JP7426922B2 - Program, device, and method for artificially generating a new teacher image with an attachment worn on a person's face - Google Patents

Description

The present invention relates to a technique for artificially generating a new teacher image. It is particularly suitable for generating a face image of a person wearing a mask or the like as a teacher image.

Machine learning engine technology that recognizes people and objects from captured images is progressing. In particular, the accuracy of face recognition for recognizing a person from a face image is rapidly improving with the development of deep learning technology. For example, Facebook announced that the accuracy of DeepFace (registered trademark), a face recognition technology using deep learning, has reached 97.35% (for example, see Non-Patent Document 1).
In addition, in order to train the learning model within a machine learning engine, it is necessary to use a large amount of teacher images. For example, Affectiva uses approximately 7 billion emotional features collected from more than 87 countries around the world. As a result, more accurate emotion recognition technology has been realized (see, for example, Non-Patent Document 2).

BACKGROUND ART Conventionally, there is a technology that learns features of a large number of facial images for each emotion and recognizes the emotion based on the characteristics (for example, see Patent Document 1). Specifically, there are the Ekman 7-class facial expression model (neutral, joy, disgust, anger, surprise, sadness, fear) and the 3-class emotional model of positive, negative, and neutral.

There is also a technology that defines multiple recognition modes based on the state of the target person, has a recognizer for each recognition mode, and applies any one recognizer according to the recognition mode during face recognition (for example, patented (See Reference 2). The current face condition of the target person includes whether or not he or she is wearing a mask, glasses, sunglasses, hat, or the like. According to this technique, a recognition mode is selected from the current closed area of the face, and a recognizer based on the selected recognition mode determines whether authentication is successful or not. That is, each recognizer is trained from teacher images with different closed regions.

Furthermore, there is also a technique of "facial expression recognition AI (Artificial Intelligence)" developed by the applicant of the present application (for example, see Non-Patent Document 3).
In the past, facial recognition technology has been used not only for unlocking devices based on successful facial recognition, but also for marketing purposes, such as automatic photo-taking functions based on smile detection, and acceptance surveys based on facial expression analysis of TV program viewers. However, because it detects many parts of the human face as clues, it can only handle faces facing forward where both eyes are clearly visible.
On the other hand, according to the machine learning "multi-angle adaptive model control technology" described in Non-Patent Document 3, it is possible to analyze facial expressions with high accuracy regardless of the orientation of the face (angle-free). . It can accurately recognize facial expressions even when the face is facing directly to the side.
According to this technology, using the de facto standard face image dataset LFW (Labeled Faces in the Wild), even facial images with a face orientation of 45 degrees or more and only one eye visible can be Facial expressions can be recognized with high accuracy.

Japanese Patent Application Publication No. 2011-150381 Japanese Patent Application Publication No. 2018-165983

Taigman, Yaniv, et al. "Deepface: Closing the gap to human-level performance in face verification." Proceedings of the IEEE conference on computer vision and pattern recognition. 2014. affectiva, [online], [searched on November 8, 2020], Internet <URL: https://affectiva.jp/reason.html> Angle-free facial expression recognition AI, [online], [searched on November 8, 2020], Internet <URL: https://www.kddi-research.jp/newsrelease/2018/080201.html> shape_predictor_68_face_landmarks.dat, [online], [searched on November 8, 2020], Internet <URL: http://dlib.net/files/> Facial point annotations, [online], [searched on November 8, 2020], Internet <URL: https://ibug.doc.ic.ac.uk/resources/facial-point-annotations/>

In recent years, since the outbreak of the new coronavirus infection, wearing masks and goggles on the face has become commonplace. When such a device is worn on the face, up to 70% of the facial area is covered. This has led to the problem that faces and expressions cannot be fully recognized. This is because, according to a general face recognition algorithm, it is necessary to capture as many features as possible from a face image.

Of course, a machine learning engine for facial recognition can be trained on a large number of teacher images of faces and facial expressions wearing masks or goggles. However, it is currently extremely difficult to collect a large amount of such teacher images of faces.

All of the techniques described in Non-Patent Documents 1 and 2 and Patent Documents 1 and 2 described above do not take into account the direction or inclination of the face at all. Furthermore, although the technique described in Non-Patent Document 3 takes into consideration the orientation and inclination of the face, when a mask or goggles are worn, the face and facial expressions cannot be sufficiently recognized.

