JP6397439B2 - Attribute estimation device - Google Patents

Description

  The present invention relates to an attribute estimation apparatus that estimates an attribute such as age from information such as an image related to a target person.

  When estimating an attribute expressed as a quantity representing a target feature, for example, when estimating the age of a person, a face image of the person is acquired and input to an age estimation means constructed by a learning process. Obtaining techniques have been proposed.

  In Patent Document 1, there are a plurality of feature quantities used for estimation, and the age estimation means using each feature quantity is composed of a model that estimates an age group in advance by a learning process. A technique for outputting and determining an age group by majority vote is disclosed. It is also disclosed that the age group may be divided when it is desired to estimate a more detailed age group.

JP 2004-318632 A

  As in Patent Literature 1, in order to obtain the age group-like character of the person of the input face image, for each age group, classifiers that make the age group correct and other age groups incorrect are classified for each age group. The classifier group prepared in (1) is constructed.

  Here, when the age group to be obtained as a result is divided as disclosed in Patent Document 1, in the learning process of the classifier of each age group, face images labeled as correct answers are reduced along with the division. In this case, since learning processing is performed for a small number of correct face images, it is closer to learning the facial features of a person in the face image rather than learning the facial features of the person in that age group. Then, for example, the face of a person whose age is completely different from the person of the correct face image, such as when one of the two parents is the correct face image and the other is the input face image. When an image is input, it may be recognized as a completely different age group.

  In order to avoid such a situation, it is sufficient to prepare a sufficient number of correct face images for each divided age group, but in reality it is difficult to collect face images for learning. It is. In particular, it is difficult to collect a lot of correct face images.

  Therefore, the present invention has been made in view of the above problems, and an object of the present invention is to provide an attribute estimation apparatus capable of estimating attributes in fine steps without increasing the number of data used for learning.

One aspect of the present invention is an information acquisition unit that acquires information about a target whose attribute is to be estimated; an attribute feature amount calculation unit that determines an attribute feature amount related to an attribute that is divided into a plurality of attribute categories regarding the target from the information; The attribute feature amount using a plurality of discriminators including a set of discriminators having an overlapping section which is an attribute section in which the learned classification sections overlap with each other, learning features of different classification sections in the attribute The attribute feature quantity overlaps by integrating the attribute likelihood obtained in the set of the classifier and attribute likelihood calculating means for obtaining the attribute likelihood representing the degree of each characteristic of the classification category An integrated estimator that obtains an overlap classification likelihood that represents a degree having a characteristic of the section and estimates the attribute classification to which the object belongs using at least the overlap classification likelihood When, with a set of classifiers, a classifier belonging to the first segment learned characteristics of each of the identified segments of predetermined ranges assigned in the first segment of the attributes, the predetermined range A classifier belonging to a second section that has learned the characteristics of each of the identification sections allocated in a different number of second sections from the first section, and the integrated estimation means includes: The overlap likelihood is obtained by integrating the attribute likelihood obtained in the classifier belonging to the category and the attribute likelihood obtained in the classifier belonging to the second category. It is an attribute estimation device.

  Moreover, it is preferable that the set of classifiers includes classifiers that have learned the characteristics of the classification sections having different widths.

  In addition, it is preferable that the integrated estimation unit obtains the overlapping category likelihood by integrating the attribute likelihood obtained for the identification category with a larger weight as the range of the identification category is wider. is there.

  Moreover, it is preferable that the said attribute is any one of a person's age or a person's hair length.

  According to the present invention, it is possible to provide an attribute estimation device that estimates an attribute in fine steps without increasing the number of data used for learning.

It is a figure explaining the outline | summary of the attribute identification process in embodiment of this invention. It is a figure explaining the integration process of likelihood distribution in embodiment of this invention. It is a figure which shows the structure of the age estimation apparatus in 1st Embodiment. It is a figure which shows the age group estimation model in 1st Embodiment. It is a figure explaining the feature-value of the face image in 1st Embodiment. It is a figure explaining the integrated process of the age group likelihood distribution and generation likelihood distribution in 1st Embodiment. It is a figure explaining another example of composition of a discriminator in a 1st embodiment. It is a figure which shows the relationship between the hair length and attribute division in 2nd Embodiment. It is a figure which shows the hair length estimation model in 2nd Embodiment. It is a figure explaining the feature-value of the hair length in 2nd Embodiment. It is a figure which shows the estimation model in the modification of this invention.

[Outline of the Invention]
Before giving an example of specific human attributes, the basic idea of the present invention will be described. FIG. 1A schematically shows a classifier group according to the present invention.

