KR20190098858A - Method and apparatus for pose-invariant face recognition based on deep learning - Google Patents

Abstract

The present invention relates to a face recognition method using a computer. According to an embodiment of the present invention, the face recognition method comprises the following steps of: (a) detecting a face area from an original image using a first deep learning algorithm and calculating detection area information including coordinates of the detected face area; (b) calculating pose information about a face pose in the detected face area using a second deep learning algorithm; (c) calculating a pose-normalized face image from the original image based on at least part of the pose information and the detection area information; and (d) recognizing a face by comparing the pose-normalized face image with a face image previously stored in a database using a third deep learning algorithm.

Description

Face recognition method and apparatus robust to deep learning based pose change {Method and apparatus for pose-invariant face recognition based on deep learning}

The present invention relates to a technology for recognizing a face of a person in an image, and more particularly, a deep learning-based face recognition method and apparatus that are robust to changes in conditions such as lighting, facial expressions, and face pose changes in a real environment without constraints. It is about.

Recently, face detection or face recognition technology is frequently dealt with in computer vision problems. Conventional face recognition methods have face recognition from images taken under defined environmental conditions (lighting, facial expressions, poses, etc.). Recently proposed methods extract faces from faces taken without limitations of environmental conditions such as real environments. Is designed to recognize.

However, in order to detect and recognize faces without the limitations of environmental conditions, some difficult problems must be overcome. For example, complex backgrounds can be so severe that it makes it difficult to detect discernible features on faces and non-faces. In addition, since the face existing in the captured image may have various sizes, the search space for face detection is very large, and the face poses vary according to the viewpoint change.

One way to solve this problem is to apply a deep learning technique to output good face recognition results for lighting changes and pose changes close to the front image.However, except for images or poses with extremely changed poses Regardless of changes that may occur in the environment (lighting, background, face size), there is still a lack of technology for face recognition with high accuracy, and there is a need for such technology.

Patent Document 1: Korean Unexamined Patent Publication No. 2015-0065445 (June 15, 2015) Patent Document 2: Korean Unexamined Patent Publication No. 2017-0053069 (published May 15, 2017)

According to an embodiment of the present invention, an object of the present invention is to provide a face recognition method capable of identifying a face with high accuracy regardless of a pose change with respect to a face existing in an image.

According to an embodiment of the present invention, as a face recognition method using a computer, (a) detecting a face region in an original image using a first deep learning algorithm and calculating detection region information including coordinates of the detected face region Doing; (b) calculating pose information about a face pose in the detected face region by using a second deep learning algorithm; (c) calculating a pose normalized face image from the original image based on at least some of the pose information and the detection region information; And (d) recognizing a face by comparing the pose normalized face image with a face image pre-stored in a database using a third deep learning algorithm. .

According to an embodiment of the present invention, a computer-readable recording medium having a program recorded thereon for executing the face recognition method on a computer may be provided.

According to an embodiment of the present invention, a deep learning-based face that is robust to pose changes capable of outputting face identification results with high accuracy despite changing conditions such as lighting, facial expressions, and facial pose changes in a real environment without constraints. A recognition method and apparatus can be provided.

1 is a block diagram of a face recognition system for performing a face recognition method according to an embodiment of the present invention;
2 is an exemplary flowchart of a face recognition method according to an exemplary embodiment;
3 is a view for explaining the rotation of the face,
4 is a diagram for explaining a preprocessing step according to one embodiment;
5 is a diagram for describing a pose estimation step, according to an exemplary embodiment;
6 is a diagram for describing an exemplary method of setting yaw difficulty according to an embodiment;
7 is a diagram for describing data used for learning a pose estimation algorithm, according to an exemplary embodiment;
8 is a diagram for describing a pose normalization step, according to an exemplary embodiment;
9 is a view for explaining a face recognition step according to one embodiment;
10 is a block diagram illustrating a face recognition apparatus according to an exemplary embodiment.

Objects, other objects, features and advantages of the present invention will be readily understood through the following preferred embodiments associated with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided so that the disclosure may be made thorough and complete, and to fully convey the spirit of the present invention to those skilled in the art.

Where the terms first, second, etc. are used herein to describe the components, these components should not be limited by these terms. These terms are only used to distinguish one component from another. The embodiments described and illustrated herein also include complementary embodiments thereof.