In response to this, the inventors of the present application have wondered if it is possible to input a teacher image in which a person's face is reflected and artificially generate an image of the person wearing an attachment as a new teacher image. I thought.

Therefore, the present invention provides a program and an artificial teacher image generation device that can input a teacher image in which a person's face is reflected and artificially generate an image of the person wearing an attachment on the person's face as a new teacher image. and a method.

According to the present invention, there is provided a program that operates a computer to input a teacher image in which a person's face is reflected, and to artificially generate an image of a person wearing an attachment on the person's face as a new teacher image. ,
a fitted object image storage means for storing in advance a plurality of worn object images with different orientations;
face area detection means for detecting a human face area from the input teacher image;
Feature point detection means for detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a wearable object image selection means for selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
The computer is characterized in that the computer functions as an artificial teacher image generating means for generating an artificial teacher image in which the selected attachment image is rotated according to the inclination of the face and superimposed on the human face area of the teacher image. .

According to another embodiment of the program of the present invention,
It is also preferable that the wearable item is a mask, glasses, goggles, or sunglasses for the computer to function.

According to another embodiment of the program of the present invention,
A computer is used as a wearing object determining means for estimating whether or not a wearing object is already worn at the feature point of the facial area output from the feature point detecting means and outputting only the facial area where no wearing object is worn. It is also preferable to make it function further.

According to another embodiment of the program of the present invention,
The wearable object image selection means selects a plurality of wearable object images having the same orientation and different orientations;
It is also preferable that the artificial teacher image generation means causes the computer to function so as to generate a plurality of artificial teacher images in which each of the plurality of wearable object images is superimposed on the human face area of the teacher image.

According to another embodiment of the program of the present invention,
The artificial teacher image generation means may cause the computer to perform image processing on the single generated artificial teacher image so that the illuminance, color, or contrast is different, thereby generating a plurality of artificial teacher images. preferable.

According to another embodiment of the program of the present invention,
It is also preferable that the artificial teacher image generation means causes the computer to function so as to cut out the wearable object image that is not visible depending on the orientation and inclination of the face when superimposing the wearable object image on the human face area of the teacher image. .

According to the present invention, there is provided an artificial teacher image generation device that inputs a teacher image in which a person's face is reflected and artificially generates an image of a person wearing an attachment on the person's face as a new teacher image,
a fitted object image storage means for storing in advance a plurality of worn object images with different orientations;
face area detection means for detecting a human face area from the input teacher image;
Feature point detection means for detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a wearable object image selection means for selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
The present invention is characterized by comprising an artificial teacher image generation means for generating an artificial teacher image in which the selected wearing object image is rotated according to the inclination of the face and superimposed on the human face area of the teacher image.

According to the present invention, there is provided an artificial teacher image generation method using a device that inputs a teacher image in which a person's face is reflected and artificially generates an image of the person wearing an attachment on the person's face as a new teacher image. ,
The device is
It has a fitted object image storage section that stores in advance a plurality of worn object images in different orientations,
A first step of detecting a human face area from the input teacher image;
a second step of detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a third step of selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
The present invention is characterized by executing a fourth step of rotating the selected attachment image according to the inclination of the face and generating an artificial teacher image superimposed on the human face area of the teacher image.

According to the program, artificial teacher image generation device, and method of the present invention, a teacher image in which a person's face is reflected is input, and an image of the person wearing an attachment on the person's face is artificially generated as a new teacher image. be able to.

1 is a functional configuration diagram of an artificial teacher image generation device according to the present invention. FIG. 3 is an explanatory diagram of a face area detection section. FIG. 3 is an explanatory diagram of a feature point detection section. It is an explanatory diagram showing direction and inclination of a face. FIG. 3 is an explanatory diagram of an attached object image storage section. It is an explanatory view showing direction and inclination of a wearing object. FIG. 3 is an explanatory diagram of an artificial teacher image generation unit. FIG. 3 is an explanatory diagram of a worn object determination section. FIG. 2 is a first functional configuration diagram of a face authentication device using an artificial teacher image. FIG. 2 is a second functional configuration diagram of a face authentication device using an artificial teacher image.

Hereinafter, embodiments of the present invention will be described in detail using the drawings.

