KR20210048623A - Apparatus and method for estimating age using modified cycle-consistent adversarial networks - Google Patents

Abstract

The present invention relates to an age estimation technique, and more particularly, to an age recognition technique using a modified cycle-generating adversarial neural network. According to one embodiment of the present invention, through the process of generating a transformed image even using an age estimation neural network which has learned a face image from one database, it is possible to increase the accuracy of age estimation for images in one database and other databases which have low relevance. An age estimation device includes a preprocessing unit, a transformation unit, and an age estimation unit.

Description

A device and method for estimating age using a modified cycle-generated adversarial neural network {APPARATUS AND METHOD FOR ESTIMATING AGE USING MODIFIED CYCLE-CONSISTENT ADVERSARIAL NETWORKS}

The present invention relates to an age estimation technique, and more particularly, to an age recognition technique using a modified cycle-generated hostile neural network.

Recently, a technology for estimating a person's age in real time by referring to a face image is required. It is actively applied in various fields such as marketing by age group, product demand survey, and consumption trend analysis. In addition, with the development of deep learning technology, recent studies on deep feature-based age estimation technology using a convolutional neural network (CNN) rather than a handcrafted feature-based method have been conducted. A lot is being practiced. In general, face images for age estimation have a lot of different characteristics according to changes in objects and environments such as race, camera, and lighting. In most existing deep feature-based age estimation methods, an age estimator is separately trained and used for each database. Therefore, there is a problem in that the recognition performance is deteriorated with respect to a heterogeneous face image that has not been learned. In addition, even if the process of learning together by combining heterogeneous databases to solve this problem, it takes a long time to learn and it is difficult to sufficiently learn about various environmental change factors included in the database, so that the estimation performance of the age estimator is degraded. have.

The background technology of the present invention is disclosed in Korean Patent Publication No. 10-1786754.

The present invention uses a modified cycle-generated hostile neural network to estimate the age of an image of another database having a low correlation with the work database through the process of generating a transformed image even when using an age-estimating neural network that has learned a face image from a work database. It provides an age estimation apparatus and method that improves accuracy.

According to an aspect of the present invention, an age estimation device is provided.

The age estimation apparatus according to an embodiment of the present invention includes a preprocessor for generating an original image by performing a preprocessing process on an input image, and a conversion unit for generating a transformed image using a modified cycle-generated hostile neural network for the original image. And an age estimating unit that estimates the age of the transformed image.

When the age difference between the age of the original image and the transformed image is equal to or greater than a threshold, the age estimating apparatus outputs the age of the original image as the age of the input image, and the age of the original image and the age of the transformed image If the difference is less than a threshold, the conversion image further includes a selection unit for outputting the age of the input image, the age estimating unit may further estimate the age of the original image.

The modified cycle-generated hostile neural network includes a generator and a discrimination unit, and the generator provides a loss function including a loss according to a distance between an image of one database and a feature vector that is an additional output of the generation unit for an image of another database. And the determination unit may be learned using a loss function in consideration of an additional loss obtained by applying an increase function to the additional output of the determination unit for the original image and the transformed image.

The generation unit may be a neural network including a layer outputting the transformed image and a layer outputting the additional output, and the determining unit may be a neural network including a layer outputting the original output and a layer outputting the additional output.

The preprocessor extracts a face area from the input image, calculates an in-plane rotation based on the center position of both eyes in the face area, and rotates the face area according to the inplay rotation angle. Then, the original image may be generated by normalizing the size of the face region.

According to another aspect of the present invention, an age estimation method in an age estimation apparatus is provided.

The age estimation method according to an embodiment of the present invention includes generating an original image by performing a preprocessing process on an input image, generating a transformed image using a cycle-generating hostile neural network modified from the original image, and the It may include estimating the age of the transformed image.