In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, the words 'comprise' and / or 'comprising' do not exclude the presence or addition of one or more other components.

In addition, the terms "... unit", "... module", "... block" and the like used to refer to the components of the present invention may mean a unit for processing at least one function or operation, which It can be implemented in hardware or software, or a combination of hardware and software.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings. In describing the specific embodiments below, various specific details are set forth in order to explain and understand the invention in more detail. However, those skilled in the art can understand that the present invention can be used without these various specific details. In some cases, it is mentioned in advance that parts of the invention which are commonly known in the description and which are not related to the invention are not described in order to prevent confusion in describing the invention.

1 is a block diagram of a face recognition system for performing a face recognition method according to an embodiment of the present invention, and FIG. 2 is an exemplary flowchart of a face recognition method according to an embodiment.

Referring to the drawings, a face recognition system according to an embodiment includes a preprocessor 100, a pose estimator 200, a pose normalizer 300, and a face recognizer 400, and one or more according to need. Database 530 may be included.

According to the face recognition system, first, a preprocessing step by the preprocessor 100 is performed in step S10. The preprocessor 100 detects a face region in the original image and calculates detection region information including coordinates of the detected face region. In one embodiment, the preprocessor 100 may detect a face region in the original image using a deep learning algorithm. The deep learning algorithm is a kind of machine learning technique, and refers to an artificial neural network algorithm including a plurality of hidden layers. In order to use the deep learning algorithm, the deep learning algorithm model must be preceded by a learning step using images including various face poses, and the following description assumes that the deep learning algorithm mentioned herein has completed this learning step. .

In an exemplary embodiment, the preprocessor 100 may detect a face region in the original image and calculate detection region information including coordinates of the detected face region. Here, the 'detection area information' may include, for example, at least some of x, y, width, and height information of the face area in the original image. In this case, the x coordinate and the y coordinate may be, for example, coordinates of the center point of the face region, or alternatively, may refer to coordinates of the upper left or lower left of the face region.

In one embodiment, the preprocessing unit 100 may normalize the face area after detecting the face area in the original image. For example, the preprocessing unit 100 may normalize the detected face region to an image having a predetermined size, and may also normalize the color distribution.

The image preprocessed by the preprocessor 100 is input to the pose estimator 200, and in operation S20, an operation of calculating pose information by the pose estimator 200 is performed. The pose estimator 200 calculates pose information regarding a face pose in the face region of the preprocessed image. In an embodiment, the pose estimator 200 may calculate the face pose information by a deep learning algorithm. In one embodiment, the pose information calculated by the pose estimator 200 may include information about a face rotation, a yaw rotation direction, and a difficulty level of the yaw. 'Yo' is used in the same sense as indicating the direction of rotation of any object in three dimensions. That is, as shown in Fig. 3, 'yaw' is a rotation about the Y axis (up and down direction of the face) (ie, turning the face left or right).

In an exemplary embodiment, the pose estimator 200 calculates information about the 'rotation of the face' as an angle. For example, if the face on the image is upright and not tilted, the face's angle of rotation is zero, and if the face is rotated clockwise or counterclockwise, it can range from +90 degrees to -90 degrees depending on the rotated angle. Can have

According to an exemplary embodiment, the pose estimator 200 may classify and calculate the information about the yaw into three types, front, left, and right. In other words, if the face faces the front, it is calculated as 'front', and if it faces the left or the right, it is calculated as 'left' or 'right'. In this embodiment, the corresponding number may be calculated by mapping the left, front, and right sides to 0, 1, and 2, respectively. In an alternative embodiment, the pose estimator 200 may calculate the information about the yaw in an angle. For example, if the front is displayed at 0 degrees and rotated to the left or right, it can be displayed as a value between +90 degrees and -90 degrees depending on the rotated angle.

The degree of difficulty is classified as the degree of difficulty according to the degree of rotation of the yaw direction of the face. For example, in one embodiment, the degree of difficulty may be divided into upper, middle, and lower points according to the degree of yaw rotation. In an alternative embodiment, the pose estimator 200 may further include information about a rotation about the pitch direction of the face, that is, the X axis (left and right directions of the face) of FIG. 3.