FIG. 1 is a functional configuration diagram of an artificial teacher image generation device according to the present invention.

The artificial teacher image generation device 1 inputs a teacher image in which a person's face is reflected, and artificially generates an image of the person's face wearing an attachment as a new teacher image.
According to FIG. 1, the artificial teacher image generation device 1 includes a teacher image storage section 10, a face area detection section 11, a feature point detection section 12, a wearable object image storage section 130, and a wearable object image selection section 13. , and an artificial teacher image generation unit 14. Furthermore, as an option, it may further include a worn object determination section 121. These functional components are realized by executing a program that causes a computer installed in the device to function. Further, the processing flow of these functional components can be understood as an artificial teacher image generation method of the device.

[Teacher image storage unit 10]
The teacher image storage unit 10 stores teacher images in which human faces are reflected. It is assumed that the person's face in the teacher image is basically not wearing a mask, glasses, goggles, or sunglasses.
According to FIG. 1, when one teacher image is input from the teacher image storage unit 10, the artificial teacher image generation device 1 eventually outputs one or more artificial teacher images. By repeating this process, it is possible to generate a large number of artificial facial images with the wearer on from facial images without the wearer.
In the embodiment of the present invention, the attachment will be described as a "mask," but it may of course be glasses, goggles, or sunglasses.

As mentioned above, the teacher image is an image in which no attachments are worn on the person's face in the teacher image, but if the optional attachment determination unit 121 functions, it is determined that the teacher image has no attachments on the person's face. May be worn. The teacher image in which the face of the person wearing the attachment is reflected is excluded by the attachment determination unit 121 from the generation target of the artificial teacher image.
With this, it is possible to generate a large number of artificial facial images with the wearer wearing only the face images without the wearer.

[Face area detection unit 11]
The face area detection unit 11 detects a human face area (for example, a bounding box) from the input teacher image. The detected face area is output to the feature point detection section 12.

FIG. 2 is an explanatory diagram of the face area detection section.
According to FIG. 2(a), the faces of two women are reflected in the teacher image.
According to FIG. 2(b), the face area detection unit 11 detects the face areas of two people.

Specifically, the face region detection unit 11 uses R-CNN (Regions with Convolutional Neural Networks) or SSD (Single Shot Multibox Detector).
R-CNN combines rectangular face regions with convolutional neural network features to detect a subset of face regions (region proposal). Next, CNN features are extracted from the region proposal. Then, the bounding box of the region proposal is adjusted using a support vector machine trained in advance using CNN features.
SSD is a general object detection algorithm using machine learning, and determines rectangular bounding boxes called default boxes. A large number of default boxes of different sizes are superimposed on one image, and a predicted value is calculated for each box. For each default box, it is possible to predict how far it is from the object and how different its size will be.

[Feature point detection unit 12]
The feature point detection unit 12 detects "feature points" (for example, 68 key points) from the detected face area, and detects "face orientation and inclination" from the feature points.

FIG. 3 is an explanatory diagram of the feature point detection section.

(Detection of feature points)
First, each part (eyes, nose, mouth, eyebrows, chin) is detected from the facial area, and feature points are extracted from the outline of that part.
For example, the official website of Dlib discloses a trained model of facial feature points (for example, see Non-Patent Document 4). This learned model was trained on a dataset with 3837 annotations (train:3148/test:689) (see, for example, Non-Patent Document 5).

(Detection of face direction and inclination)
According to FIG. 3(a), the direction of the face is "to the right" when viewed from the front, and the inclination of the face is "down 30 degrees" when viewed from the front perpendicular.
According to FIG. 3(b), the face is facing forward, and the inclination of the face is the front perpendicular.

FIG. 4 is an explanatory diagram showing the direction and inclination of the face.