  In the present invention, for an attribute indicating one of the characteristics to be estimated, it is estimated where it is likely to belong to the attribute classification into which the attribute is classified. In the present invention, the attribute is estimated by two or more classifier groups. Each classifier group has classifiers for a predetermined number of attribute categories.

  The first classification classifier group 11 includes seven classification classification classifiers A to G. The second classification category classifier group 12 includes six classification category identifiers H to M.

  The identification classification classifier A included in the first identification classification classifier group 11 is constructed by learning processing with data having attribute classifications “0” and “1” as correct answers and data in other attribute classifications as incorrect answers. ing. The identification classification discriminator B is constructed by learning processing with data having attribute classifications “2” and “3” as correct answers and data of other attribute classifications as incorrect answers. In the same manner, the classification classifier G is constructed by a learning process with data having attribute classifications “12” and “13” as correct answers and data of other attribute classifications as incorrect answers.

  In addition, the classification classifier H included in the second classification classifier group 12 performs a learning process with data having attribute classifications “1” and “2” as correct answers and data in other attribute classifications as incorrect answers. Has been built. The discriminating classifier I is constructed by learning processing with the data having the attribute classifications “3” and “4” as correct answers and the data of other attribute classifications as incorrect answers. In the same manner, the identification classification discriminator M is constructed by a learning process with data having attribute classifications “11” and “12” as correct answers and data of other attribute classifications as incorrect answers.

  As described above, the first identification classification classifier group 11 and the second identification classification classifier group 12 are classified such that there are sections in which the attribute classifications overlap (hereinafter referred to as overlapping classifications). It is constructed by a learning process using learning data prepared as correct answers for data belonging to a category. For example, the attribute classification of the identification classification identifier A of the first identification classification identifier group 11 is “0” and “1”, and the attribute classification of the identification classification identifier H of the second identification classification identifier group 12 is “1”. "And" 2 ", and attribute classification" 1 "is an overlapping classification.

  And as shown in FIG.1 (b), an input image is input into the identification classification discriminator contained in the 1st identification classification discriminator group 11 and the 2nd identification classification discriminator group 12, and each identification classification | category is input about an input image. The discriminator learns the characteristics of the data and arranges the scores that are output as representing the likelihood of the attribute classification (hereinafter referred to as the identification classification), and obtains the normalized first identification classification likelihood distribution and second identification classification likelihood distribution. . Thereafter, both are integrated to obtain an integrated likelihood distribution.

  The integration of the first discrimination category likelihood distribution and the second discrimination category likelihood distribution will be described with reference to FIG. FIG. 2 schematically shows a first discrimination category likelihood distribution 30 and a second discrimination category likelihood distribution 31 when the attribute category of the input data is “5”.

  Since the correct attribute category of the input data is “5”, the likelihood output by the discrimination category discriminator D is high. However, on the other hand, the likelihood output by the classification category discriminator F also shows a high value, which is higher than the likelihood output by the classification category identification device C. If an identification classification classifier included only in the first identification classification classifier group 11 is used as in the past, the input data is determined to be the attribute classification “10” or “11” which is the identification classification indicating the highest likelihood. Is done. Actually, since the correct attribute classification of the input data is “5”, this is an error. On the other hand, referring to the second classification category likelihood distribution 31, the classification category classifier J shows the highest likelihood.

  Therefore, in the present invention, the integrated likelihood distribution 32 is obtained by integrating the first identification category likelihood distribution 30 and the second identification category likelihood distribution 31 for each attribute category. For example, in the integration method, a normalization process is performed in which an average of likelihood is obtained for each attribute section and divided by the sum of the averages. The normalization process may be omitted. As a result of this integration processing, the likelihoods of the attribute categories “10” and “11” that the discrimination category discriminator F makes correct are greatly reduced, and the likelihood of the attribute category “5” that is correct is the highest, and the input data Is correctly obtained as “5”.

  That is, in the present invention, by preparing a plurality of discriminating class discriminator groups each including discriminators that have learned discriminating classifications including overlapping segments, the attribute classification obtained in the discriminating class discriminator group is integrated to thereby overlap the classification The overlap classification likelihood representing the degree of having the above feature is obtained. Then, the attribute category to which the object belongs is estimated using this overlap category likelihood. Thereby, it is possible to prevent the estimation result from being erroneous by calculating the maximum likelihood in the attribute classification other than the correct answer for some reason.