The age estimation method further includes estimating an age for the original image; And when the difference between the age of the original image and the age of the transformed image is greater than or equal to a threshold value, the age of the original image is output as the age of the input image, and a threshold value is applied to the age difference between the age of the original image and the transformed image. If it is less than, the step of outputting the age of the converted image as the age of the input image may further include.

The modified cycle-generated hostile neural network includes a generator and a discrimination unit, and the generator provides a loss function including a loss according to a distance between an image of one database and a feature vector that is an additional output of the generation unit for an image of another database. And the determination unit may be learned using a loss function in consideration of an additional loss obtained by applying an increase function to the additional output of the determination unit for the original image and the transformed image.

The generation unit may be a neural network including a layer outputting the transformed image and a layer outputting the additional output, and the determining unit may be a neural network including a layer outputting the original output and a layer outputting the additional output.

The generating of the original image by performing a preprocessing process on the input image includes: extracting a face region from the input image, and an in-plane rotation angle based on the center positions of both eyes in the face region. Calculating ); It may include rotating the face region according to the in-play rotation angle, and generating the original image by normalizing the size of the face region.

According to another aspect of the present invention, there is provided a computer program that executes the above age estimation method and is recorded on a computer-readable recording medium.

According to an embodiment of the present invention, even if a neural network for estimating faces from one database is used, the accuracy of age estimation for an image of another database having a low correlation with the one database can be improved through the process of generating a transformed image. have.

In addition, according to an embodiment of the present invention, even if the age estimation neural network is trained with only an image of a database corresponding to one style, age estimation can be accurately performed, thereby reducing the cost required for learning the age estimation neural network.

1 is a block diagram schematically illustrating an age estimation apparatus according to an embodiment of the present invention.
2 is a diagram illustrating a preprocessing process performed by an age estimation apparatus according to an embodiment of the present invention.
3 is a diagram illustrating a generation unit of a modified cycle-generating hostile neural network of an age estimation apparatus according to an embodiment of the present invention.
4 is a table illustrating the structure of a generation unit of a modified cycle-generating hostile neural network of an age estimation apparatus according to an embodiment of the present invention.
5 is a diagram for explaining a loss function of a generator of a modified cycle-generated absolute neural network of an age estimation apparatus according to an embodiment of the present invention.
6 is a diagram illustrating a structure of a determination unit of a modified cycle-generating hostile neural network of an age estimation apparatus according to an embodiment of the present invention.
7 is a table illustrating the structure of a discriminator of a modified cycle-generating hostile neural network of an age estimation apparatus according to an embodiment of the present invention.
8 is a diagram for explaining a loss function of a generator of a modified cycle-generating absolute neural network of an age estimation apparatus according to an embodiment of the present invention.
9 is a diagram illustrating the structure of an age estimating neural network of an age estimating apparatus according to an embodiment of the present invention.
10 is a diagram illustrating a process of comparing the age of an original image and a transformed image by an age estimation apparatus according to an embodiment of the present invention.
11 is a diagram illustrating an original image input to an age estimating device and a transformed image generated by the age estimating device according to an embodiment of the present invention.
12 is a flowchart illustrating a method of estimating age by an age estimating apparatus according to an embodiment of the present invention.
13 is a table illustrating the accuracy of age estimation for face images of two databases using an age estimation neural network.
14 is a table illustrating accuracy when the age estimation apparatus estimates age for face images of two databases according to an embodiment of the present invention.

In the present invention, various modifications may be made and various embodiments may be provided. Specific embodiments are illustrated in the drawings and will be described in detail through detailed description. However, this is not intended to limit the present invention to a specific embodiment, it should be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, when it is determined that a detailed description of a related known technology may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted. In addition, the singular expressions used in the specification and claims are to be construed as meaning "one or more" in general, unless otherwise stated.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, and in the description with reference to the accompanying drawings, the same or corresponding components are assigned the same reference numbers, and redundant descriptions thereof will be omitted. It should be.

