JP2013058060A - Person attribute estimation device, person attribute estimation method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a person attribute estimation device that can estimate a person attribute from feature quantity related to wrinkle shapes and has robustness against a change in a photographic environment.SOLUTION: A person attribute estimation device 1 includes: face detection means 110 for detecting a face region from images of a camera 2 photographing a person browsing an advertising medium 2a; a region of interest setting means 111 for specifying the position of a facial organ on the face region by applying a face model created in advance to the face region and setting a first region of interest to be region where wrinkles are easy to be generated as a region of interest for determining feature quantity to be used for estimating the person attribute from the specified position of the facial organ in the face region; feature quantity extraction means 112 for extracting gradient histogram as feature quantity of the face region from edge images in the first region of interest; and person attribute estimation means 120 for referring to a learning device 130 to specify the person attribute classification to which the feature quantity extracted by the feature quantity extraction means 112 belongs.

Description

  The present invention relates to a technique for estimating a person attribute (age / gender) of a person who viewed an advertisement medium when measuring the advertisement effect of the advertisement medium.

  Digital signage that displays advertisements using displays such as liquid crystal displays and projectors is beginning to be installed in various places. By using digital signage, not only will it be possible to provide rich content using video and audio, but it will also be possible to efficiently deliver advertisements according to the location of digital signage. Expansion is expected.

  Since digital signage is used as an advertising medium, it is necessary to measure objective advertising effects in digital signage, such as TV audience ratings and the number of newspapers and magazines sold. When measuring, as disclosed in Patent Document 1, a person who has viewed digital signage is analyzed as an index of the advertising effect of digital signage by analyzing a video captured by a camera that captures a person browsing digital signage Will be installed.

  When counting the number of people who viewed digital signage for each person attribute category, it is necessary to estimate the person attributes (age and gender) of the person included in the video taken by the camera, As an invention for estimating a person attribute of a person, for example, in Patent Document 2, an apparatus for automatically estimating attributes such as sex and age of a target person from a face image of the target person taken from an arbitrary face direction Is disclosed.

  In the device disclosed in Patent Document 2, a feature vector is created from a face image of a captured human image, and after estimating a plurality of attribute similar persons having similar gender and age using the feature vector, the attribute similar person's The gender of the person corresponding to the person image is estimated based on the gender, the attribute similar person is limited by the estimated gender, and the gender of the person is estimated.

  Further, in Patent Document 3, the feature amount such as the shade of the whole or a part of the color of the characteristic part, the number of wrinkles, etc. from the characteristic part that is the facial characteristic such as both ends of the mouth, the corner of the eye, the corner of the eye, and the nose tip. It is described that a person attribute (gender / age) is estimated by extracting and obtaining a score with a reference feature amount prepared in advance.

  Furthermore, Non-Patent Document 1 uses a feature amount effective for gender / age estimation independent of face appearance, such as skin texture related to wrinkles and spots, color information hue information in lip area and cheek area, and function approximation capability. A method has been proposed that enables continuous output of estimated age considering not only gender classification but also application to various fields by using a multi-layered neural network with estimator.

JP 2011-70629 A JP 2003-242486 A JP 2008-282089 A

Hironori Takimoto and three others, “Gender age estimation from face images not affected by posture changes”, Denki Theory C, Vol. 127, No. 7, 2007, p.1022-p.1029

  Certainly, although the technique disclosed in Non-Patent Document 1 is useful, there is still room for improvement. For example, in the skin texture according to Non-Patent Document 1, the average and deviation of the edge strength in each region such as the forehead, the corner of the eye, and the legal age line (bodily line) are defined as wrinkle feature values. Accordingly, the wrinkle shape (the direction and depth of the wrinkle) changes. Therefore, if the feature amount related to the wrinkle shape is used as the wrinkle feature amount, improvement in estimation accuracy of human attributes can be expected. Further, since the hue feature according to Non-Patent Document 1 does not consider the change of the shooting environment, if the robustness can be given to the change of the shooting environment, the estimation accuracy of the person attribute can be expected to be improved.

  Therefore, the present invention has been made in view of such problems, and the object of the present invention is to apply the feature amount related to the wrinkle shape to the estimation of the person attribute, and in addition to the change of the shooting environment. It is to provide a person attribute estimation device, a person attribute estimation method, and a program having robustness.