Next, the pose normalization step S30 is executed by the pose normalization unit 300. The pose normalization unit 300 calculates a pose normalized face image from the original image based on the detection region information calculated by the preprocessor 100 and the pose information calculated by the pose estimator 200. Here, the pose normalized face image is a normalized face region for easy face recognition. The pose normalization unit 300 may perform pose normalization by, for example, standing the face vertically so that the rotation angle of the face is 0 degrees, and inverting the image so that the face faces the front or the left. At this time, for pose normalization, face rotation or inversion may be performed using face rotation and yaw information calculated by the pose estimator 200.

In addition, the pose normalization unit 300 may extract the face region in the original image by using the detection region information calculated by the preprocessor 100, and may also normalize the color distribution if necessary.

The pose normalization unit 300 transmits the pose normalized image to the face recognition unit 400, and in step S40, the face recognition unit 400 compares the pose normalized image with the face image stored in the database 530. To recognize (identify) the face. In one embodiment, the face recognition unit 400 may perform a face recognition operation by a deep learning algorithm.

In an embodiment, the face recognition unit 400 may use yaw difficulty information calculated by the pose estimator 200. For example, when using a deep learning algorithm composed of a plurality of layers (layers), the face recognition unit 400 includes at least one layer having a plurality of sublayers having different weights according to the degree of difficulty. In this case, a feature vector representing a face feature is configured to be delivered to one of the plurality of sublayers according to the degree of difficulty, and the feature vector processed in this sublayer is a feature value of a predetermined vector space regardless of the degree of difficulty. It is mapped to Therefore, even though a face detected in the original image has various poses, a feature value that is robust to a pose change can be extracted while passing through a layer composed of the plurality of sublayers, and the extracted feature value is stored in the database 530. Compare faces to identify faces.

Meanwhile, the above-described face recognition system may include a plurality of differently learned deep learning algorithms. For example, the preprocessor 100 includes a first deep learning algorithm for detecting a face region in the original image, and the pose estimator 200 executes a second deep learning algorithm for calculating pose information in the preprocessed image. The facial recognition unit 400 may include a third deep learning algorithm for extracting a face from a pose normalized image and comparing the image with an image in a database.

In one embodiment, some of these deep learning algorithms may be implemented with a convolution neural network (CNN) algorithm. In one embodiment, the convolutional neural network algorithm may include one or more convolutional layers, one or more active functions, and one or more pooling layers, where each convolutional layer may consist of one or more convolutional filters. The structure of such a convolutional neural network may vary according to specific embodiments of the present invention. Since the convolutional layer, the active function, and the pooling layer constituting the neural network are well known in the art, detailed description thereof will be omitted.

In one embodiment, the first to third deep learning algorithm is composed of different layers because the target and the output value is different for each deep learning algorithm. For example, the first deep learning algorithm of the preprocessor 100 is an algorithm for detecting a face region in the original image, and may include, for example, an image pyramid or cascade method in order to minimize a search space. have.

Since the second deep learning algorithm of the pose estimator 200 needs to calculate a plurality of pose information such as the rotation of the face, the rotation direction of the yaw, and the difficulty of the yaw, the multi-task (the rotation of the face) based on a single deep learning algorithm. , The rotation direction of the yaw, and the difficulty of the yaw) can be configured to calculate each pose information. At this time, the rotation of the face in the pose information is a matter of regression, and the direction of yaw rotation and the difficulty of the yaw is a matter of classification. It is a good idea to configure the task so that it can be performed.

The third deep learning algorithm of the face recognizer 400 extracts a feature vector of the face from the pose normalized image and compares the extracted feature vector with face images previously stored in the database 530 to identify the face. In one embodiment, a feature extraction neural network extracting a feature vector of a face and a machine learning algorithm (eg, artificial neural network, cosine similarity), a support vector machine (SVM) capable of measuring face similarity Machine), nearest neighbor rules (Nearest Neighbors, etc.). The feature extraction neural network may include, for example, one or more convolutional layers and one or more pooling layers, and the convolutional layer is composed of one or more convolutional filters and is constructed using various active functions (functions of ReLU, PReLU, Tanh, Sigmoid, etc.). can do.

Each operation of the face recognition system will now be described in detail with reference to FIGS. 4 to 9.