According to FIG. 4A, the distances between the "end point of the left cheek" and the "end point of the right cheek" are calculated from the coordinate points of the point at the center of the nose.
If the distance between the end point of the right jaw and the center of the nose is longer than the distance between the end point of the left jaw and the center of the nose by at least a first predetermined threshold, the face direction is determined to be "towards the left". judge.
Leftward: Distance between the end point of the left jaw and the center of the nose < Distance between the end point of the right jaw and the center of the nose Similarly, the distance between the end point of the left jaw and the center of the nose is If the distance is longer than the distance to the center of the nose by a first predetermined threshold value or more, the face direction is determined to be "to the right."
Right direction: Distance between the end point of the left jaw and the center of the nose > Distance between the end point of the right jaw and the center of the nose Then, the distance between the end point of the left jaw and the center of the nose is If the difference from the distance to the center is small and equal to or less than the second predetermined threshold (if the two distances almost match), the face orientation is determined to be "front facing".
In addition, as another embodiment, the face direction is not limited to "rightward facing", "leftward facing", and "frontward facing", but may also have an angle depending on the direction. In that case, the angle may be determined according to the ratio of the distance between the end point of the right jaw and the center of the nose to the distance between the end point of the left jaw and the center of the nose. For example, even if the object is ``to the right'', it can be detected in stages such as ``a little to the right,'' ``45 degrees to the right,'' and ``completely to the right.''

According to FIG. 4(b), the angle between the front perpendicular and the nose center line is expressed as a slope. Here, it is expressed as an inclination of 30 degrees clockwise with respect to the front perpendicular.

[Weared object image storage unit 130]
The wearable object image storage section 130 stores in advance a plurality of wearable object images according to the orientation of the face.

FIG. 5 is an explanatory diagram of the attached object image storage section.

According to FIG. 5, images of a wearable object "mask" are stored in the wearable object image storage section 130. Here, a front-facing mask image and a right-facing mask image are accumulated. Of course, it is also preferable that a plurality of mask images having the same orientation but different colors and shapes are stored. In the case of masks, there are various shapes such as cloth, non-woven fabric, flat type, pleat type, three-dimensional type, etc.
Of course, as another embodiment, even if the mask images are of the same type, they may be accumulated in stages such as "slightly rightward,""45 degrees rightward," and "completely rightward."

[Weared object image selection unit 13]
The wearable item image selection unit 13 selects a wearable item image according to the orientation of the face from the wearable item image storage unit 130.
Furthermore, the wearable object image selection unit 13 may select a plurality of wearable object (mask) images that have the same face orientation but differ from each other.

[Artificial teacher image generation unit 14]
The artificial teacher image generation unit 14 generates an artificial teacher image in which the selected wearable object image is rotated according to the inclination of the face and superimposed on the human face area of the teacher image.

FIG. 6 is an explanatory diagram showing the orientation and inclination of the worn object.

According to FIG. 6(a), the face is "facing right" and "not tilted". At this time, a right-facing mask image is selected from the wearable object image storage section 130. Then, for the right-facing mask image, the mask is set according to the left side distance (distance between the left cheek feature point and the center of the nose) and right side distance (distance between the right cheek feature point and the nose center) of the face. The size of the image is rescaled (enlarged/reduced). Thereby, the width and height of the attachment image are adjusted to match the width and height of the wearing area of the face (the area surrounded from the center of the nose and the end points of the lower, left, and right jaws).

According to FIG. 6(b), the face is "facing right" and "tilted 30 degrees clockwise." Here, unlike in FIG. 6(a), the face is tilted. For this purpose, the mask image is further tilted 30 degrees clockwise with respect to the nose center line (the straight line between the nose center and the chin).

FIG. 7 is an explanatory diagram of the artificial teacher image generation section.

According to FIG. 7, the artificial teacher image generation unit 14 generates the teacher image so that the orientation and inclination of the feature points of the human face in the teacher image match the orientation and inclination of the wearable object image (for example, a mask image). The image of the attached object is superimposed on the face area of the person.

Further, as another embodiment, when a plurality of wearable object images are selected by the wearable object image selection section 13, the artificial teacher image generation section 14 generates a wearable object image on a person's face reflected in one teacher image. A plurality of artificial teacher images may be generated by superimposing them. For example, by superimposing mask images of different types, colors, and shapes on a person's facial area, it is possible to generate a large number of padded artificial teacher images.

Furthermore, as another embodiment, the artificial teacher image generation unit 14 performs image processing on one generated artificial teacher image so that the illuminance, color, or contrast (including segmentation processing) is different, and generates a plurality of artificial teacher images. It may also be possible to generate an artificial teacher image. This makes it possible to generate a variety of artificial teacher images.

Furthermore, as another embodiment, when superimposing the wearable object image on the human face area of the teacher image, the artificial teacher image generation unit 14 cuts out the wearable object image that is not visible depending on the orientation and inclination of the face. It may be something. For example, in order to imagine a realistic usage scenario, for a left-right mask, a hook belt that goes over one ear on the side facing the face is cut.