[First embodiment (age estimation apparatus)]
As a first embodiment, a case where the present invention is applied to an age estimation apparatus will be described. That is, the object is a person, the attribute is age, and the attribute classification is an age group. Moreover, the 1st classification division classifier group 11 shown in FIG. 1 becomes a 1st age group identification device group, and the 2nd identification classification classifier group 12 becomes a 2nd age group identification device group. The target information is a face image.

  FIG. 3 shows a block diagram of the age estimating apparatus 100. The age estimation device 100 is connected to an imaging device 50 for acquiring an input image, in this embodiment a human face image, and is configured to output an estimation result to an external device (not shown).

  The imaging device 50 is a so-called camera and includes an imaging element such as a CCD or C-MOS and a lens. The resolution of the imaging device 50 may be determined according to the attribute to be estimated in the age estimation device 100, the installation environment of the imaging device 50, the processing capability of the estimation processing unit 120, and the like, for example, 1600 × 1200 pixels. The imaging device 50 captures an image of the face or head of a person whose age is to be estimated, and inputs the image to the age estimation device 100 as an input image.

  The age estimation device 100 is composed of a CPU / MPU, various memories, peripheral circuits, and the like, extracts a face image from an input image acquired by the imaging device 50, estimates the age of a person shown in the image, and outputs it. . The age estimation apparatus 100 includes an image acquisition unit 110, an estimation processing unit 120, a storage unit 130, and an output unit 140.

  The image acquisition unit 110 that operates as an information acquisition unit is an interface circuit connected to the imaging device 50, acquires an input image captured by the imaging device 50, and outputs the input image to the estimation processing unit 120. Instead of the imaging device 50, an input image captured in the past may be acquired from a storage medium such as a hard disk. Moreover, you may acquire the input image transmitted from the external device via the network.

  The output unit 140 is an interface for outputting the age group estimated for the person in the input image by the estimation processing unit 120 to an external device (not shown) and its control circuit. The external device may be a display device (monitor) or a storage device for storing a separate estimation history. Alternatively, the estimated age group may be transmitted to the server machine via the network.

  The storage unit 130 includes a large-scale storage unit such as a semiconductor memory such as a ROM and a RAM, a hard disk, and the like, and an estimation processing unit realized by a program (not shown) for controlling the age estimation apparatus 100 and a program module. 120 programs (not shown) and various data are stored, and information is input to and output from the estimation processing unit 120. Various data includes at least an age group estimation model 131.

  FIG. 4 schematically shows the age group estimation model 131. The age group estimation model 131 includes a first age group classifier group 132, a second age group classifier group 133, and a generation classifier group 134.

  The first age group discriminator group 132 includes seven age group discriminators A, age group discriminators B... The age group discriminator A is constructed by a learning process in which a face image of a person corresponding to an age classification of 10 to 19 years is correct, and other age groups are incorrect. The age group discriminator B is constructed by a learning process in which the face image of a person corresponding to the classification category of ages 20 to 29 is correct and the other age groups are incorrect. Similarly, the age classifier C to the age group classifier G are those of persons corresponding to the classification categories of ages 30 to 39, 40 to 49, 50 to 59, 60 to 69, and 70 to 79, respectively. It is constructed by a learning process in which the face image is correct and the others are incorrect.

  The second age group discriminator group 133 is composed of six of an age group discriminator H, an age layer discriminator I... The age group discriminator H is constructed by a learning process in which the face image of a person corresponding to the classification category of ages 15 to 24 is correct and the other age groups are incorrect. The age group discriminator I is constructed by a learning process in which the face image of the person corresponding to the classification category of the ages 25 to 34 is correct and the other age groups are incorrect. Similarly, the correct images of the face images of the persons corresponding to the classification categories of the ages 35 to 44, 45 to 54, 55 to 64, and 65 to 74, from the age group classifier J to the age group classifier M, It is constructed by a learning process in which other than that is incorrect.

  Furthermore, the age group estimation model 131 in the present embodiment also includes a generation discriminator group 134. The generation classifier group 134 includes a young generation classifier N, a grand generation classifier O, and an old age classifier P. The young generation classifier N is constructed by a learning process in which a face image of a person corresponding to a classification category of 10 to 34 years of age is correct, and other age groups are incorrect. The senior generation classifier O is constructed by learning processing with a face image corresponding to a person in the classification category of ages 35-54 as the correct answer and other age groups as incorrect answers. The old age generation classifier P is constructed by learning processing with a face image corresponding to a person in the classification category of ages 55 to 80 as a correct answer and other age groups as incorrect answers. Compared with the first age group classifier group 132 and the second age group classifier group 133, the generation classifier group 134 has a wider range of correct age groups and is copied to the input image according to each name. Outputs the generational quality of a person.