4 illustrates a pretreatment step S10 by the preprocessor 100 according to an embodiment. Referring to the drawing, the preprocessor 100 detects a face region from an original image and normalizes the detected image. In an embodiment, the preprocessor 100 may detect a face region in the original image using the first deep learning algorithm. The first deep learning algorithm may include a convolutional neural network including at least one convolutional layer, at least one active function, and at least one pooling layer, and a neural network for identifying and detecting a face region and extracting detection region information.

In one embodiment, the detection area information may include at least some of x coordinate, y coordinate, width, and height information of the face area in the original image. Here, the x coordinate and the y coordinate may be, for example, the coordinates of the center point of the face area, and in an alternative embodiment, may refer to the coordinates of any reference point, such as the upper left coordinate or the lower left coordinate.

On the other hand, since the background screen of the original image photographed in the real environment is different and the face size in the original image is different, it is desirable to solve the problem of rapidly increasing the number of search spaces. To this end, in one embodiment, the search space may be reduced by using an image pyramid, a cascade, a region recommendation network, or the like.

In addition, since the size and color of the face region detected in the original image are different, the size and color distribution of the detected face region can be normalized and processed to have a constant size and color distribution.

5 illustrates a pose estimation step S20 by the pose estimator 200, according to an exemplary embodiment. Referring to the drawing, in step S20, the pose estimator 200 calculates pose information regarding a face posture (pose) in the face region of the preprocessed image. In an embodiment, the pose estimator 200 may calculate pose information by using a second deep learning algorithm.

In one embodiment, the pose information includes information about a rotation angle of a face, a rotation direction of a yaw, and a difficulty level of the yaw. The rotation angle of the face may have a value between +90 degrees and -90 degrees based on the face not being tilted upright in the image. The yaw direction may have a value indicating one of the left side, the front side, and the right side (for example, 0, 1, and 2, respectively) with respect to the front side.

The degree of difficulty is classified as the degree of difficulty according to the degree of rotation of the yaw direction of the face. For example, in one embodiment, the degree of difficulty may be divided into upper, middle, and lower points according to the degree of yaw rotation. In this regard, Figure 6 shows an exemplary criterion for classifying the degree of difficulty according to the degree of facial rotation. In FIG. 6, a plurality of photographs are listed, and photographs of which yaw directions are gradually changed from a photograph facing the front to a scene facing 90 degrees to the left. In one embodiment, the degree of difficulty may be classified into upper, middle, and lower. In this case, the difficulty of the 'low' (Easy) is a case in which both eyes are completely visible as shown in the upper three faces of FIG. 6. The middle difficulty (middle) is a case where one eye is completely visible and the other eye is partially visible, as in the middle two faces of FIG. 6, and the difficulty level 'hard' is a lower two face of FIG. 6. As shown in this example, only one eye is visible and the other eye is not visible at all.

As described above, according to an exemplary embodiment, the second deep learning algorithm of the pose estimator 200 may classify the difficulty of the image into upper, middle, and lower regions by using a feature vector representing a feature of a face in the preprocessed image. have.

On the other hand, when designing the second deep learning algorithm of the pose estimator 200, it is important to construct training data for effectively performing the pose estimation according to the present invention. For example, if there are not many face pictures according to the direction of rotation of the face with respect to the same person, that is, if there are only faces in a specific yaw direction (viewing direction to the left of 90 degrees, for example) as shown in the center picture in FIG. A plurality of learning data having different rotations of the face may be generated by randomly rotating clockwise and counterclockwise based on the face to be extracted and extracting a face region. The learning data generated in this way may be set to have the same rotation direction and difficulty of the yaw as the original picture (the center picture of FIG. 7).

8 illustrates a pose normalization step S30 by the pose normalization unit 300 according to an exemplary embodiment. Referring to the drawing, the pose normalization unit 300 calculates a pose normalized face image from the original image based on the rotation angle of the face, the rotation direction of the yaw, and the detection region information. In an embodiment of the pose normalization step, the original image is rotated so that the rotation angle of the face of the face region of the original image is 0 degrees using the rotation angle of the face calculated by the pose estimator 200.