As another embodiment, if the wearable object image selected by the wearable object image selection section 13 is already in accordance with the inclination of the face, the artificial teacher image generation section 14 converts the fitted object image into a teacher image. The image is simply superimposed on the face area of the person.
For example, if the wearable object image stored in the wearable object image storage section 130 is a three-dimensional model, the wearable object image selection section 13 rotates the three-dimensional model according to the orientation and inclination of the face, and photographs the image at that position. A two-dimensional image of the attached object may be output.

[Weared object determination unit 121]
According to the embodiment described above, it is assumed that the person's face reflected in the teacher image input to the face area detection unit 11 is not wearing any attachments. Therefore, the feature point detection unit 12 can obtain feature points of the entire face even if the face has orientation and inclination.
On the other hand, the wearable object determination unit 121 is an optional unit that estimates whether or not a wearable item is already worn at the feature point of the facial area output from the feature point detection unit 12. Only the facial area where no attachment is worn is output. That is, the face area where the attachment is already worn is excluded.

FIG. 8 is an explanatory diagram of the attached object determination section.

According to FIG. 8, it is assumed that, for example, a teacher image in which a person wearing a mask is reflected is input to the face area detection unit 11. The face area detection unit 11 detects a face area wearing a mask, and the feature point detection unit 12 detects feature points only for face areas other than the closed area of the mask. In the case of a trout, the feature points around the mouth cannot be detected.
Specifically, the wearing object determination unit 121 uses a support vector trained in advance using the feature points of a face not wearing a mask (68 points) and the feature points of a face wearing a mask (<68 points) as training data. It may be a machine. Thereby, the wearable item determination unit 121 can classify and estimate whether or not a wearable item is already being worn in the face area output from the feature point detection unit 12. Then, only the facial region where no attachment is worn is output.
Note that the wearable object determination unit 121 is not limited to a support vector machine, and may be any machine learning engine that can classify two categories, ie, presence or absence of a wearable object, or may be a simple branch determination function of feature point group coordinates.

FIG. 9 is a first functional configuration diagram of a face authentication device using an artificial teacher image.

According to FIG. 9, the face recognition device 2 inputs a recognition target image and outputs a recognized user ID. The recognition target image input to the face recognition device 2 includes a human face, and the face may or may not be wearing an accessory.
As described above with reference to FIG. 1, the face recognition engine 15 is trained in advance using a mixture of teacher images and padded artificial teacher images.
As the face recognition engine 15, for example, Facenet (registered trademark) of Google (registered trademark) can be used. As a result, learning is performed using multidimensional vectors (128/256/521 bytes) converted from the face region. It is possible to estimate the user ID whose multidimensional vector has the closest Euclidean distance.

The face recognition engine 15 may be, for example, the facial expression recognition AI described in Non-Patent Document 3 mentioned above. Facial expression recognition AI can analyze facial expressions with high accuracy regardless of the orientation of the face, but it can also recognize facial expressions even when wearing an accessory. Facial expression recognition AI is trained in two stages, separating facial orientation and facial expression recognition estimation. In particular, three types of facial expression recognition models (positive, negative, and neutral) are applied depending on the direction of the face (up, down, left, right, middle) and inclination. According to the present invention, it is possible to generate an artificial teacher image in which the user is wearing clothing according to the direction and inclination of the face. As a result, even if the target user is wearing something on his or her face, facial expressions can be estimated as much as possible.

FIG. 10 is a second functional configuration diagram of a face authentication device using an artificial teacher image.

In FIG. 10, similarly to FIG. 9, the face recognition device 2 inputs the recognition target image and outputs the recognized user ID. The recognition target image input to the face recognition device 2 includes a human face, and the face may or may not be wearing an accessory.
According to FIG. 10, the face recognition engine 151 is trained in advance using a teacher image in which a face is not wearing an attachment, and the face recognition engine 152 is trained in advance using an image in which a face wearing an attachment is reflected. It was trained in advance using artificial teacher images that were inflated to .