  In addition to the first age group discriminator group 132 and the second age group discriminator group 133, the generation discriminator group 134 is also used in combination, so that the age group is completely different, like a parent and child whose faces are very similar, but with a face. Even when there is a possibility of outputting wrong results due to common characteristics, the generation discriminator 134 first identifies a correct generation roughly, and additionally estimates the age group to improve the estimation accuracy. be able to.

  The estimation processing unit 120 includes a so-called computer such as a memory, a peripheral circuit thereof, and a microprocessor thereof, executes various processes on the input image input from the image acquisition unit 110, and outputs the processing result to the output unit 140. To do.

  For this purpose, the estimation processing unit 120 includes a feature amount calculation unit 121, a first age group likelihood calculation unit 122, a second age group likelihood calculation unit 123, a generation likelihood calculation unit 124, and an integrated estimation unit 125. Each means is a functional module realized by software operating on a microprocessor.

  The feature amount calculation unit 121 acquires an input image from the image acquisition unit 110, extracts a face image, and then calculates an attribute feature amount for estimating an age group. A known technique may be used to extract a face image from an input image. For example, an oval shape that looks like a face may be detected after performing a background difference, or an identifier that has learned facial features may be prepared to sequentially scan the input image to detect the shape that looks like a face.

  The attribute feature amount is not particularly limited. For example, as shown in FIG. 5, after extracting a feature point 301 from the face image 300, a Haar-Like feature value 302, a distance 303 between feature points, a Gabor feature value 304, a HOG (Histogram of) What is necessary is just to make it the feature-value vector which was suitably selected from Oriented Gradient) feature-value 305 and LBP (Local Binary Pattern) feature-value (not shown) as a component.

  The feature amount calculating unit 121 outputs the calculated attribute feature amount to the first age group likelihood calculating unit 122, the second age group likelihood calculating unit 123, and the generation likelihood calculating unit 124.

  The first age group likelihood calculating means 122 reads the first age group discriminator group 132 stored in the storage unit 130 using the attribute feature quantity, calculates and arranges the likelihood for each age group, Find the normalized likelihood distribution. For example, using the age group discriminator A of the first age group discriminator group 132, the likelihood that the attribute feature amount represents the likelihood of the person of the input face image being 10-19 years old is obtained. Similarly, the likelihoods representing the likelihood of each age group are obtained by the age group discriminators B to G. As a result, as shown in FIG. 6, a first age group likelihood distribution 400 obtained by normalizing the likelihood obtained by the age group discriminator A to the age group discriminator G is obtained. FIG. 6 shows the result of obtaining the likelihood for each identification category for the face image of a person whose actual age is 39 years.

  Similarly, the second age group likelihood calculation means 123 uses the age group discriminator H to the age group discriminator M included in the second age group discriminator group 133, and the attribute feature amount indicates the likelihood of each age group. Find the likelihood to represent. Thereby, as shown in FIG. 6, a second age layer likelihood distribution 410 obtained by normalizing the likelihood obtained by the age group discriminator H to the age group discriminator M is obtained.

  Similarly, the generation likelihood calculation means 124 uses the young generation discriminator N, the grand generation discriminator O, and the old age discriminator P included in the generation discriminator group 134, and the attribute feature amount represents the generation likelihood. Find the likelihood. As a result, as shown in FIG. 6, a generation likelihood distribution 420 obtained by normalizing the likelihoods obtained by the young generation classifier N, the grand generation classifier O, and the old age classifier P is obtained.

  The integrated estimation unit 125 integrates the first age group likelihood distribution 400, the second age group likelihood distribution 410, and the generation likelihood distribution 420, and inputs the age group that is the attribute category indicating the highest likelihood as the input face. Estimated as the age group of the person in the image. The integration method is the same as the method described in the outline description of the invention.

  The integrated estimation unit 125 determines the integrated likelihood distribution 430 by integrating the first age group likelihood distribution 400, the second age group likelihood distribution 410, and the generation likelihood distribution 420 for each attribute category. For example, the integration method obtains the average of the likelihoods of the first age group likelihood distribution 400, the second age group likelihood distribution 410, and the generation likelihood distribution 420 for each attribute category, and divides by the sum of all the averages. Normalization is performed to obtain integrated likelihood. The normalization process may be omitted.

