation multiple times for one sample.

  The estimated age distribution of each class (classified according to age in this embodiment) is calculated from the age estimation result obtained by the subject experiment. FIG. 14 shows distributions of a 10-year-old class, a 42-year-old class, and a 65-year-old class as an example of the estimated age distribution. It can be seen that the distribution shape (variation) of the estimated age differs for each class. In general, variation in estimated age for infants and boys is small, and variation in estimated age in the 40s to 50s tends to be large. The estimated age distribution of each class is stored in the allowable range storage unit 8. The data stored in the allowable range storage unit 8 may be data in which the distribution is expressed in the form of a histogram, or when an appropriate distribution such as a normal distribution is applied as shown in FIG. In the case of distribution, it may be dispersion).

  Next, a process for calculating the reliability of the estimated age using the estimated age distribution will be described. Here, a case where an estimation result of “42 years old” is obtained by the attribute estimation unit 4 is taken as an example.

  When receiving the estimated age “42 years old” from the attribute estimating unit 4, the allowable range calculating unit 7 reads the estimated age distribution of the class to which the age belongs from the allowable range storage unit 8. Next, as shown in FIG. 15A, the allowable range calculation unit 7 determines that the cumulative probability centered on “42 years old” in the estimated age distribution (the proportion of samples included in the hatched portion) is a predetermined threshold value α. Calculate the lower and upper limits when%. In the example of FIG. 15A, the lower limit “37 years” and the upper limit “47 years” are obtained, and the reliability (allowable range) of the estimated age “42 years” is obtained as “± 5 years”.

  When the threshold value α is constant, the allowable range changes according to the shape of the estimated age distribution. For example, if the distribution shape is wide as in the 42-year-old class shown in FIG. 15A, the allowable range is widened (the reliability is low). Further, if the distribution shape is steep as in the 5-year-old class shown in FIG. 15B, the allowable range is narrowed (reliability is increased). That is, the reliability of the age with many misestimations in the subject experiment for creating the estimated age distribution is low, and the reliability of the age with few misestimations is high. Thus, according to the method of the present embodiment, there is an advantage that the reliability reflecting the tendency of human estimation error can be obtained.

  The threshold value α may be set to an appropriate value according to the purpose of use of the apparatus and the required accuracy. Typically, it may be set to 1σ (about 68%). It is also preferable that a user using the apparatus can change the set value of α. The larger the value of α, the wider the allowable range. In the example of FIG. 15A, the shape of the estimated age distribution is symmetrical, so that the plus and minus errors are the same. However, if the estimated age distribution is not symmetrical, the plus error and the minus error are not necessarily the same value.

  In this embodiment, the estimated age distribution is obtained for each age, but it is not always necessary to obtain the distribution for all ages. For example, classify by age group, such as 0-3 years old, 4-9 years old, 10-19 years old, 20-34 years old, and so on, and obtain an estimated age distribution for each class (age group) Or just one distribution common to all ages. When dividing by age group, the age range is narrowed for ages with small estimation errors such as infancy and boyhood, and relatively large age estimation errors such as 40s and 50s. Should have a wide age range.

In this embodiment, as shown in FIG. 15A, the allowable range calculation unit 7 calculates the allowable range each time from the estimated age distribution. However, an allowable range for each class (for each age) is calculated in advance from the threshold α and the estimated age distribution of each class, and the value is stored in the allowable range storage unit 8 as an allowable range table (see FIG. 10). It is also preferable to keep it. Thereby, the allowable range can be obtained only by referring to the table, and the allowable range calculation process can be speeded up. When the user can change the threshold value α, a table may be prepared in advance for the plurality of threshold values α.

(Example 2)
The second embodiment is a method that uses an estimation result by a machine. In Example 1, the estimated age distribution was calculated from the results of the subject experiment, whereas Example 2 differs in that the age estimation result by the machine is used. Other configurations are the same as those of the first embodiment. Here, “machine” refers to a device that automatically estimates age from learning data. For example, the age estimation device according to the present invention can be used. In order to improve the reliability of estimation, it is also preferable to use a plurality of types of machines having different age estimation algorithms.

  When a machine is used, it is possible to process a large number of samples as compared with human estimation, so there is an advantage that the shape of the estimated age distribution is closer to the true value. In addition, when the age estimation device according to the present invention is used to create an estimated age distribution, the accuracy of reliability is improved. This is because the tendency of the estimation error of the age estimation device itself is reflected in the reliability.

(Example 3)
The third embodiment is a method using the score of the discriminator.