  A first invention for solving the above-mentioned problem is to apply a face detection means for detecting a face area from an image of a camera set to photograph a person and a face model created in advance to the face area. The position of the facial organ on the face area is specified, and the first region of interest that is an area where wrinkles are likely to occur is obtained as a region of interest for obtaining a feature value used for estimating the human attribute from the position of the specified facial organ. The region of interest setting means for setting the face region, the feature amount extracting means for extracting a gradient histogram from the edge image in the first region of interest as the feature amount, and the feature extracted from a plurality of learning face images A personal attribute estimation unit that refers to a learning device that is a result of classifying a quantity into human attribute categories, and that uses the feature quantity extracted by the feature quantity extraction unit to identify a personal attribute category corresponding to the face region A person attribute estimation device characterized by comprising.

  In the first invention described above, the gradient histogram is obtained by taking the edge direction as the horizontal axis, the edge strength as the vertical axis, and the edge strength as a cumulative histogram. The gradient histogram generated from the edge image in the region of interest indicates the object shape in the region of interest. Here, the first region of the face where wrinkles are likely to occur is the region of interest. The gradient histogram generated by the feature quantity extraction unit of the extraction layer indicates the wrinkle shape in the image in the first region of interest, and the feature quantity related to the wrinkle shape can be applied to estimation of the person attribute.

  Further, a second invention for solving the above-described problem is the person attribute estimation device described in the first invention, wherein the region of interest setting means includes a skin reference color in addition to the first region of interest. A second region of interest corresponding to the lip and a third region of interest corresponding to the lips are set in the face region, and the second region of interest and the third region of interest are added as the feature amount to the gradient histogram. A human attribute estimation device that extracts a color distance between regions.

  According to the second invention described above, since the color distance between the second region of interest and the third region of interest is a relative value, even if the imaging condition varies, the second region of interest and the third region of interest. Since the color distance between the regions of interest does not vary, the human attribute estimation device of this embodiment has robustness against variations in shooting conditions.

  Furthermore, a third invention for solving the above-described problem is the person attribute estimation device described in the first invention or the second invention, wherein a plurality of learning faces prepared for each desired person attribute category It is a human attribute estimation device characterized by comprising learning device generation means for extracting the feature value from an image and creating the learning device by machine learning.

  According to the third aspect described above, the learning device used for estimating the person attribute can be created. As a method for creating the learning device, SVM using nonlinear separation using a kernel function may be used.

  Furthermore, the fourth invention is a face detection step for detecting a face area from an image of a camera set to shoot a person, and a face model created in advance is applied to the face area to thereby detect the face area. As the region of interest for determining the feature amount used for estimating the human attribute from the specified position of the facial organ, the first region of interest that is prone to wrinkles is defined as the facial region. A region-of-interest setting step to be set; a feature amount extracting step of extracting a gradient histogram from an edge image in the first region of interest as the feature amount; and the feature amount extracted from a plurality of learning face images A process including a person attribute estimation step of referring to a learning device that is a result of classification into a plurality of classifications and identifying a classification of a person attribute corresponding to the face region using the feature quantities extracted by the feature quantity extraction unit. A person attribute estimating method characterized in that Yuta performs.

  Further, a fifth invention for solving the above-described problem is the person attribute estimation method described in the fourth invention, wherein, in the region of interest setting step, in addition to the first region of interest, a skin reference color and A second region of interest corresponding to the lip and a third region of interest corresponding to the lips are set as the face region, and in the feature amount extraction step, the feature amount is added to the gradient histogram and the second region of interest. It is a person attribute estimation method characterized by extracting a color distance between a region and the third region of interest.

  Furthermore, the sixth invention is a face detection step for detecting a face area from an image of a camera set to shoot a person, and a face model created in advance is applied to the face area to thereby detect the face area. As the region of interest for determining the feature amount used for estimating the human attribute from the specified position of the facial organ, the first region of interest that is prone to wrinkles is defined as the facial region. A region-of-interest setting step to be set; a feature amount extracting step of extracting a gradient histogram from an edge image in the first region of interest as the feature amount; and the feature amount extracted from a plurality of learning face images A process including a person attribute estimation step of referring to a learning device that is a result of classification into a plurality of classifications and identifying a classification of a person attribute corresponding to the face region using the feature quantities extracted by the feature quantity extraction unit. A computer program to be executed by the Yuta.