  In the integration process performed by the integration estimation unit 125, a weight may be given such that the intermediate value of the identification classification is the highest value, and the value decreases as the boundary of the identification classification is approached. This is because, when the input image is correct when the age is close to the boundary of the identification classification, the adjacent identification classification often shows high likelihood, and which classification classification is correct, the reliability is low. Because it is considered.

  It is preferable that the weight is 1.0 as the maximum value in the middle of the identification section and closer to 0 as the boundary of the identification section is approached. For example, weights are set so that when a graph is drawn, a shape similar to a triangular wave or normal distribution is obtained.

  In the present embodiment, in the first age group likelihood distribution 400 and the second age group likelihood distribution 410, the likelihood for each identification category in 10-year increments obtained by combining two attribute segments in 5-year increments is obtained. In the generation likelihood distribution 420, the likelihood for each identification section obtained by combining five, four, and five attribute sections in increments of five years is obtained, whereas in the integrated likelihood distribution 430, the likelihood for overlapping sections is calculated. Can be used to obtain the overlapping section likelihood in increments of 5 years.

  In the first age group likelihood distribution 400, the likelihood of the correct age group discriminator C is high, and the likelihood of the age group discriminator D having the correct age group as the correct answer is also high. However, the likelihood of the age group discriminator F having a completely different age group as a correct answer is higher, and if the estimation result is referred to with reference to the likelihood of the peak of the first age group likelihood distribution 400, the actual age is 39. An old person is estimated as an age group of 60 to 69 years old. On the other hand, in the second age group likelihood distribution 410, the likelihood of the correct age group discriminator J is the highest. Furthermore, also in the generation likelihood distribution 420, the likelihood of the grand generation discriminator O which is a correct generation is the highest.

  In the integrated likelihood distribution 430 obtained by integrating the first age group likelihood distribution 400, the second age group likelihood distribution 410, and the generation likelihood distribution 420, the correct attribute category (age group) of 35 to 39 years old is obtained. The likelihood (overlapping division likelihood) shows the maximum value, and the integrated estimation means 125 estimates the person as “35 to 39 years old” and outputs the result to the output unit 140.

  Note that the integrated estimation unit 125 may output the median value of the attribute classification indicating the likelihood of the maximum value (here, “37 years old”) as the estimated age. Further, the integrated estimation unit 125 may output a weighted average value with the likelihood of the age group as a weight, as the estimation result, using the representative value of each attribute category. The representative value of the attribute classification is an average value of the upper limit and the lower limit of the attribute classification. For example, in the attribute classification “30 to 39 years old”, the representative value is 34.5 years old. Further, in the calculation of the weighted average, only the upper predetermined number of attribute categories having a high likelihood may be used.

  In the present embodiment, in the age group discriminator A to the age group discriminator M included in the first age group discriminator group 132 and the second age group discriminator group 133, each 10-year-old discriminating section is set as a learning region. Includes a set of discriminators having a 5-year-old overlapping section. If the age group identifier is simply 5 years old, the correct learning face image will be reduced accordingly, so learning rather than the facial features of the age group corresponding to that class. There is an unavoidable tendency to construct an age classifier that learns the facial features of a person's own face image. In that case, it is dragged to the output of the age group discriminator who learned about a person with a similar face rather than the age of the person shown in the input face image, and it is estimated that it falls into the wrong attribute classification (age group) The possibility of being made increases. On the other hand, in the present embodiment, while the identification division of each age group discriminator is 10 years old, a plurality of age group discriminators having overlapping divisions are prepared, and the likelihood integrated for each overlapping division is set. By calculating, it is possible to increase the accuracy of age estimation in the attribute classification in increments of 5 years.

  Furthermore, by providing a young generation classifier N, a middle age classifier O, and an old age classifier P that learn the facial features of each generation by increasing the width of the classification category, age group identification of 10 years old It is possible to avoid generational erroneous estimation that cannot be suppressed even by the combination of the first age group discriminator group 132 and the second age group discriminator group 133 including the device.

  Note that the attribute classification is not limited to 5 years. For example, it is possible to estimate the age with a 2-year-old width by the same process, with the attribute classification set in increments of 2 years. Moreover, it is not necessary to fix the identification classification of each age group classifier to 10 years old. Under the condition that the identification classification between the age classifiers includes an overlapping classification, the width of the identification classification may be adjusted according to the number of face images used for learning for each age group. For example, when the lower limit number of correct face images (for example, 5% of the total number of face images) used for learning for each classification category is set and the number of correct face images does not reach the lower limit number The width of the identification section may be widened. Further, the width of the identification section may be narrowed as the number of correct face images increases.

