JP6969878B2 - Discriminator learning device and discriminator learning method - Google Patents

Description

The present invention relates to a classifier learning device and a classifier learning method for learning a classifier that estimates a person's attributes based on a person image including at least a part of the person.

In recent years, in commercial facilities such as shopping malls and theme parks, for the purpose of displaying digital signage advertisements according to the customer base and investigating the customer base, the person attributes of the person have been changed based on the captured image of the person taken by a surveillance camera. Estimating is being done. As such a technique, a technique for estimating a person's attributes (for example, age, gender, etc.) based on a face image extracted from a captured image using a discriminative model (discriminative model) is known (Patent Document 1, Patent Document 1). 2. See Patent Document 3).

Japanese Unexamined Patent Publication No. 4-264985 Japanese Unexamined Patent Publication No. 2015-125731 Special Table 2007-5210

However, in the above-mentioned prior art of Patent Document 1-3, since the person attribute is estimated based on the face image, there is a problem that the person attribute cannot be estimated when the face image cannot be extracted from the captured image. was there. Therefore, it is conceivable to estimate the person attribute based on a person image (for example, a whole body image) including at least a part of the person instead of the face image, but the clothes and appearance (appearance) of the person included in the person image are determined. Since it varies greatly depending on the country, region, season, age, etc., it was not possible to accurately estimate the person attributes based on the person image.

The present disclosure has been devised in view of such problems of the prior art, and provides a classifier capable of accurately estimating the person attribute of a person based on a person image including at least a part of the person. The main purpose is to provide a discriminator learning device and a discriminator learning method.

The classifier learning device of the present disclosure includes a processor and a communication unit that executes communication with the image pickup device, and the processor is at least other than the face of an estimation target person to be estimated from the image pickup device for gender or age. A whole-body image including a part and a face image including the face of the estimated target person are acquired, and the input whole-body image constructed in advance by machine learning is analyzed and the person included in the input whole-body image is included. for human image discriminator for outputting the estimation result of sex or age, the sex or age of the estimated target person estimated by entering the systemic image estimation target person was previously constructed by machine learning, it is input the person information for identifier to infer gender or age of a person included a face image the input face image by analyzing, sex or age of the estimated target person by inputting the face image of the estimated target person Is estimated, and it is determined whether or not the estimation result based on the whole body image and the estimation result based on the face image match, and it is determined that the estimation result based on the whole body image and the estimation result based on the face image match. If so, the estimation result determined to match is used as a correct answer label, and the whole body image is additionally trained by the person image classifier.

According to the present disclosure, it is possible to provide a classifier capable of accurately estimating the person attribute of a person based on a person image including at least a part of the person.

A block diagram showing an outline of a classifier learning system including a classifier learning device according to the present disclosure. Block diagram showing the schematic configuration of the classifier learning device Flow diagram showing the flow of classifier learning process The figure which shows an example of the whole body image and the face image extracted from the captured image. Figure showing an example of estimation result of person attribute Flow diagram showing the flow of person attribute estimation processing

The first invention made to solve the above-mentioned problems includes a processor and a communication unit that executes communication with an image pickup device, and the processor is an estimation target person who is an estimation target of a person attribute from the image pickup device. A person image including at least a part of the above is acquired, a person attribute of the estimated target person is estimated based on the person image using a person image classifier constructed in advance by machine learning, and a person constructed in advance by machine learning. Using the information classifier, the person attribute of the estimation target person is estimated based on the person information which is different from the person image, and the estimation result based on the person image and the estimation result based on the person information match. If it is determined whether or not the estimation result matches the estimation result based on the person image and the estimation result based on the person information, the person image is added as training data and the person image classifier is used. It is characterized by learning.

According to the classifier learning device according to the first invention, when the estimation result based on the person image and the estimation result based on the person information other than the person image match, the person image of the person to be estimated is added as learning data. Since the learning of the classifier for the person image is performed, it is possible to realize highly accurate online learning (additional learning). This makes it possible to provide a classifier capable of accurately estimating the person attributes of the person based on the person image of the person.