  Furthermore, a seventh invention for solving the above-described problem is the person attribute estimation method described in the sixth invention, wherein, in the region of interest setting step, in addition to the first region of interest, a skin reference color and A second region of interest corresponding to the lip and a third region of interest corresponding to the lips are set as the face region, and in the feature amount extraction step, the feature amount is added to the gradient histogram and the second region of interest. A computer program characterized by extracting a color distance between a region and the third region of interest.

  As described above, according to the present invention, it is possible to estimate a person attribute of a feature amount related to a wrinkle shape, and in addition, a person attribute estimation device and a person attribute estimation method having robustness against changes in a shooting environment And programs can be provided.

The figure explaining the system configuration which has arranged the person attribute estimating device. The figure explaining the hardware which a person attribute estimation apparatus has. The figure explaining the structure of the layer of the application mounted in a person attribute estimation apparatus. The figure explaining operation | movement of the face detection means of a feature-value extraction layer. FIG. 6 is a supplementary diagram for explaining the operation of the feature amount extraction layer face detection means. The figure explaining an example of an object field. The figure explaining the operation | movement which a feature-value extraction means extracts a gradient histogram. The figure explaining the gradient histogram which a feature-value extraction means extracts. The figure explaining the operation | movement in which the feature-value extraction means extracts the color distance between the 2nd region of interest and the 3rd region of interest. The figure explaining the effect of this invention.

  Now, preferred embodiments of the present invention will be described in detail with reference to the drawings. The embodiment described below is only one embodiment of the present invention, and the present invention is not limited to the embodiment described below, and various modifications and changes are possible.

  FIG. 1 is a diagram illustrating a system configuration in which a person attribute estimation device 1 according to the present embodiment is arranged. The person attribute estimation device 1 illustrated in FIG. 1 is an advertisement medium 2a (digital signage in FIG. 1) that displays an advertisement, a camera 2 that photographs a person 3 who is browsing the advertisement medium 2a, and a camera 2 that captures images. And a person attribute estimation device 1 that estimates the person attributes (age and gender) of the person 3 included in the video.

  In the present embodiment, the advertising medium 2a is a medium that presents an advertisement to the public in a public place. In FIG. 1, the advertising medium 2a is used as a digital signage that displays an advertisement using a liquid crystal display. Although illustrated, the advertisement medium 2a may be a show window, an advertisement signboard, a specific product display shelf in a store, or the like.

  FIG. 2 is a diagram illustrating hardware included in the person attribute estimation device 1. As shown in FIG. 2, the person attribute estimation device 1 which is a device realized using a computer has a CPU (hardware) connected to a chipset 1b mounted with a memory interface, a graphic board, and the like. The chip set 1b is connected with a memory 1c, a display 1d, a keyboard / mouse 1e, a LAN interface 1f, a hard disk 1g, and a camera interface 1h.

  A computer program for operating the person attribute estimation apparatus 1 is mounted on the hard disk 1g of the person attribute estimation apparatus 1, and this computer program operates on an operating system that provides basic functions and on the operating system. In addition to a library related to image recognition (for example, OpenCV), an application that provides a function for estimating a person attribute using a library related to image recognition is implemented.

  FIG. 3 is a diagram for explaining the configuration of the application layer implemented in the person attribute estimation device 1. As shown in FIG. 3, an application layer providing a function for estimating a person attribute using a library related to image recognition includes an interface layer 10 to which an image taken by the camera 2 is input. By analyzing the image input to the interface layer 10, the feature amount extraction layer 11 that extracts the feature amount used for estimating the person attribute from the image and the feature amount obtained from the learning face image are used in advance. With reference to the generated person attribute learning layer 13 having the learning device 130 and the learning device 130 having the person attribute learning layer 13, the person attribute estimation layer 12 that estimates the person attribute from the feature amount extracted by the feature amount extraction layer 11. Is included.