  As described above, the age group can be estimated more finely than in the past while avoiding erroneous estimation by making maximum use of the learning face image prepared in constructing the age estimation apparatus 100.

  As the age group estimation model 131, only the first age group classifier group 132 and the second age group classifier group 133 may be used, and the generation classifier group 134 may not be used. In this case, although the number of age group discriminators decreases, the width of the age step (width of the attribute division step) of the second age group discriminator group 133 is made wider than that of the first age group discriminator group 132. Also good. As a result, as in the case where the generation discriminator group 134 is used in combination, it can be prevented that the age group is completely different. In other words, if there are few correct images prepared for each attribute category, the features of the image itself tend to be learned rather than the features common to each attribute category. This is because each feature can be learned.

  In addition, the integrated estimation unit 125 calculates the average of the first age group likelihood distribution 400 and the second age group likelihood distribution 410 for each attribute category, but the present invention is not limited to this.

  For example, it is preferable to weight the likelihood according to the width of the attribute division assigned to each discriminator and perform the integration process. As shown in FIG. 7, when the first age group discriminator group 132 and the second age group discriminator group 150 are constructed, the integrated estimation unit 125 uses the age group discriminator C and the age group discriminator Q for the attribute classification 151. Synthesize the likelihood output from. Since the age group range of the age group discriminator C is 10 years old and the age group range of the age group discriminator Q is 25 years old, it is preferable to weight according to the ratio of both age groups when integrating. is there. That is, the weighted average may be obtained by multiplying the likelihood output by the age group discriminator C by the weight 1.0 and the likelihood output by the age group discriminator Q by the weight 2.5. Similarly, for the attribute classification 152, the age group range of the age group discriminator E is 10 years old, and the age group range of the age group discriminator R is 20 years old. Is weighted by 1.0, and the likelihood output by the age classifier R is multiplied by a weight of 2.0 to obtain a weighted average. Then, the integrated estimation unit 125 similarly obtains a weighted average of likelihoods in consideration of weighting for other age groups, and normalizes the whole to obtain an integrated likelihood distribution.

[Second Embodiment (Hair Length Estimation Device)]
As a second embodiment, a case where the present invention is applied to a hair length estimation apparatus will be described. That is, the attribute to be estimated is the length of human hair (head hair), and the attribute classification is the length of the hair with a unit of centimeter. Moreover, the 1st classification division classifier group 11 shown in FIG. 1 becomes a 1st hair length identification group, and the 2nd identification classification classifier group 12 becomes a 2nd hair length identification group.

  FIG. 8 shows the relationship between hair length and attribute classification. When the length of the hair is in the range of 0 cm (ie, skin head) to 1 cm, it is defined up to attribute category 0, up to attribute category 6 which is long hair of 40 cm or more. The length of the hair is defined as the length of the dominant part in the glance, although there are some differences depending on the position on the scalp.

  FIG. 9 schematically shows a hair length estimation model 700 corresponding to the age group estimation model. The hair length estimation model 700 includes a first hair length classifier group 710 and a second hair length classifier group 720. Each classifier group is constructed by a learning process in which the upper body image of a person with hair of each length is correct and the upper body image of a person with other hair length is incorrect according to the hair length. .

  The hair length discriminator A is constructed by a learning process in which the upper body image of a person whose hair length is 0 to 1 cm is correct and the other is incorrect. The hair length discriminator B is constructed by a learning process in which the upper body image of a person whose hair length is 1 to 10 cm is correct and the rest is incorrect. Similarly, the hair length classifier C to the hair length classifier H are images of the upper body of a person whose hair length is 10 cm to 30 cm, 30 cm to 50 cm, 0 cm to 5 cm, 5 cm to 20 cm, 20 cm to 40 cm, 40 cm or more, respectively. It is constructed by learning processing with the correct answer and the others as incorrect answers. Then, the hair length classifier A to the hair length classifier H output the likelihood that the hair length of the person in the input image is each classification category.

  Next, components of the hair length estimation apparatus according to the present embodiment and the operation thereof will be described. In the block diagram of the age estimation apparatus according to the first embodiment shown in FIG. 3, the hair length estimation apparatus replaces “age group” with “hair length”, and generates “generation likelihood calculation means” and “generation identifier”. This is achieved by omitting the “group”.

  The feature amount calculation unit 121 of the estimation processing unit 120 extracts an upper body image (up from the chest) of the person from the input image. Further, the feature amount calculation unit 121 may extract a head image above the chin height. FIG. 10 shows an example of the extracted upper body image 600. Further, the feature amount calculation unit 121 extracts a facial feature point 601 from the upper body image 600 by a well-known method, determines a face image region 602 (indicated by a dotted line in FIG. 10) based on the feature point, and determines the face image region. A hair region 603 is defined around 602. It is preferable that the lower end of the hair region 603 is defined so as to extend below the chin and around the shoulder so that the long hair of a woman can be captured.