Further, in the second invention, when it is determined in the first invention that the estimation result based on the person image and the estimation result based on the person information do not match, the person image is converted into the person image. It is characterized in that the discriminator for a person image is trained by adding it as learning data having a lower reliability than when the estimation result based on the estimation result and the estimation result based on the person information match.

According to the classifier learning device according to the second invention, when the estimation result based on the person image and the estimation result based on the person information other than the person image do not match, the person image of the person to be estimated has the reliability. Since the classifier for a person image is learned by adding it as low learning data, it is possible to realize more accurate online learning.

Further, the third invention is characterized in that, in the second invention, the estimation result based on the person information is used as a correct answer label.

According to the classifier learning device according to the third invention, when the person information is more reliable than the person image, the person information is used as a correct label to realize more accurate online learning. Is possible.

Further, the fourth invention is characterized in that, in any one of the first invention to the third invention, the person information is a face image included in the person image.

According to the classifier learning device according to the fourth invention, since the face image included in the person image is used as the person information, the person information can be easily acquired.

Further, in the fifth aspect of the invention, in the fourth aspect, the processor further estimates the person attribute of the estimation target person based on the person image, and the face image is extracted from the person image. Estimates the person attribute of the person based on the face image using the person information classifier, and if the face image is not extracted from the person image, the person image classifier is used. It is characterized in that the person attribute of the estimation target person is estimated based on the person image.

According to the classifier learning device according to the fifth invention, even if the face image cannot be extracted from the person image, the person attribute of the person can be estimated based on the person image. Further, since the person image classifier is constructed by highly accurate online learning, it is possible to estimate the person attribute based on the person image with high accuracy.

Further, the sixth invention is characterized in that, in any one of the first invention to the fifth invention, the person attribute includes the gender and age of the person.

According to the classifier learning device according to the sixth invention, gender or age can be estimated as a person attribute.

Further, the seventh invention is characterized in that, in any one of the first invention to the sixth invention, the person image is a whole body image of the person to be estimated.

According to the discriminator learning device according to the seventh invention, a whole body image of a person can be used as a person image. Since the whole body image is information that can be easily acquired even when the person information such as the face image cannot be obtained, for example, when the person turns his back, it is possible to secure many opportunities to estimate the person attribute.

Further, the eighth invention uses a step of acquiring a person image including at least a part of the person to be estimated to be an estimation target of a person attribute from an image pickup apparatus, and a person image classifier constructed in advance by machine learning. The estimation target person is estimated based on the person information which is different from the person image by using the step of estimating the person attribute of the estimation target person based on the person image and the person information classifier constructed in advance by machine learning. The person attribute of the person is estimated, the estimation result based on the person image and the estimation result based on the person information are determined to match, the estimation result based on the person image and the estimation result based on the person information. When it is determined that the above is the same, the person image is added as training data to include a step of learning the person image classifier.

According to the discriminator learning method according to the eighth invention, when the estimation result based on the person image and the estimation result based on the person information other than the person image match, the person image of the person to be estimated is added as learning data. Since the learning of the classifier for the person image is performed, it is possible to realize highly accurate online learning (additional learning). This makes it possible to provide a classifier capable of accurately estimating the person attributes of the person based on the person image of the person.

Further, in the ninth invention, when it is determined in the eighth invention that the estimation result based on the person image and the estimation result based on the person information do not match, the person image is used as the person image. It is characterized by further having a step of learning the person image classifier by adding it as training data having lower reliability than the case where the estimation result based on the person information and the estimation result based on the person information match.

According to the discriminator learning method according to the ninth invention, when the estimation result based on the person image and the estimation result based on the person information other than the person image do not match, the person image of the person to be estimated has the reliability. Since the classifier for a person image is learned by adding it as low learning data, it is possible to realize more accurate online learning.

Further, the tenth invention is characterized in that, in the ninth invention, the estimation result based on the person information is used as a correct answer label.