  The discriminator is a machine (program) that outputs, as a score, the probability (attribute probability) that a feature amount (evaluation data) extracted from an image belongs to a class. The attribute estimation unit 4 of this apparatus needs to perform multi-class identification according to the resolution of age estimation. Such multi-class identification may be realized by using a multi-class classifier, or may be realized by a combination of a plurality of two-class classifiers. Various classifiers such as SVM (support vector machine), AdaBoost (adaboost), LDA (linear discriminant analysis) have been proposed, and any classifier can be used in the present invention. Also, a conventionally known method can be adopted for the method of combining the two class discriminators. For example, the method of obtaining the attribution probability from the two-class classifier of each age and adopting the age of the classifier that output the highest attribution probability (called 1 vs All), or the first classifier over 30 years old If it is less than 30 years old, determine whether it is 15 years old or more with the second discriminator, and combine two class discriminators that bisect the age range as There is a method of gradually narrowing down the age range (called a tree structure).

  In the third embodiment, the belonging probability of the class selected as the correct answer (for example, the class having the largest belonging probability) is used as the reliability of the estimated age as it is. For example, in the example of FIG. 3, attribution probabilities of 10%, 30%, 40%, and 20% are output for four classes of the 20s, 30s, 40s, and 50s, respectively. The allowable range calculation unit 7 adopts “40%”, which is the highest attribution probability, as the reliability of the estimated age “42 years old”. A display example of the estimation result in this case is shown in FIG.

Example 4
The fourth embodiment is also a method using the score of the discriminator. In the third embodiment, the attribution probability is used as the reliability of the estimated age as it is, but in the fourth embodiment, the reliability is obtained from the attribution probability by referring to the table. Other configurations are the same as those of the third embodiment.

FIGS. 16A and 16B show an example of a table stored in the allowable range storage unit 8. FIG. 16A is an example of a table for converting the attribution probability into ± X years old format reliability, and FIG. 16B is a table for converting the attribution probability into Y% format reliability. It is an example. For example, when the age estimation result of age “42 years old” and attribution probability “80%” is obtained from the discriminator, the allowable range calculation unit 7 refers to the table in FIG. Gain confidence of 2 to +3 years old. Therefore, the output of the estimation result is "42 years old (-2 years old to +3 years old)" or "
40 to 45 years old ". If the attribution probability value is not in the table as “75%”, interpolation may be performed from neighboring values (for example, 70% and 80%).

  These tables may be obtained from learning data or may be created artificially. In the case of obtaining from learning data, as in the second embodiment, age estimation by this apparatus is performed using sample data of various ages, and the attribution probability and estimated age of each sample are obtained. And about the sample whose attribution probability was about 0.1 (for example, 0.05-0.15), the distribution of estimated age is calculated | required and the estimation error with respect to real age is calculated | required. For example, as in the first embodiment, a range in which the cumulative probability is the threshold value α, a range of 1σ, or the like can be selected as the estimation error. The lower limit value and upper limit value of the estimation error thus obtained are set to an allowable range corresponding to the attribution probability 0.1. Similarly, the permissible range corresponding to attribution probabilities 0.2, 0.3,..., 0.9 can be obtained. In general, when the attribution probability is low, the error in the estimated age becomes large, so the allowable range becomes large. Conversely, the higher the attribution probability, the smaller the allowable range.

  It is also preferable to use a different table for each class. This is because the estimation error varies depending on the age as described above.

  Further, not only the belonging probability of one class but also the reliability can be determined based on the belonging probability of a plurality of classes as shown in FIG. In FIG. 16C, a set of N class membership probabilities is associated with a reliability (allowable range). This table may be obtained from the learning data or may be created artificially. Note that it is not necessary to use the belonging probabilities of all classes in determining the reliability. For example, two classes may be selected in descending order of belonging probability, and the reliability may be obtained from a set of belonging probabilities of the two classes.

(Example 5)
The fifth embodiment is also a method for obtaining the reliability from the belonging probability which is the score of the discriminator. In the fourth embodiment, a table is used, but in the fifth embodiment, a mapping is used to convert the attribution probability to the reliability. Other configurations are the same as those in the fourth embodiment.

  FIGS. 17A and 17B show examples of mappings stored in the allowable range storage unit 8. FIG. 17A is an example of a map for converting the attribution probability into ± X-year format reliability, and FIG. 17B is a map for converting the attribution probability into Y% format reliability. It is an example. For example, when an age estimation result of age “42 years old” and attribution probability “80%” is obtained from the discriminator, the allowable range calculation unit 7 refers to the mapping of FIG. Gain confidence of 3 years old.

  Although FIG. 17 shows a linear mapping, a non-linear or discontinuous mapping may be used. As a non-linear mapping, a sigmoid function can be considered. These maps may be obtained from learning data or may be created artificially. In the case of obtaining from learning data, the relationship between the attribution probability and the estimation error may be obtained in the same manner as in the fourth embodiment, and an appropriate function may be applied thereto. It is also preferable to use different mappings for each class.