  Next, the feature amount calculation unit 121 uses the pixels inside the hair region 603 to extract attribute feature amounts used for hair length estimation. The attribute feature quantity is preferably a feature quantity vector having HOG feature quantity 604, Haar-like feature quantity 605, and LBP feature quantity 606 as components. Also, only one of these feature amounts may be used.

  Each of the first hair length likelihood calculating means 122 and the second hair length likelihood calculating means 123 uses the first hair length identifier group 710 and the second hair length identifier group 720 stored in the storage unit. Then, the first hair length likelihood distribution and the second hair length likelihood distribution of the attribute feature amount obtained in the feature amount calculation means 121 are obtained.

  The integrated estimation unit 125 obtains an integrated likelihood distribution by integrating the first hair length likelihood distribution and the second hair length likelihood distribution, and outputs the attribute classification (hair length) indicating the highest likelihood as an estimation result. Output to the unit 140. For example, if the likelihood is highest when the attribute classification (hair length) is 30 to 40 cm, an estimation result of “semi-long” is output.

  As described above, the hair length estimation apparatus according to the second embodiment is a set of hair length classifiers that have learned about identification classifications having overlapping classifications, similar to the age estimation apparatus according to the first embodiment. By preparing the first hair length discriminator group 710 and the second hair length discriminator group 720, the hair length can be reduced more finely than in the past while avoiding false estimation without increasing the number of human images used for learning. Can be estimated.

  In addition, like the generation discriminator in the first embodiment, a hair length discriminator learning a wider discriminating segment is prepared, and the likelihood is integrated by combining three or more types of discriminator groups. You may make it do.

[Modification]
In any of the above embodiments, the entire range of the attribute value range (between the maximum and minimum values of 10 to 80 years in the first example and 0 to 50 cm in the second example) (Although it is assigned to the age group division in the first embodiment and the hair length division in the second embodiment), it is not necessarily limited thereto.

  For example, as shown in FIG. 11, a first discriminating class discriminator group 11 and a second discriminating class discriminator group 12 may be configured. The first identification category classifier group 11 assigns attribute categories 0 to 5 to the identification category identifier AC, and assigns attribute categories 6 to 13 to the identification category identifier DG. On the other hand, the second classification category discriminator group 12 assigns the attribute categories 5 to 6 to the identification category discriminator J, and no discriminator is provided for other attribute categories. In the first embodiment, this is a suitable configuration when it is not necessary to estimate all age groups equally. In other words, when we want to estimate age group in detail focusing on attribute category 5 and attribute category 6, but it is sufficient to estimate other attribute categories 0 to 4 and attribute categories 7 to 13 as young and old, respectively. It is.

  In this case, in the second identification category classifier group 12, there are attribute categories 0 to 4 and 7 to 13 to which no identification category identifier is assigned. In the age group estimation device, no age group discriminator is provided in the range of ages 10 to 35 years and 45 to 80 years. Therefore, the integration estimation unit 125 does not perform the integration process for the range of the attribute classification, and the likelihood output by each classifier of the first classification classification classifier group 11 (first age group classifier group 132). It is assumed that normalization processing is performed on the entire likelihood distribution as it is as an integrated likelihood.

  In addition, the attribute classification (age group) in which both classification classification classifier groups (age group identification group) have the same attribute classification range and one classification classification classifier group (age group identification class) does not exist By making it absent, it is possible to integrate likelihoods having the same properties for the entire range of attribute classification.

  In any of the embodiments, the classifier and its learning method are not particularly limited. For example, a well-known method that is often used in a two-class identification problem such as support vector machine or logistic regression can be employed.

When using a support vector machine, instead of using the likelihood output by each discriminator as it is, the likelihood is input to the sigmoid function shown in Equation (1) and converted to a value in the 0 to 1 interval class. It is desirable to use the posterior probability. Here, A is called a gradient parameter, B is called a bias parameter, and is a parameter that determines the shape of the sigmoid function, and an optimum value is obtained using a verification sample. For example, a known method such as that disclosed in the document “Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods” In Advances in Large Margin Classifiers. MIT Press, 1999 may be appropriately employed.