  From here, the structure of the layer of the application mounted in the person attribute estimation apparatus 1 is demonstrated in detail. As shown in FIG. 3, the interface layer 10 includes an image input unit 100 realized using the camera interface 1 h. The image output to the feature amount extraction layer 11 may be each frame included in the moving image captured by the camera 2, or may be a frame at regular time intervals (for example, every minute).

  Next, the feature quantity extraction layer 11 will be described. As illustrated in FIG. 3, the feature amount extraction layer 11 includes a face detection unit 110 that detects a rectangular face region from an image input from the camera 2, and a feature for each face region detected by the face detection unit 110. A region-of-interest setting unit 111 that detects a target region from which the amount is extracted, and a unit that extracts a feature amount of the face region detected by the face detection unit 110. A feature quantity extraction unit 112 that extracts the feature quantity as a feature quantity of the face area is included.

  From here, each means contained in the feature-value extraction layer 11 is demonstrated. FIG. 4 is a diagram for explaining the operation of the face detection unit 110 in the feature quantity extraction layer 11, and FIG. 5 is a supplementary diagram for explaining the operation of the face detection unit 110 in the feature quantity extraction layer 11.

  In the present embodiment, when the face detection unit 110 of the feature quantity extraction layer 11 detects a face area from an image input from the interface layer 10, first, the background from the image input to the image input unit 100 of the interface layer 10 is detected. The process which removes is performed (S1). As a method for removing the background, a background image is prepared, and if the difference between the pixel at a certain position in the image input from the interface layer 10 and the pixel at the same position in the background image is less than a threshold value, “0” ( Black), if it is equal to or greater than the threshold value, the process of “1” (white) may be performed.

  When the background is removed from the image input from the interface layer 10, the face detection unit 110 of the feature quantity extraction layer 11 removes the background from the image input from the interface layer 10 as illustrated in FIG. The subsequent area is extracted as a blob (small blob) 4a, and further, an area surrounded by a rectangle depending on the size of the blob 4a is extracted as a blob area 4b (S2).

When the face detection unit 110 of the feature quantity extraction layer 11 extracts the blob (blob) 4a and the blob area 4b, the face detection process 110 applies the blob area 4b to the image before removing the background, and performs face detection processing from the image in the blob area 4b. To detect face area candidates (S3). For the face detection process, a method using Haar-Like feature values disclosed in Non-Patent Document 1 below may be used.
<Non-Patent Document 1> Paul Viola and Michel J. Jones: "Rapid Object Detection Using a Boosted Cascade of Simple Features", Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.511-518, (2001 )

  As shown in FIG. 5B, when the blob area 4b is applied to the image before the background is removed, and face detection processing is performed from the image in the blob area 4b, a plurality of face areas detected from the same face 4c and the face area 4d that is erroneously detected as a face other than the face is included, in this embodiment, the face detection unit 110 of the feature amount extraction layer 11 detects the face in the extracted image in the blob area 4b. After performing the processing, as shown in FIG. 5C, in the two face areas, the distance between the center of one face area and the center of the other face area ≦ the distance from the center of one face area to the end. Check whether the distance + the distance from the center to the end of the other face area, etc., and determine whether the candidate face area 4c overlaps, and select the face area of the maximum size from the face areas determined to overlap. Detect from the same face, such as by selecting A plurality of face areas 4b that are estimated to be included in one face area 4e (S4), or as shown in FIG. 5D, the face areas (here, face areas 4e and d) are included. Referring to the ratio of blob 4a, a process of deleting a candidate for face area 4d that is erroneously detected other than the face is performed (S5), and the finally remaining face area (here, face area 4e) is used as a face detection result, The finally remaining face area (here, the face area 4e) is normalized to a predetermined size (S6), and this process is terminated.