1 is a block diagram schematically illustrating an age estimating apparatus according to an embodiment of the present invention, and FIG. 2 is a diagram illustrating a preprocessing process performed by an age estimating apparatus according to an embodiment of the present invention, and FIG. 3 is A diagram illustrating a generation unit of a modified cycle-generated hostile neural network of an age estimating apparatus according to an embodiment of the present invention, and FIG. 4 is a diagram illustrating a modified cycle-generating hostile neural network of the age estimating apparatus according to an embodiment of the present invention. A table illustrating the structure of a generator, and FIG. 5 is a diagram for explaining a loss function of a generator of a modified cycle-generated absolute neural network of an age estimation apparatus according to an embodiment of the present invention, and FIG. 6 is an embodiment of the present invention. A diagram illustrating a structure of a determination unit of a modified cycle-generated hostile neural network of an age estimation apparatus according to an embodiment, and FIG. 7 is a diagram illustrating a determination unit of a modified cycle-generating hostile neural network of an age estimating apparatus according to an embodiment of the present invention. A table illustrating the structure, and FIG. 8 is a diagram for explaining a loss function of a generator of a modified cycle-generated absolute neural network of an age estimation apparatus according to an embodiment of the present invention, and FIG. 9 is an embodiment of the present invention. FIG. 10 is a diagram illustrating a structure of an age estimating neural network of an age estimating apparatus according to FIG. 10, and FIG. 10 is a diagram illustrating a process of comparing ages of an original image and a transformed image by the age estimating apparatus according to an embodiment of the present invention.

Referring to FIG. 1, an age estimation apparatus according to an embodiment of the present invention includes a preprocessor 110, a conversion unit 120, an age estimation unit 130, and a selection unit 140.

The preprocessor 110 receives an input image photographed by a person subject to age estimation, and performs preprocessing on the input image. The preprocessor 110 extracts a face area from the input image. In this case, the preprocessor 110 may extract a face area from the input image using a dlib facial box detector. The preprocessor 110 calculates an in-plane rotation based on the center positions of both eyes in the facial area extracted as in Equation 1 below, and rotates the face area according to the inplay rotation angle, The original image is generated by normalizing the size of the face area.

)는 오른쪽 눈의 중심좌표이고, (

)는 왼쪽 눈의 중심좌표를 나타낸다. 전처리부(110)는 계산된

를 바탕으로 얼굴 영역의 너비 및 높이를 재설정하여, 얼굴 영역의 전처리를 수행한다.At this time, (

) Is the center coordinate of the right eye, and (

) Represents the central coordinate of the left eye. The preprocessor 110 is calculated

Based on, the width and height of the face area are reset, and pre-processing of the face area is performed.

For example, the preprocessor 110 may perform a preprocessing process of extracting, rotating, and normalizing a face region as shown in FIG. 2. The preprocessor 110 transmits the original image to the conversion unit 120.

The conversion unit 120 converts the original image into a transformed image using a modified cycle-generating adversarial network (MODIFIED CYCLE-CONSISTENT ADVERSARIAL NETWORKS). At this time, the modified cycle-generating hostile neural network may be a neural network that learns through images of two unrelated groups (for example, group A and group B) and generates an image matching the image of group A to group B. have.

At this time, in a general cycle-generating hostile neural network, if the two sets of data are clearly different from each other, learning does not proceed normally. The modified cycle-generating hostile neural network according to an embodiment of the present invention includes a generator and a discriminator that generate two outputs using a plurality of residual blocks.

Referring to FIG. 3, the generation unit includes an encoder and a decoder. The encoder of a general cycle-generating hostile neural network performs a function of finding useful feature values by reducing the feature dimension, but it is difficult to learn about generating an image because it is a structure based on a neural network. The encoder of the modified cycle-generating hostile neural network according to an embodiment of the present invention generates N additional feature vectors (hereinafter referred to as additional outputs). That is, as illustrated in FIG. 4, the encoder outputs an additional output from the 1st output (Flatten) layer (referred to as the first output layer) after 9 residual blocks, and a final output layer (referred to as the second output layer). An output (referred to as the original output) that is a converted image can be generated at Two deconveolution layers may be located after the first output layer.