According to the classifier learning method according to the tenth invention, when the person information is more reliable than the person image, the person information is used as a correct label to realize more accurate online learning. Is possible.

Further, the eleventh invention is characterized in that, in any one of the eighth invention or the tenth invention, the person information is a face image included in the person image.

According to the discriminator learning method according to the eleventh invention, since the face image included in the person image is used as the person information, the person information can be easily acquired.

Further, in the twelfth invention, in the eleventh invention, the person attribute of the estimation target person is further estimated based on the person image, and when the face image is extracted from the person image, the person. Using the information classifier, the step of estimating the person attribute of the person based on the face image, and when the face image is not extracted from the person image, the person image classifier is used. It is characterized by further including a step of estimating the person attribute of the person to be estimated based on the person image.

According to the discriminator learning method according to the twelfth invention, even if the face image cannot be extracted from the person image, the person attribute of the person can be estimated based on the person image. Further, since the person image classifier is constructed by highly accurate online learning, it is possible to estimate the person attribute based on the person image with high accuracy.

Further, the thirteenth invention is characterized in that, in any one of the eighth invention to the twelfth invention, the person attribute includes the gender and age of the person.

According to the discriminator learning method according to the thirteenth invention, gender or age can be estimated as a person attribute.

Further, the fourteenth invention is characterized in that, in any one of the eighth invention to the thirteenth invention, the person image is a whole body image of the estimation target person.

According to the discriminator learning method according to the fourteenth invention, a whole body image of a person can be used as a person image. Since the whole body image is information that can be easily acquired even when the person information such as the face image cannot be obtained, for example, when the person turns his back, it is possible to secure many opportunities to estimate the person attribute.

Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the present embodiment, for example, in a commercial facility such as a shopping mall or a theme park, a case where a person attribute of a person included in the captured image is estimated based on an captured image obtained by capturing a person with an image pickup device will be described. In this embodiment, gender is estimated as a person attribute. Further, in the present embodiment, a whole body image is used as the person image, and a face image is used as the person information.

FIG. 1 is a configuration diagram showing an outline of a classifier learning system 1 including a classifier learning device 3 according to the present disclosure. As shown in FIG. 1, the classifier learning system 1 mainly includes an image pickup device 2 and a classifier learning device 3. The image pickup device 2 and the classifier learning device 3 are connected to each other via a network 4 such as a dedicated line or a LAN (Local Area Network).

The image pickup device 2 is a surveillance camera (video camera) having a known shooting function and communication function, and is installed on a wall, a ceiling, a pole, or the like in various places of a commercial facility, and a predetermined surveillance area is a still image or a moving image. Image in time series. The captured image (still image or moving image) generated by the imaging of the imaging device 2 is appropriately transmitted to the classifier learning device 3 via the network 4. The form, function, arrangement, quantity, and the like of the image pickup apparatus 2 are not particularly limited as long as the predetermined monitoring area can be imaged, and various changes can be made.

The classifier learning device 3 is composed of a computer having a known configuration, and a whole body image extracted from the captured image acquired from the image pickup device 2 is added as learning data to estimate a person attribute (gender) based on the whole body image. Learn the vessel (discrimination model). This learning is online learning (additional learning) in which new learning data is added to the classifier possessed by the classifier learning device 3 in advance to perform learning. Further, the classifier learning device 3 can also estimate the person attribute (gender) based on the whole body image extracted from the captured image acquired from the image pickup device 2 by using the above-mentioned classifier.

FIG. 2 is a block diagram showing a schematic configuration of the classifier learning device 3. The classifier learning device 3 has a known hardware configuration, and collectively executes each process of classifier learning and person attribute estimation based on a predetermined control program (for example, a classifier learning program and a person attribute estimation program). CPU 12, RAM 12, which is a volatile memory that functions as a work area of the processor 11, ROM 13, which is a non-volatile memory for storing control programs and data executed by the processor 11, and a storage unit 14 including an HDD and a flash memory. It has a display unit 15 including a liquid crystal monitor and the like, an input unit 16 including an input device such as a keyboard and a mouse, and a communication unit 17 for executing communication via a network 4.