Next, the region of interest setting unit 111 of the feature quantity extraction layer 11 will be described. In the present embodiment, the region-of-interest setting unit 111 of the feature quantity extraction layer 11 applies a face model created in advance to the face region detected by the face detection unit 110, so that the facial organs (eyes, nose, The area of the face where wrinkles are likely to occur as the region of interest for determining the feature quantity used for estimating the human attribute using the position of the identified facial organs (eyes, nose, lips, etc.) A process of setting a certain first region of interest, a second region of interest that is a facial region as a skin reference color, and a third region of interest that is a lip region is performed as a facial region. As a technique for applying the face model to the face area, ASM (Active Shape Model) disclosed in the following Non-Patent Document 2 or the like and applying the face model to the face area by Affine transformation may be used.
<Non-Patent Document 2> TF Cootes, D. Cooper, CJ Taylor and J. Graham, "Active Shape Models-Their Training and Application." Computer Vision and Image Understanding. Vol. 61, No. 1, Jan. 1995, pp 38-59.

  FIG. 6 is a diagram illustrating an example of setting the region of interest. In FIG. 6, as a first region of interest which is a face region where wrinkles are likely to occur, a total of five regions including a forehead region 5a, left and right current regions 5b and c, and left and right legal age line regions 5d and 5e, The nose region 5f is set as the second region of interest that is the face region to be colored, and the lip region 5g is set as the third region of interest that is the lip region, and the region of interest setting means 111 of the feature quantity extraction layer 11 is set. Finds these regions of interest in the face region detected by the face detection means 110 from the relative positional relationship between the region of interest shown in FIG. 6 and the facial organs (eyes, nose, lips, etc.).

  As described above, in this embodiment, the face model created in advance is applied to the face area, the position of the face organ in the face area is detected, and the region of interest is set as the face area using the position of the face organ. The region of interest in the face region can be accurately detected not only for the front face but also for the oblique face, and the person attribute estimation device 1 of the present embodiment is robust against the orientation of the photographed face. Will have sex.

  Next, the feature quantity extraction unit 112 of the feature quantity extraction layer 11 will be described. In Non-Patent Document 1, the average and deviation of the edge strength in each region such as the forehead, the corner of the eye, and the legal age line (Hoyu Line) are defined as wrinkle feature amounts. Focusing on the fact that the wrinkle shape (the direction and depth of the wrinkles) changes accordingly, the feature amount extraction unit 112 of the feature amount extraction layer 11 detects the edge image in the first region of interest detected by the region of interest setting unit 111. Is extracted as a feature amount indicating the wrinkle shape (the direction and depth of the wrinkle) in the first region of interest detected by the region-of-interest setting unit 111.

  Furthermore, paying attention to the difference in redness of lips depending on age and sex, in this embodiment, the feature amount extraction unit 112 of the feature amount extraction layer 11 detects the region of interest setting unit 111 as a feature amount indicating redness of the lips. The color distance between the second region of interest and the third region of interest is extracted. Note that since the color distance between the second region of interest and the third region of interest is a relative value, the color distance between the second region of interest and the third region of interest is the same even if the shooting conditions vary. Since there is no change, the person attribute estimation device 1 of the present embodiment has robustness against changes in shooting conditions.

  FIG. 7 is a diagram for explaining an operation in which the feature quantity extraction unit 112 of the feature quantity extraction layer 11 extracts the gradient histogram. This procedure is executed for each face area detected by the face detection unit 110.

When the gradient histogram is extracted, the wrinkle feature amount extraction unit 112 of the feature amount extraction layer 11 first detects edges in the X direction and the Y direction in the image in the first region of interest that is the target region of the face where wrinkles are likely to occur. An image is calculated (S10). As a method for calculating the edge image in the X direction and the Y direction in the first region of interest, a Sobel filter that is one of spatial filters may be used. Here, in order to calculate the edge image in the X direction and the Y direction. The operators shown in Table 1 may be used as Sobel filter operators.

When the edge image in the X direction and the Y direction in the image in the first region of interest, which is the target region of the face where wrinkles are likely to occur, is calculated, the feature amount extraction unit 112 of the feature amount extraction layer 11 uses Equation 1. The gradient strength / gradient direction is calculated for each pixel of the edge images in the X and Y directions (S11).

  When the gradient strength / gradient direction is calculated for each pixel of the edge images in the X direction and the Y direction, the feature amount extraction unit 112 of the feature amount extraction layer 11 uses the gradient strength / gradient direction as the gradient direction, Y A gradient histogram with the axis as the gradient strength is generated (S12), and this procedure is terminated.