  In addition, the reliability can be determined based on not only the probability of belonging to one class but also the probability of belonging to a plurality of classes. In that case, a mapping representing the relationship between a set of N membership probabilities (N-dimensional vector) and reliability may be used. Alternatively, N attribution probabilities can be converted into one feature value (scalar), and a map representing the relationship between the feature value and the reliability can be used.

(Example 6)
The sixth embodiment is also a method for obtaining the reliability from the score of the discriminator. However, while the belonging probabilities obtained as a result of multi-class identification are used in the third to fifth embodiments, the sixth embodiment is different in that the reliability is calculated from the score obtained from the two-class classifier.

  First, for each class (which may be age or age), the distribution of the identification score is obtained from the learning data. For example, for a 42-year-old 2-class classifier, a large number of learning data consisting of 42-year-old samples is identified, and the relationship of “identification score-frequency” as shown in FIG. 18 is obtained. In this score distribution, the horizontal axis indicates the value of the identification score, and the vertical axis indicates the frequency of the sample having the identification score (a value normalized so that the maximum frequency is 100%). The shape of the score distribution is stored in the reliability storage unit 6 in the form of a histogram or distribution parameters (in the case of the shape shown in FIG. 18, a sigmoid function or the like can be used). Note that the score y of the two-class classifier takes a value between −1 and 1, and takes a value closer to 1 as the possibility that the sample belongs to the class is higher, and becomes −1 as the possibility that the sample does not belong is higher. A value close to is assumed.

  Next, a process for calculating the reliability of the estimated age will be described. Here, a case where an estimation result “42 years old” is obtained by the attribute estimation unit 4, that is, a case where the 42-year-old 2-class classifier outputs the largest identification score will be described as an example. When the reliability calculation unit 5 receives the estimated age “42 years old” and the identification score (for example, 0.7) from the attribute estimation unit 4, it reads out the score distribution data of the class of the age from the reliability storage unit 6. . Next, the reliability calculation part 5 calculates | requires the frequency corresponding to the identification score 0.7 with reference to score distribution, and makes the frequency the reliability of the said estimation age. In the example of FIG. 18, the reliability of the estimated age of 42 is 90%.

  Note that a table, mapping (see FIG. 8), or the like can be used to convert the reliability in the Y% format into the ± X format. When combining two class discriminators, it is possible to use only the reliability obtained by the discriminator of the finally selected class, but through the discrimination results of a plurality of discriminators like a tree structure. If the final class is narrowed down, it is considered that multiple classifiers are affecting the estimation results, so the product or sum of the reliability of all or some of the multiple classifiers included in the path is calculated. It is also preferred to use to calculate the final confidence.

(Example 7)
In the seventh embodiment, the k-NN method is used for calculation of reliability.

  FIG. 19 shows the distribution of each class in the feature amount space (in this case, the classes are classified by age, such as 20's, 30's, and 40's). This distribution is obtained in advance from learning data consisting of a large number of samples.

  In the estimation process, the reliability calculation unit 5 maps the feature quantity of the estimation target to the feature quantity space of FIG. 19 and extracts sample points existing in the vicinity of the estimation target data. At this time, if the setting is such that k neighboring sample points are extracted, k sample points are extracted in order from the closest distance from the estimation target data. Or if it is the setting which extracts the sample point of the near space of the radius r, all the sample points which exist in the radius r centering on estimation object data will be extracted. Note that the values of k and r can be set to arbitrary values.

Next, the reliability calculation unit 5 obtains the number of sample points belonging to the same class as the estimation target data (class in the thirties in FIG. 19) among the extracted sample points. When the extracted sample points are A, of which B sample points are of the same class as the estimation target, the reliability calculation unit 5
Reliability (%) = B / A × 100
The reliability is calculated as follows. For example, when A = 11 and B = 9, the reliability is about 82%. In order to convert the reliability in the Y% format to the ± X format, a table, mapping, or the like can be used.

(Example 8)
The eighth embodiment is a method for calculating the reliability of each class by regression mapping.

  As shown in FIG. 20A, a map f that outputs an estimated age y with respect to a feature value x is learned using learning data composed of a large number of samples. A general regression method such as multiple regression, CCA (canonical correlation analysis), Ridge regression, or RVM (Relevance Vector Machine) can be used for learning the map f. In addition, since several tens to several hundreds of feature quantities are used in age estimation, the feature quantity x is a multidimensional vector, but in FIG. 20, it is shown in one dimension for the sake of simplicity.