The face recognition device 2 includes a face area detection section 11 , a feature point detection section 12 , a worn object determination section 121 , a face recognition engine 151 , and a face recognition engine 152 .
The face area detection unit 11 detects a human face area from the input recognition target image, as described above.
The feature point detection unit 12 detects feature points from the detected face area, as described above.
Similarly to the above-mentioned unit, the wearable item determination unit 121 also estimates whether or not a wearable item is already worn at a feature point in the face area. A facial image without wearing items is input to the face recognition engine 151, and a facial image with wearing items is input to the face recognition engine 152. Recognition accuracy can be improved by separately estimating the case without an attached item and the case with an attached item.

As described in detail above, according to the program, artificial teacher image generation device, and method of the present invention, a teacher image in which a person's face is reflected is input, and a new image in which an attachment is worn on the person's face is generated. It can be artificially generated as a training image.
This makes it possible to simply train an existing face recognition engine with the artificial teacher images generated in an inflated manner according to the present invention, without having to collect a large number of teacher images in which the person wears the object on a person's face. It becomes possible to recognize facial images in which the wearer is wearing a mask.

Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and viewpoint of the present invention can be easily made by those skilled in the art. The above description is merely an example and is not intended to be limiting in any way. The invention is limited only by the claims and their equivalents.

1 Artificial teacher image generation device 10 Teacher image storage section 11 Face area detection section 12 Feature point detection section 121 Wearing object determination section 130 Wearing object image storage section 13 Wearing object image selection section 14 Artificial teacher image generation section 15 Face recognition engine 151 Wearing Facial recognition engine for use without objects 152 Face recognition engine for use with attached items 2 Face recognition device

Claims (8)

A program that operates a computer to input a teacher image in which a person's face is reflected, and to artificially generate an image of the person wearing an object as a new teacher image,
a fitted object image storage means for storing in advance a plurality of worn object images with different orientations;
face area detection means for detecting a human face area from the input teacher image;
Feature point detection means for detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a wearable object image selection means for selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
The computer is characterized in that the computer functions as an artificial teacher image generating means for generating an artificial teacher image in which the selected attachment image is rotated according to the inclination of the face and superimposed on the human face area of the teacher image. program. 2. The program according to claim 1, wherein the wearable item causes the computer to function as if it were a mask, glasses, goggles, or sunglasses. A computer is used as a wearing object determining means for estimating whether or not a wearing object is already worn at the feature point of the facial area output from the feature point detecting means and outputting only the facial area where no wearing object is worn. The program according to claim 1 or 2, further comprising a function. The wearable object image selection means selects a plurality of wearable object images having the same orientation and different orientations;
4. The artificial teacher image generating means causes the computer to function to generate a plurality of artificial teacher images in which each of the plurality of wearable object images is superimposed on the human face area of the teacher image. The program described in any one of the paragraphs. The artificial teacher image generation means performs image processing on the single generated artificial teacher image so that illuminance, color, or contrast differs, and causes the computer to function to generate a plurality of artificial teacher images. The program according to any one of claims 1 to 4. The artificial teacher image generation means is characterized in that when superimposing the wearable object image on the human face area of the teacher image, the computer functions to cut out the wearable object image that is not visible according to the orientation and inclination of the face. The program according to any one of claims 1 to 5. An artificial teacher image generation device that inputs a teacher image in which a person's face is reflected and artificially generates an image of a person wearing an attachment on the person's face as a new teacher image,
a fitted object image storage means for storing in advance a plurality of worn object images with different orientations;
face area detection means for detecting a human face area from the input teacher image;
Feature point detection means for detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a wearable object image selection means for selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
an artificial teacher image generating means for generating an artificial teacher image in which the selected attachment image is rotated according to the inclination of the face and superimposed on the human face area of the teacher image; Image generation device. An artificial teacher image generation method using a device that inputs a teacher image in which a person's face is reflected and artificially generates an image of a person wearing an attachment on the person's face as a new teacher image, the method comprising:
The device is
It has a fitted object image storage section that stores in advance a plurality of worn object images in different orientations,
A first step of detecting a human face area from the input teacher image;
a second step of detecting feature points from the detected face area and detecting the orientation and inclination of the face from the feature points;
a third step of selecting a wearable object image according to the orientation of the face from the wearable object image storage means;
and a fourth step of rotating the selected wearable object image according to the inclination of the face and generating an artificial teacher image superimposed on the human face area of the teacher image. Artificial teacher image generation method.

Images