(A는 입력 영상, B는 참조 영상)로 학습을 진행 할 때는 두 개의 generator 모델이 사용되고,

(B는 입력 영상, A는 참조 영상)로 학습을 진행 할 때도 위와 마찬가지로 두 개의 generator 모델이 사용될 수 있다.In this case, the conversion unit 120 includes four independent generation units of the above-described structure. Each generation unit may be independently learned without sharing a weight parameter. For example, for example

(A is an input image, B is a reference image) when learning is performed, two generator models are used,

(B is an input image, A is a reference image), as in the above, two generator models can be used.

의 학습을 진행할 때 손실 함수(loss function)을 나타낸다.In this case, the generation unit includes a loss because each of the four generation units performs different tasks. Equation 2 below is

Represents a loss function when learning is performed.

는 A 를 입력으로 받아 generator에서 생성된 이미지를 나타낸다.

Represents the image generated by the generator by receiving A as an input.

In this case, the generation unit may perform learning in a direction of minimizing log likelihood by using a least-squares loss method as shown in Equation 3 below, rather than a method of maximizing log likelihood.

의 과정에서 인코더의 후단에서 제1 출력을 생성하고, 제1 출력 간의 거리 값에 대한 감소 함수를 적용하는 손실 함수를 수행할 수 있다. 이상적인 사이클 생성적 적대 신경망은

로 진행이 되면 두 번째 생성부의 출력이 원래의 A입력과 같아지는 것이다. 하지만, 실제로 학습을 하면 두 번째 생성부의 출력은 원래의 A와는 다른 출력이 된다. 본 발명의 일 실시 예에 따른 생성부는

의 과정 중

의 경우와

가 완벽하게 변환되도록 각각의 생성자의 제1 출력 레이어에서 출력된 특징 벡터들의 거리를 산출한다. 즉, 수학식 4 및 도 5와 같이 생성부는 A를 입력 받은 생성부의 인코더 끝 부분에 출력된 특징 벡터와 B를 입력 받은 생성부의 인코더 끝 부분에 출력된 특징 벡터 사이의 거리가 가까워 지도록 학습을 한다. At this time, the two generating units

In the process of, a loss function of generating a first output at a rear end of the encoder and applying a reduction function to a distance value between the first outputs may be performed. The ideal cycle-generated hostile neural network is

When progressing to, the output of the second generation unit becomes the same as the original A input. However, when actually learning, the output of the second generation unit becomes a different output from the original A. The generator according to an embodiment of the present invention

In the process of

With the case of

The distances of the feature vectors output from the first output layer of each generator are calculated to be completely transformed. That is, as shown in Equations 4 and 5, the generation unit learns so that the distance between the feature vector output to the end of the encoder of the generator receiving A and the feature vector output to the end of the encoder receiving B becomes close. .

는 A를 입력 받은 생성부의 인코더 끝 부분에 출력된 특징 벡터와 B를 입력 받은 생성부의 인코더 끝 부분에 출력된 특징 벡터 사이 간의 거리이고, k는 손실 함수 값의 상수를 의미한다. 또한,

의 과정을 수행하는 다른 2개의 생성부도 상술한 바와 동일한 방식으로 학습될 수 있다.At this time,

Is the distance between the feature vector output to the encoder end of the generator receiving A and the feature vector output to the encoder end of the generator receiving B, and k denotes a constant of the loss function value. Also,

The other two generation units performing the process of can also be learned in the same manner as described above.

In addition, the generation unit additionally uses the loss according to Equation 5.

는

에서 두 번째 생성부에서 출력된 변환 영상이고,

는

에서 두 번째 생성부에서 출력된 변환 영상이다.

Is

Is the converted image output from the second generation unit in,

Is

This is the converted image output from the second generation unit in.