  Next, the reliability (upper limit / lower limit of the allowable range) of each age is calculated from the learning data and the map f. For example, when obtaining the reliability for β-year-old, as shown in FIG. 20A, age estimation is performed for each of the β-year-old learning data by mapping f, and the range of variation of these estimation results (FIG. 20 ( In (a), β-q years to β + p years) is defined as the confidence level of β years. That is, the learning data having the largest estimated age is the upper limit of the allowable range, and the learning data having the smallest estimated age is the lower limit of the allowable range. Confidence for other ages is calculated in the same way.

  FIG. 20B shows a modified example of the reliability calculation. In this example, age estimation by mapping f is performed using a part of the learning data of β years old (data within the range of threshold value α% from the center of distribution in the distribution of feature value x). The range of variation (β-t years old to β + s years old) is defined as the confidence level of β years old. According to this method, since outliers of the learning data are excluded, more appropriate reliability can be obtained.

Example 9
In the ninth embodiment, when the photographing condition of the face / human body or the photographing condition of the image is bad, the process of lowering the reliability of age estimation is performed. This is because, when an image with poor shooting conditions is used, the reliability of the obtained feature value is lowered, and as a result, the reliability of the age estimation result is lowered. Examples of factors that reduce the reliability of the feature amount include the following.

-Size of face and human body: When the size of the face or human body in the image is small, the amount of information that can be extracted from the face or human body is small, so that the age estimation accuracy may be lowered.
-Face or human body direction: When the face is facing the front, the facial organ can be detected and the feature quantity can be accurately extracted, so that the age estimation accuracy is improved. However, when the face is facing sideways or facing up or down, face organs may be hidden and the accuracy of age estimation may be reduced.
-Organ reliability: When the facial organ is deformed or the facial organ is hidden, the organ reliability in the facial organ detection process decreases. In that case, the accuracy of age estimation may be reduced because the accuracy of the position of the facial organs and the information contained in the part is low.
Shadows and whiteout: When a face image is unclear due to shadows or whiteout, the accuracy of age estimation may be reduced because the accuracy of face organ detection and feature quantity extraction is reduced.
-Image blur: If the blur amount of the image is large, the accuracy of detecting the facial organs and extracting the feature amount is lowered, so that the age estimation accuracy may be lowered.
-Lighting conditions: When the lighting is too dark, too bright, or outdoors (not indoors).
Noise: If there is a lot of noise in the image, the accuracy of age estimation may be reduced because the accuracy of face organ detection and feature quantity extraction is reduced.

  In this embodiment, a table in which reliability is set for a plurality of shooting conditions is prepared in advance. For example, the reliability for the estimated age of 42 years is ± 4 years in the case of the photographing condition “lighting condition: bright indoor, face size: large, face direction: front”, “lighting condition: outdoor, face direction: right In the case of the shooting condition of “downward”, it is set in advance such as ± 7 years old.

  When an image to be estimated is given, all or a part of the above factors are evaluated. This evaluation may be automatically performed by image processing, or a person may input the evaluation. When the evaluation result is obtained, the corresponding reliability is obtained by referring to the table. As a result, it is possible to obtain reliability that reflects the imaging conditions.

  In addition, a configuration in which the reliability calculated by the reliability calculation unit 5 is corrected according to the evaluation result of the imaging condition, instead of calculating the reliability independently. For example, as shown in FIG. 21, the reliability can be corrected according to the interval between eyes. In the example of FIG. 21, when the eye interval is 50 pixels or more, the face size is sufficiently large and reliability correction is not performed. Reduce reliability. When the interval between the eyes is 20 pixels, the reliability in the ± X-year format is ± (X + 3) years, and the reliability in the Y% format is (Y × 0.01)%. Similarly, when the face direction is the front (angle = 0 degree), the reliability is not corrected, and the correction can be performed such that the reliability gradually decreases when the angle exceeds a predetermined value.

  As described above, specific examples of calculation of the reliability (allowable range) of the present invention have been described by giving a plurality of examples. However, the present invention is not limited to these examples, and within the scope of the technical idea. It can be modified as appropriate. For example, the configurations of the above-described embodiments can be appropriately combined.

Claims (4)

Means for detecting a face from an image;
Means for estimating the age of the detected face;
Means for calculating an allowable range representing an error in the estimated estimated age;
The estimated age and the allowable range
X ± Y (X: estimated age [age], ± Y: tolerance [age])
Output means for outputting in the form of
An age estimation device comprising: The age estimation apparatus according to claim 1 , wherein the allowable range is determined based on reliability of estimation processing. The allowable range, age estimation apparatus according to claim 1 or 2, characterized in that it is determined on the basis of the estimated age. The age estimation device according to any one of claims 1 to 3 , wherein the estimated age and the allowable range are displayed in the vicinity of a face in the image.

Images