Each function of the classifier learning device 3 can be realized by the processor 11 executing a predetermined control program in the hardware configuration shown in FIG. The classifier learning device 3 is not limited to a computer, and other information processing devices (for example, a server or the like) capable of performing the same function can also be used. Further, at least a part of the function of the classifier learning device 3 may be replaced by processing by other known hardware.

FIG. 3 is a flow chart showing the flow of the classifier learning process in the classifier learning device 3. The following processing is performed by the processor 11 of the classifier learning device 3, and is started when a captured image of a person is acquired from the image pickup device 2 which is an external device.

First, a whole-body image of a person is extracted from the captured image (step ST101). Specifically, a region including the whole body of a person (whole body region) is detected from the captured image, and the detected whole body region is cut out from the captured image to extract the whole body image from the captured image. Detection of the whole body region is performed by using a known method such as a method using a statistical or non-statistical classifier or a method using template matching.

Subsequently, the face image of the person is extracted from the whole body image (step ST102). Specifically, a region including a person's face (face region) is detected from the whole body image, and the detected face region is cut out from the whole body image to extract a face image from the whole body image. The face region is detected by using a known method such as a method using a statistical or non-statistical classifier or a method using template matching. The face image may be directly extracted from the captured image instead of the whole body image.

Since the face image is less affected by the country / region, season, age, etc. than the full-body image, the reliability in estimating the human attribute is higher than that of the full-body image.

FIG. 4 is a diagram showing an example of a whole body image and a face image extracted from a captured image. In the example of FIG. 4A, the person included in the captured image 21 is an adult female, and the captured image 21 is extracted from the whole body image 22 of the adult female, and the whole body image 22 is used to capture the face image 23 of the adult female. Has been extracted. In the example of FIG. 4B, the person included in the captured image 31 is a male child, the captured image 31 is extracted from the full-body image 32 of the male child, and the full-body image 32 is the facial image 33 of the male child. Has been extracted.

In the next step ST103, the person attribute (gender) is based on the whole body image extracted from the captured image by using the whole body image classifier (person image classifier) constructed by machine learning using the whole body image of the person as learning data. ) Is estimated. Specifically, for example, a method using a Gabor feature amount, a method using an LBP (Local Binary Pattern) feature amount, a method using a color feature amount, and other known methods are used to capture the contour (or one of the contours) of a person from a whole body image. Part) or slightly color is detected, and the person attribute (gender) is estimated based on the detected contour and color. For example, if it is determined from the contour of a person that the person is wearing a skirt, the person can be presumed to be a female. Similarly, it is possible to estimate the person attribute (gender) based on the color, appearance (appearance), hairstyle, etc. of the clothes.

The whole-body image classifier (human image classifier) is a classifier constructed by extracting image features from various whole-body images used as learning data and analyzing them statistically or non-statistically. As the discriminator for the whole body image, for example, an SVM (Support Vector Machine), a neural network, or the like can be used. The discriminator for the whole body image is not particularly limited as long as the person attribute can be estimated based on the whole body image, and various types of discriminators can be used.

In the following step ST104, the person attribute (gender) is estimated based on the face image by using the face image classifier (person information classifier) that uses the face image of the person as learning data. Specifically, facial feature points such as eyes, nose, and mouth are detected from facial images using known methods such as a method using a Gabor feature, a method using an LBP feature, and a method using a color feature. , The person attribute (gender) is estimated based on the detected facial feature points.

The face image classifier is a classifier constructed by extracting image features from various face images used as learning data and analyzing them statistically or non-statistically. As the face image classifier, for example, an SVM, a neural network, or the like can be used. The face image classifier is not particularly limited as long as the person attribute can be estimated based on the face image, and various types of classifiers can be used.