The generation unit applies all the losses shown in Equations 2 to 5 as shown in Equation 6.

That is, the generation unit may be learned by including a loss function including a loss according to a distance between the additional output of the generation unit for the A image and the B image.

의 경우 하기의 수학식 7과 같이 손실을 포함할 수 있다.The conversion unit 120 includes a total of four discrimination units like the generation unit. As shown in FIG. 6, an image generated by each generation unit is input to each determination unit. The discrimination unit

In the case of, loss may be included as shown in Equation 7 below.

At this time, the determination unit may perform learning in a direction that minimizes log likelihood as shown in Equation 8 below, rather than a method of maximizing log likelihood.

In addition, the determination unit performs additional output through the last layer (the 2nd output of FIG. 7) of the determination unit other than the original output, which is a feature vector output by a general cycle-generating hostile neural network. The loss for the additional output can be expressed as in Equation 9.

은 증가 함수에 사용되는 상수 값이며, 도 8에 예시된 바와 같이

는

로 진행할 때 영상 A 에 관한 판별부의 마지막 출력을 뜻하고

는

에서의 A에 대한 변환 영상(fake image)에 관한 판별부의 마지막 출력이다. At this time,

Is a constant value used in the increment function, as illustrated in FIG. 8

Is

When proceeding to, it means the last output of the discrimination section about image A,

Is

This is the final output of the discrimination unit about the fake image for A in.

Accordingly, the determination unit may include a loss such that the distance between the feature vectors output from each of the two layers as illustrated in FIG. 7 increases.

Equation 10 below shows the loss of the above-described discrimination unit.

That is, the determination unit according to an embodiment of the present invention may be trained using a loss function in consideration of an additional loss by applying an increase function to the final output of the determination unit for the original image and the transformed image.

Therefore, since the modified cycle-generated hostile neural network according to an embodiment of the present invention extracts an additional feature vector and uses a loss function considering the additional feature vector, it transforms the image so that the styles between the two groups of images become similar. In addition, it can be learned to convert images according to age. In addition, the transform unit 120 uses the modified cycle-generated hostile neural network to convert the original image, which is a face image of one group, to a face image similar to the age of the original image while corresponding to the style of the face image of another group with little correlation. Converted images can be created.

The age estimating unit 130 estimates the age corresponding to the original image and the transformed image using comparative CNN for age estimation (CCNNAE, referred to as an age estimating neural network). In this case, the age-estimating neural network has a structure as shown in FIG. 9 and may be learned through a loss function configured by measuring similarity between two images in a mini-batch. That is, the age estimation neural network can be trained using an increase function and a decrease function that make the distance value smaller when the age is the same for two images in a mini-batch, and increase the distance value when the age difference is large. I can. In the age-estimating neural network, the last layer is a fully-connected layer, and a feature vector output through the layer can be extracted. The age estimating unit 130 may input the original image to the age estimation neural network and estimate an age of a vector having the closest distance among the output target feature vector and the learned feature vector as the age of the original image. Also, the age estimating unit 130 may estimate the age of the transformed image through the above-described method. The age estimation unit 130 transmits the estimated age of the original image and the transformed image to the selection unit 140.

The selection unit 140 calculates a difference between the estimated age of the original image and the transformed image. When the difference exceeds a preset threshold, the selection unit 140 determines that the converted image has not been normally converted and outputs the age of the original image. In addition, when the difference is less than or equal to a preset threshold, the selection unit 140 determines that the converted image has been normally converted and outputs the age of the converted image. For example, as shown in FIG. 10, when the age of the original image is 20, the age of the transformed image is 38, and the threshold is 10, since the age difference between the original image and the transformed image is 18, the selection unit 140 You can select the age of and print it out.

11 is a diagram illustrating an original image input to an age estimating device and a transformed image generated by the age estimating device according to an embodiment of the present invention.