  FIG. 8 is a diagram for explaining a gradient histogram extracted by the feature amount extraction unit 112 of the feature amount extraction layer 11. The gradient direction of the gradient histogram may be all directions (0 degrees to 360 degrees). However, in this embodiment, the feature amount extraction unit 112 of the feature amount extraction layer 11 divides all directions (0 degrees to 360 degrees) into 10 parts. A gradient histogram in 36 directions is generated.

  The gradient histogram generated from the X-direction and Y-direction edge images of the image in the target region indicates the object shape in the target region. Here, the first region of interest where wrinkles are likely to occur is the target region. Therefore, the gradient histogram generated by the feature amount extraction unit 112 of the feature amount extraction layer 11 indicates a wrinkle shape in the image in the first region of interest.

  Next, an operation in which the feature amount extraction unit 112 of the feature amount extraction layer 11 extracts a color distance between the second region of interest and the third region of interest will be described. FIG. 9 is a diagram for explaining the operation of extracting the color distance between the second region of interest and the third region of interest in the feature amount extraction unit 112 of the feature amount extraction layer 11. This procedure is detected by the face detection unit 110. It is executed for each face area.

When extracting the lip feature value, the feature value extraction unit 112 of the feature value extraction layer 11 extracts a color in a color space containing a lot of redness components from the image in the reference skin region and the image in the third region of interest. After obtaining the histogram (S20), the difference between the color histograms is extracted as a color distance (S21), and this procedure is terminated. In the present embodiment, a color space containing a lot of redness components is defined as three color spaces of H (hue), S (saturation), and R (red), and the batchary distance according to Equation 2 is set as a difference between the color histograms. Although calculated, the Mahalanobis distance may be calculated as a difference between histograms.

  As described above, the feature amount extraction unit 112 of the feature amount extraction layer 11 creates a gradient histogram in 36 directions for each of the five first regions of interest as the feature amount of the face image, and H (hue) , S (saturation) and R (red) for each of the three color spaces, one facial region is used to extract the color distance between the second region of interest and the third region of interest (here, the batcha rear distance). The feature quantity extracted from is a 183 dimensional vector.

  Next, the learning device 130 of the person attribute learning layer 13 will be described. The person attribute learning layer 13 includes a learning device 130 used for estimating a person attribute (here, age / sex) from the feature amount (here, gradient histogram and color distance) obtained from the feature amount extraction layer 11. A learning device generating means 131 is provided for creating the learning device 130 by machine learning.

  The learning device 130 of the person attribute learning layer 13 is generated in advance before the learning device generating means 131 estimates the person attribute. In this embodiment, the learning device generation means 131 of the person attribute learning layer 13 generates the learning device 130 by SVM (Support Vector Machine).

  When the learning device 130 is generated using SVM (Support Vector Machine), classification of person attributes (here, from 0 to 19 years old, from 20 to 34 years old, for each of four age categories according to gender sex) Prepare a plurality of learning face images for each of the 35 to 49 years old and 50 years old or more), extract the above-mentioned 183 dimensional feature amount for each learning face image, and use the kernel function The result obtained by separating the feature amount extracted from each learning face image into eight human attribute sections by the SVM using the non-linear separation is used as the learning device 130.

  Next, the person attribute estimation layer 12 will be described. The person attribute estimation means of the person attribute estimation layer 12 refers to the learning device 130 of the person attribute learning layer 13, and the person is obtained from the feature amounts (here, the wrinkle feature amount and the lip feature amount) obtained from the feature amount extraction layer 11. A process of estimating attributes (here, age and sex) is performed. Specifically, the person attribute estimation means of the person attribute estimation layer 12 refers to the learning device 130 of the person family name learning layer when estimating the person attribute (here, age and sex), and the feature amount extraction layer 11 The person attribute classification corresponding to the vector space to which the 183 dimensional feature value obtained from the above belongs is specified as the person attribute of the face area targeted by the feature value.

  Finally, the effect of the present invention will be described. FIG. 10 is a diagram for explaining the effects of the present invention. FIG. 10 (a) shows the age estimation result, FIG. 10 (b) shows the gender estimation result, and the percentage display in FIG. ) No. 2 indicates that the identification rates for correctly estimating women from the age of 20 to 34 are 87%, 82%... At each angle.