FIG. 5 is a diagram showing an example of the estimation result of the person attribute in step ST103 and step ST104. 5 (a) shows the estimation result based on the whole body image 22 of FIG. 4 (a) and the estimation result based on the face image 23, and FIG. 5 (b) shows the estimation result based on the whole body image 32 of FIG. 4 (b). The result and the estimation result based on the face image 33 are shown. In the example of FIG. 5A, the detection result based on the whole body image 22 was “female”, and the detection result based on the face image 23 was also “female”. In the example of FIG. 5B, the detection result based on the whole body image 32 was "female", but the detection result based on the face image 23 was "male".

In the next step ST105, do the two estimation results of the estimation result based on the whole body image (estimation result of the whole body image classifier) and the estimation result based on the face image (estimation result of the face image classifier) match each other? Judge whether or not. If it is determined that the two estimation results match each other (step ST105: Yes), the process proceeds to step ST106, and if it is determined that the two estimation results do not match each other (step ST105: No), the process proceeds to step ST107. .. In the example of FIG. 5, in the case of FIG. 5A, since the two estimation results match each other, the process proceeds to step ST106, and in the case of FIG. 5B, the two estimation results do not match each other, so the process proceeds to step ST107. move on.

In step ST106, the whole body image is added as highly reliable learning data, and online learning (additional learning) of the discriminator for the whole body image is performed. The correct label uses estimation results based on facial and whole body images. In the example of FIG. 5A, the whole body image 22 is added as highly reliable learning data, and the correct label is “female”.

In step ST107, the whole body image is added as learning data with low reliability, and online learning (additional learning) of the discriminator for the whole body image is performed. The correct label uses the estimation result based on the facial image. In the example of FIG. 5B, the whole body image 32 is added as learning data with low reliability, and the correct answer label is “male”.

As mentioned above, the reliability of the face image is higher than that of the whole body image, so if the estimation result based on the face image and the estimation result based on the whole body image do not match, the estimation result based on the face image is trusted and the correct answer label is given. Used as. If the estimation result based on the face image and the estimation result based on the whole body image do not match, the reliability of the whole body image, which is the learning data, matches the estimation result based on the face image and the estimation result based on the whole body image. Lower than if. If the estimation result from the face image has high reliability by itself, the reliability may not be lowered.

As described above, according to the above-mentioned classifier learning process, the reliability of the whole body image, which is the learning data, is changed based on whether or not the estimation result based on the whole body image and the estimation result based on the face image (personal information) match. This makes it possible to realize highly accurate online learning. This makes it possible to provide a classifier capable of accurately estimating the person's attribute based on the whole body image of the person.

If the estimation result based on the whole body image and the estimation result based on the face image match as in the example of FIG. 5A, the learning effect is considered to be small, and the whole body image is excluded from the training data. You may try to do it. If the size of the face image or the orientation of the face included in the face image is out of the predetermined allowable range, the whole body image including the face image is excluded from the training data, or the reliability of the whole body image is removed. You may try to lower.

FIG. 6 is a flow chart showing the flow of the person attribute estimation process in the classifier learning device 3. The following processing is performed by the processor 11 of the discriminator learning device 3 as in the case of the discriminator learning process described above, and is started when a captured image of a person is acquired from the image pickup device 2.

First, a whole-body image of a person is extracted from the captured image (step ST201). Subsequently, an attempt is made to extract a person's face image from the whole body image (step ST202). Since the method of extracting the whole body image and the method of extracting the face image are the same as those of step ST101 and step ST102 of the discriminator learning process described with reference to FIG. 3, the description thereof will be omitted.

In the next step ST203, it is determined whether or not the face image has been extracted. If it is determined in step ST203 that the face image has been extracted (step ST203: Yes), the process proceeds to step ST204, and if it is determined that the face image has not been extracted (step ST203: No), the process proceeds to step ST205. When a person is imaged by the image pickup device 2, if the person is not facing the image pickup device 2 (for example, the back is turned to the image pickup device 2), or the face area of the person is a shield or another person. If it is obstructed by such as, the face area of the person is not shown in the captured image, so that the face image is not extracted.