Referring to FIG. 11, the age estimation apparatus inputs the original image of A into the age estimation apparatus when there is a database A (eg, yellow-in) and a database B (eg, black) including faces by race. In this case, a transformed image converted into a face corresponding to the database B may be generated. Conversely, when the original image of B is input to the age estimating device, the age estimating apparatus may generate a converted image that may belong to A. In this case, the age estimation apparatus may generate a converted image corresponding to the age of each original image, as illustrated in FIG. 11.

12 is a flowchart illustrating a method of estimating age by an age estimating apparatus according to an embodiment of the present invention. The subject of each step to be described below is a process performed by each functional unit constituting the age estimating device or the subject of each step is collectively referred to as an age estimating device for concise and clear explanation.

Referring to FIG. 12, in step 1210, the age estimating apparatus performs pre-processing on the input image to generate an original image. For example, the age estimation apparatus may extract a face area from the input image. The age estimation apparatus may generate an original image by calculating an in-plane rotation angle based on the center positions of both eyes in the extracted face area, rotating the face area according to the in-play rotation angle, and normalizing the size of the face area.

In step 1220, the age estimation apparatus converts the original image into a transformed image using a modified cycle-generating adversarial network (MODIFIED CYCLE-CONSISTENT ADVERSARIAL NETWORKS). The modified cycle-generated hostile neural network includes a generator and a discrimination unit. The generation unit may be learned by including a loss function including a loss according to a distance between an image of one database and a feature vector that is an additional output of the generation unit for an image of another database. In addition, the determination unit may be learned using a loss function in consideration of an additional loss by applying an increase function to the final output of the determination unit for the original image and the transformed image. The detailed structure and learning method of the modified cycle-generating hostile neural network has been described above with reference to FIGS. 3 to 8.

In step 1230, the age estimating apparatus estimates the age of the original image and the transformed image using an age estimating neural network. For example, the age estimating apparatus may input the original image to the age estimating neural network and estimate an age of a vector having the closest distance among the output target feature vector and the learned feature vector as the age of the original image.

In step 1240, the age estimating apparatus determines whether the age difference between the original image and the transformed image is equal to or greater than a preset threshold.

In step 1240, when the age difference is greater than or equal to a preset threshold, in step 1250, the age estimating device outputs the age of the original image.

In step 1240, when the age difference is less than a preset threshold, the age estimating device outputs an age of the converted image in step 1260.

13 is a table illustrating the accuracy of age estimation for face images of two databases using an age estimating neural network, and FIG. 14 is a table illustrating age estimation for face images of two databases by an age estimating apparatus according to an embodiment of the present invention. This is a table illustrating the accuracy of the case.

Referring to FIG. 13, when an age estimation neural network is trained through a work database (MORPH) and age according to an image of another database (MegaAge) is estimated, an MAE, which is the difference between the actual age and the estimated age, was calculated to be 8.6. .

Referring to FIG. 14, when the age estimation neural network is trained through a work database (MORPH) and a transformed image corresponding to an image of another database (MegaAge) is generated through the conversion unit 120 to estimate the age, the actual age The difference between and estimated age was calculated as 7.12.

Therefore, the age estimating apparatus according to an embodiment of the present invention estimates the age of an image of another database having a low correlation with the work database through the process of generating a transformed image even using an age estimating neural network that has learned a face image of a work database. The accuracy of is high.

In addition, the age estimating apparatus according to an embodiment of the present invention can accurately estimate the age even when learning of the age estimating neural network is performed only with images from a database corresponding to one style, thereby reducing the cost required for learning the age estimating neural network. I can.

The above-described method of estimating age may be implemented in computer-readable code on a computer-readable medium. The computer-readable recording medium is, for example, a removable recording medium (CD, DVD, Blu-ray disk, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk). I can. The computer program recorded on the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet and installed in the other computing device, thereby being used in the other computing device.