  Moreover, in the attribute estimation apparatus described so far, a classifier is constructed by using a Haar-like feature amount that is a feature amount extracted from an image, but estimation can also be performed using information other than the image. For example, the attribute estimation device can be similarly constructed by using audio information instead of image information.

  Specifically, when performing age estimation, an age group estimation model 131 composed of a first age group estimator group and a second age group estimator group is prepared according to the first embodiment, and the correct answer Voice data uttered by a speaker whose age is known is prepared as learning data, and an age group estimator is trained using correct learning data for each age group. At this time, a well-known Mel-Frequency Cepstral Coefficients (MFCC) can be used as the feature quantity used for constructing the estimator. That is, in the learning process of the estimator for each age group, the MFCC extracted from the speech data is used as a feature value, and a mixed Gaussian model (GMM: Gaussian Mixture Model) in each age group is generated in the feature value space. Then, the posterior probability that the input voice data belongs to each age group is obtained from the GMM, and the same integration process may be performed using the value of the posterior probability as the attribute likelihood described so far.

  The method of constructing an estimator learned using only correct data as described above and using the attribute likelihood of each age group can be processed in the same manner even if the data to be handled is image information.

  In the above embodiment, the attribute estimation target is human. However, the present invention is not limited to this, and it is estimated to which attribute category the attributes of various objects such as animals, vehicles, and aircraft belong. Can be applied to technology. As attributes, for example, various attributes such as animal hair color, vehicle and aircraft size can be estimated.

  DESCRIPTION OF SYMBOLS 11 1st classification division classifier group, 12 2nd classification division classifier group, 30 1st classification division likelihood distribution, 31 2nd classification division likelihood distribution, 32 Integrated likelihood distribution, 50 Imaging device, 100 Age estimation apparatus 110 image acquisition means, 120 estimation processing section, 121 feature quantity calculation means, 122 first age group likelihood calculation means (first hair length likelihood calculation means), 123 second age group likelihood calculation means (second hair) Long likelihood calculation means), 124 generation likelihood calculation means, 125 integrated estimation means, 130 storage unit, 131 age group estimation model, 132 first age group classifier group, 133 second age group classifier group, 134 generation identification Group, 140 output unit, 150 second age group classifier group, 151,152 attribute classification, 300 face image, 301 feature point, 302 Haar-like feature amount, 303 distance between feature points, 304 Ball feature amount, 305 HOG feature amount, 400 First age group likelihood distribution, 410 Second age group likelihood distribution, 420 generation likelihood distribution, 430 Integrated likelihood distribution, 600 Upper body image, 601 Feature point, 602 Face image Region, 603 hair region, 604 HOG feature, 605 Haar-like feature, 606 LBP feature, 700 hair length estimation model, 710 first hair length classifier group, 720 second hair length classifier group.

Claims (4)

An information acquisition means for acquiring information about an object whose attribute is to be estimated;
Attribute feature amount calculating means for obtaining an attribute feature amount related to an attribute that is divided into a plurality of attribute categories with respect to the object from the information;
Using the plurality of discriminators including a set of discriminators having overlapping sections that are the attribute sections in which the learned classification sections overlap with each other, learning the characteristics of different classification sections in the attribute, the attribute feature amount is Attribute likelihood calculating means for obtaining an attribute likelihood representing the degree of each characteristic of the identification section;
By integrating the attribute likelihoods obtained in the set of classifiers, a duplicate segment likelihood representing the degree that the attribute feature amount has the features of the overlapped segment is obtained, and at least using the overlapped segment likelihood Integrated estimation means for estimating the attribute classification to which the object belongs;
With
The set of classifiers includes a classifier belonging to a first category that has learned the characteristics of each of the classification categories assigned a predetermined range of the attribute in a first category, and the predetermined range as the first category. A classifier belonging to the second section that has learned the characteristics of each of the identification sections assigned in a different number of second sections, and
In the integrated estimation means, the overlapping category is obtained by integrating the attribute likelihood obtained in the classifier belonging to the first category and the attribute likelihood obtained in the classifier belonging to the second category. An attribute estimation apparatus characterized by obtaining a likelihood . The attribute estimation device according to claim 1 ,
The attribute estimation device, wherein the set of classifiers includes classifiers that have learned the characteristics of the classification sections having different widths. The attribute estimation device according to claim 2 ,
The integrated estimation means obtains the overlapping category likelihood by integrating the attribute likelihood obtained for the identification category with a larger weight as the range of the identification category is wider. Estimating device. The attribute estimation device according to any one of claims 1 to 3 ,
The attribute estimation apparatus characterized in that the attribute is one of a person's age or a person's hair length.

Images