In step ST204, the person attribute (gender) is estimated based on the face image by using the above-mentioned face image classifier. In step ST205, the person attribute is estimated based on the whole body image by using the above-mentioned whole body image classifier. Since the method of estimating the person attribute based on the face image and the method of estimating the person attribute based on the whole body image are the same as those in steps ST103 and ST104 of the discriminator learning process described with reference to FIG. 3, the description thereof will be omitted.

As described above, according to the above-mentioned person attribute estimation process, even if the face image cannot be extracted from the whole body image, the person attribute of the person can be estimated based on the whole body image. Further, as described above, since the discriminator for the whole body image is constructed by highly accurate online learning, it is possible to estimate the person attribute based on the whole body image with high accuracy.

Although the present disclosure has been described above based on specific embodiments, these embodiments are merely examples, and the present disclosure is not limited to these embodiments. Further, all the components of the discriminator learning device and the discriminator learning method according to the present disclosure shown in the above embodiment are not necessarily all indispensable, and may be appropriately selected as long as they do not deviate from the scope of the present disclosure. It is possible.

For example, in the present embodiment, the whole body image and the face image are extracted from the captured image acquired from the image pickup device 2, but the whole body image and the face image may be acquired from an external device (not shown) other than the image pickup device 2. good.

Further, although gender is estimated as a person attribute in the present embodiment, the person attribute is not limited to gender and may be a year group (age band), race, or the like.

Further, in the present embodiment, when the estimation result based on the face image and the estimation result based on the whole body image do not match, the learning is performed as learning data with lower reliability than when both match, but the learning is limited to this. It's not a thing. For example, when simplicity is required rather than learning accuracy, such as when the performance of the classifier is low, learning is performed as learning data with the same degree of reliability as when both match and when they do not match. May be good. Further, when it is desired to exclude uncertain learning data from the learning target, the whole body image classifier may be trained only when both match. Further, since what kind of data should be trained also changes depending on the maturity of learning of the discriminator, the conditions and reliability for training the estimation result in the discriminator for whole body image may be changed at any time. ..

Further, in the present embodiment, there are cases where the estimation result based on the face image and the estimation result based on the whole body image match, and the case where the estimation result based on the face image and the estimation result based on the whole body image do not match. The discriminator to be trained may be a different discriminator. By doing so, it is possible to output which training data the classifier based on the estimation result based only on the whole body image is estimated, and by extension, notify the reliability of the estimation result. It becomes possible.

Further, in the present embodiment, the face image is used as the person information other than the whole body image, but instead of the face image, for example, a beacon signal including the person attribute information may be used. In this case, it is preferable to receive the beacon signal from a smart phone or the like possessed by the person and use the person attribute information included in the beacon signal as the person information.

Further, although the whole body image is used in this embodiment, it is also possible to use information other than the whole body image instead of the whole body image. For example, even when an image in which a part such as a toe or a fingertip is missing from a whole body image is used, an estimation result substantially equivalent to that of the whole body image can be obtained. In addition, it is possible to estimate the person attribute even when using an image in which many parts are missing from the whole body image such as the hairstyle and the shape of the upper body.

That is, the face image and the whole body image in this embodiment are merely examples. The idea of the present disclosure is various combinations as long as it is a combination of information (for example, a face image) that can obtain a highly reliable estimation result with few detectable situations and information that has many detectable situations (for example, a whole body image). Applicable to. However, since the whole body image in the present embodiment is information used as an alternative when information (for example, a face image) capable of estimating the person attribute of the captured person with high accuracy cannot be obtained, the whole body is used. Even when other information is used in place of the image, it is desirable that such other information is information obtained from the captured image.

The discriminator learning device and the discriminator learning method according to the present disclosure make it possible to provide a discriminator capable of accurately estimating the person attribute of a person based on a person image including at least a part of the person. It is useful as an instrument learning device and a discriminator learning method.