  In the age estimation result of FIG. 10, the age estimation result of the younger group (under 19 years old) has a very poor identification rate. This is because the younger group used in generating the person attribute learner 130 (19 It is considered that the number of learning face images of 15 years old or younger is as small as 15 people. If the learning device 130 is generated using more learning face images, it is considered that the identification rate is improved.

DESCRIPTION OF SYMBOLS 1 Person attribute estimation apparatus 10 Interface layer 100 Image input means 11 Feature-value extraction layer 110 Face detection means 111 Area of interest setting means 112 Feature-value extraction means 12 Person attribute estimation layer 120 Person attribute estimation means 13 Person attribute learning layer 130 Learner 131 Learning device generation means 2 Camera 2a Advertising medium 3 Person

Claims (7)

  Face detection means for detecting a face area from an image of a camera set to photograph a person, and a face model created in advance is applied to the face area to identify the position of a facial organ on the face area A region-of-interest setting means for setting, as the face region, a first region of interest, which is a region where wrinkles are likely to occur, as a region of interest for obtaining a feature amount to be used for estimating human attributes from the position of the identified facial organ; As the feature amount, a feature amount extraction unit that extracts a gradient histogram from an edge image in the first region of interest and a result of classifying the feature amount extracted from a plurality of learning face images into individual attribute categories Human attribute estimation characterized by comprising personal attribute estimation means for referring to a learning device and identifying a classification of a person attribute corresponding to the face area using the feature quantity extracted by the feature quantity extraction means Location.   The region-of-interest setting means sets, in addition to the first region of interest, a second region of interest that is a region used as a skin reference color and a third region of interest corresponding to the lips as the face region, The human attribute estimation apparatus according to claim 1, wherein a color distance between the second region of interest and the third region of interest is extracted as a quantity in addition to the gradient histogram.   2. The learning apparatus according to claim 1, further comprising: a learning device generating unit that extracts the feature amount from a plurality of learning face images prepared for each desired person attribute category and creates the learning device by machine learning. Or the person attribute estimation apparatus of Claim 2.   A face detection step for detecting a face area from an image of a camera set to shoot a person, and a position of a facial organ on the face area is specified by applying a face model created in advance to the face area. A region-of-interest setting step of setting, as the face region, a first region of interest that is an area where wrinkles are likely to occur, as a region of interest for obtaining a feature amount used for estimating a human attribute from the position of the identified facial organ; As the feature amount, a feature amount extraction step of extracting a gradient histogram from an edge image in the first region of interest, and a result of classifying the feature amount extracted from a plurality of learning face images into categories of person attributes With reference to the learning device, the computer executes a process including a person attribute estimation step of specifying a person attribute classification corresponding to the face region using the feature quantity extracted by the feature quantity extraction unit. People attribute estimating method comprising.   In the region of interest setting step, in addition to the first region of interest, a second region of interest that is a region used as a skin reference color and a third region of interest corresponding to the lips are set as the face region, and the feature 6. The person attribute according to claim 5, wherein, in the quantity extraction step, a color distance between the second region of interest and the third region of interest is extracted as the feature amount in addition to the gradient histogram. Estimation method.   A face detection step for detecting a face area from an image of a camera set to shoot a person, and a position of a facial organ on the face area is specified by applying a face model created in advance to the face area. A region-of-interest setting step of setting, as the face region, a first region of interest that is an area where wrinkles are likely to occur, as a region of interest for obtaining a feature amount used for estimating a human attribute from the position of the identified facial organ; As the feature amount, a feature amount extraction step of extracting a gradient histogram from an edge image in the first region of interest, and a result of classifying the feature amount extracted from a plurality of learning face images into categories of person attributes Referring to the learning device, the computer is caused to execute a process including a person attribute estimation step of specifying a person attribute classification corresponding to the face region using the feature quantity extracted by the feature quantity extraction unit. Because of the computer program. In the region of interest setting step, in addition to the first region of interest, a second region of interest that is a region used as a skin reference color and a third region of interest corresponding to the lips are set as the face region, and the feature The quantity extraction step includes extracting a color distance between the second region of interest and the third region of interest as the feature amount in addition to the gradient histogram. Computer programs listed in

Images