In the above, even if all the components constituting the embodiments of the present invention are described as being combined into one or operating in combination, the present invention is not necessarily limited to these embodiments. That is, as long as it is within the scope of the object of the present invention, one or more of the components may be selectively combined and operated.

Although the operations are illustrated in a specific order in the drawings, it should not be understood that the operations must be executed in the specific order shown or in a sequential order, or all illustrated operations must be executed to obtain a desired result. In certain situations, multitasking and parallel processing may be advantageous. Moreover, the separation of various components in the above-described embodiments should not be understood as necessarily requiring such separation, and the described program components and systems are generally integrated together into a single software product or may be packaged into multiple software products. It should be understood that there is.

So far, the present invention has been looked at around the embodiments. Those of ordinary skill in the art to which the present invention pertains will be able to understand that the present invention may be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from a descriptive point of view rather than a limiting point of view. The scope of the present invention is shown in the claims rather than the above description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

Claims (11)

A preprocessor for generating an original image by performing a preprocessing process on the input image;
A transform unit for generating a transformed image using a modified cycle-generating hostile neural network for the original image; And
An age estimating unit for estimating an age for the transformed image;
Age estimation device comprising a.
The method of claim 1,
When the age difference between the age of the original image and the transformed image is greater than or equal to a threshold value, the age of the original image is output as the age of the input image, and the age difference between the age of the original image and the transformed image is less than the threshold value. In case, the selection unit for outputting the age of the converted image as the age of the input image;
But further include,
The age estimating unit further estimates the age of the original image.
The method of claim 1,
The modified cycle-generating hostile neural network includes a generation unit and a discrimination unit,
The generation unit is learned including a loss function including a loss according to a distance between an image of one database and a feature vector that is an additional output of the generation unit for an image of another database,
Wherein the determination unit is trained using a loss function in consideration of an additional loss obtained by applying an increase function to an additional output of the determination unit for the original image and the transformed image.
The method of claim 3,
The generator is a neural network including a layer outputting the transformed image and a layer outputting the additional output,
Wherein the determination unit is a neural network including a layer outputting an original output and a layer outputting the additional output.
The method of claim 1,
The pretreatment unit
Extracting a face area from the input image,
Calculate an in-plane rotation based on the center position of both eyes in the face area,
Rotate the face region according to the inplay rotation angle,
And generating the original image by normalizing the size of the face region.
In the method of estimating the age by the age estimation device,
Generating an original image by performing a preprocessing process on the input image;
Generating a transformed image from the original image using a modified cycle-generated hostile neural network; And
Estimating an age for the transformed image;
Age estimation method comprising a.
The method of claim 6,
Additionally estimating an age for the original image; And
When the age difference between the age of the original image and the transformed image is greater than or equal to a threshold value, the age of the original image is output as the age of the input image, and the age difference between the age of the original image and the transformed image is less than the threshold value. In case, outputting the age of the converted image as the age of the input image;
Age estimation method further comprising a.
The method of claim 6,
The modified cycle-generating hostile neural network includes a generation unit and a discrimination unit,
The generation unit is learned including a loss function including a loss according to a distance between an image of one database and a feature vector that is an additional output of the generation unit for an image of another database,
Wherein the determination unit is trained using a loss function taking into account an additional loss obtained by applying an increase function to an additional output of the determination unit for the original image and the transformed image.
The method of claim 8,
The generator is a neural network including a layer outputting the transformed image and a layer outputting the additional output,
Wherein the determination unit is a neural network including a layer outputting an original output and a layer outputting the additional output.
The method of claim 6,
The step of generating an original image by performing a preprocessing process on the input image,
Extracting a face area from the input image;
Calculating an in-plane rotation based on the center positions of both eyes in the face area;
Rotating the face region according to the inplay rotation angle; And
Generating the original image by normalizing the size of the face area;
Age estimation method comprising a
A computer program recorded in a computer-readable recording medium after executing the method for estimating the age of any one of claims 6 to 10.

Images