1 Discriminator learning system 2 Imaging device 3 Discriminator learning device 4 Network 11 Processor 12 RAM
13 ROM
14 Storage unit 15 Display unit 16 Input unit 17 Communication unit 21, 31 Captured image 22, 32 Whole body image (human image)
23, 33 Face image (personal information)

Claims (8)

With the processor
It is equipped with an image pickup device and a communication unit that executes communication.
The processor
From the image pickup apparatus, a whole-body image including at least a part other than the face of the estimation target person who is the estimation target of gender or age , and a face image including the face of the estimation target person are acquired.
Machine learning was previously constructed by, the person image for classifier analyzes the input whole body imaging and outputs the estimation result of the sex or age of the person included in the input whole body imaging, whole body imaging of the estimated target person By inputting, the gender or age of the person to be estimated is estimated, and
Previously constructed by machine learning, for personal information discriminator to estimate the gender or age of a person included in analyzes the input face image is the input face image, and inputs a face image of the estimated target person By estimating the gender or age of the person to be estimated,
It is determined whether or not the estimation result based on the whole body image and the estimation result based on the face image match.
When it is determined that the estimation result based on the whole body image and the estimation result based on the face image match, the estimation result determined to match is used as the correct answer label, and the whole body image is used as the person image classifier. A classifier learning device characterized by additional learning. When it is determined that the estimation result based on the whole body image and the estimation result based on the face image do not match, the estimation result based on the face image is used as the correct answer label, and the whole body image is used as the person image classifier. The classifier learning device according to claim 1, wherein additional learning is performed. When it is determined that the estimation result based on the whole body image and the estimation result based on the face image do not match, the estimation result based on the whole body image and the estimation result based on the face image match the whole body image. The discriminator learning device according to claim 2, wherein the discriminator for a person image is additionally trained as learning data having a lower reliability than in the case of the above. The processor
If it is the face image is extracted from the pre-Symbol whole body imaging, using a discriminator the person information, estimates the gender or age of the estimated target person based on the face image,
When said from whole body imaging face image is not extracted, claim 1 using a classifier for the person image, and estimates the gender or age of the estimated target person based on the whole body image Or the classifier learning device according to any one of claims 3. A step of acquiring a full-body image including at least a part other than the face of the estimation target person who is the estimation target of gender or age and a face image including the face of the estimation target person from the image pickup apparatus.
A whole-body image of the estimated target person is used in a person image classifier that analyzes the input whole-body image and outputs the estimation result of the gender or age of the person included in the input whole-body image, which is constructed in advance by machine learning. And the step of estimating the gender or age of the person to be estimated by inputting
Previously constructed by machine learning, for personal information discriminator to estimate the gender or age of a person included in analyzes the input face image is the input face image, and inputs a face image of the estimated target person By doing so, the step of estimating the gender or age of the person to be estimated,
A step of determining whether the estimation result based on the whole body image and the face estimated results based on the images match,
When it is determined that the estimation result based on the whole body image and the estimation result based on the face image match, the estimation result determined to match is used as the correct answer label, and the whole body image is used as the person image classifier. A discriminator learning method characterized by having a step for additional learning. When it is determined that the estimation result based on the whole body image and the estimation result based on the face image do not match, the estimation result based on the face image is set as the correct answer level, and the whole body image is used as the person image classifier. classifier learning method according to claim 5, further comprising the step of additional learning. When it is determined that the estimation result based on the whole body image and the estimation result based on the face image do not match, the estimation result based on the whole body image and the estimation result based on the face image match the whole body image. The discriminator learning method according to claim 6 , wherein the discriminator for a person image is additionally trained as learning data having a lower reliability than the case where the data is performed. When the face image from the previous SL systemic image is extracted, a step of using a classifier the person information, estimates the gender or age of the estimated target person based on the face image,
When said from whole body imaging face image is not extracted, further comprising the step of using a classifier for the person image, estimates the gender or age of the estimated target person based on the whole body image The classifier learning method according to any one of claims 5 to 7.

Images