Thereafter, the face region is extracted from the original image based on the detection region information calculated by the preprocessor 100, and the image is inverted so that the face faces the front or the left based on the yaw direction calculated by the pose estimator 200. Let's do it. For example, the image may be inverted so that the face looking to the front is not inverted, and the face looking to the right is inverted to the left. The reason why the front and one side face image is used as an input image is to minimize the change according to the direction of the face since the left-right shape is very similar to the nose of the face.

Then, the pose normalization converts the rotated and / or inverted face area into an image of a preset size and normalizes color distribution.

Meanwhile, in the above description, the face region is extracted from the original image based on the detection region information, and then the inversion according to the yaw direction is performed. However, in an alternative embodiment, the image inversion according to the yaw direction is first performed and then detected. The face region may be extracted based on the region information. Also in alternative embodiments, the image may be reversed as needed such that the face faces front or right when inverted based on the yaw direction.

9 illustrates a face recognition step S40 by the face recognition unit 400 according to an exemplary embodiment. Referring to the drawing, in step S40, the face recognition unit 400 extracts a feature vector of a face from a pose normalized image and compares the extracted feature vector with face images previously stored in the database 530. Can be identified.

In one embodiment, the face recognition unit 400 uses a third deep learning algorithm. The third deep learning algorithm may include a deep learning algorithm 410 for feature extraction, a fully connected (FC) layer 420, and an algorithm 430 for face recognition.

The deep learning algorithm for feature extraction may be an algorithm for extracting a feature vector of a face from a pose normalized image. In one embodiment, the deep learning algorithm for feature extraction may include one or more convolutional layers, one or more active functions, and one or more pooling layers. Here, the feature vector is a set of feature values of the face. Features of the face may include different face features such as, for example, the position of the pupil, the distance between the pupils, the shape of the eyes, the position and / or height of the nose, the position of the lips, the length of the lips, the position of the masses, the position of the ears, the distance between the ears It may be selected from the components that distinguish it from the face.

The FC layer 420 is a layer (layer) for adding different weights according to the degree of difficulty to the feature vector extracted by the feature extraction deep learning algorithm 410. In the illustrated embodiment, the FC layer 420 may include the first FC layer 421, the second FC layer 422, and the first FC layer 421 that respectively add different first to third weights according to the lower, middle, and upper levels of the difficulty. And a third FC layer 423.

According to this configuration, for example, when the degree of difficulty of the pose normalized image is 'low', the feature vector extracted by the feature extraction algorithm 410 is input to the first FC layer 421. As another example, the feature vector is input to the second FC layer 422 when the yaw difficulty of the pose normalized image is 'medium', and the feature vector is input to the third FC layer 423 when the yaw difficulty is 'up'. Will understand the input.

Each of the first to third FC layers 421, 422, 423 is configured to add different weights to the feature vector. In a preferred embodiment, for at least one feature component of the feature vector extracted by the feature extraction algorithm 410, if the feature component is processed in the first to third FC layers 421, 422, 423, respectively, it is irrelevant to the difficulty level. The first to third weights are respectively set so as to map to the same feature value of the predetermined vector space.

As described above, the third deep learning algorithm of the face recognition unit 400 according to the embodiment of the present invention includes an FC layer 420 that adds different weights to the feature vector according to the degree of difficulty, and the FC layer 420 ) Is configured to map at least some feature components of the feature vector to the same feature values in a predetermined vector space irrespective of the difficulty, for a plurality of pose normalized faces having different yaw directions from the same person. Even with these various poses, it is possible to extract feature values robust to the pose change while passing through the FC layer 420.

The face recognition algorithm 430 is an algorithm that recognizes (identifies) the face of the original image by comparing the feature vector output from the FC layer 420 with data stored in the database 530. The classification neural network for face recognition is performed. It may be configured as. In one embodiment, the deep learning algorithm 430 for face recognition may be implemented as an input layer, two or more hidden layers, and an output layer in the same or similar manner to a general artificial neural network algorithm.

In one embodiment, the face recognition algorithm 430 is based on machine learning, such as artificial neural networks, cosine similarity, support vector machines (SVMs), nearest neighbors (Nearest Neighbors), and the like. Similarity measurement methods may be used to perform face recognition that is robust to pose changes.

10 is a block diagram illustrating an exemplary system configuration for performing a face recognition method according to an exemplary embodiment. Referring to the drawings, the facial recognition system 500 according to an exemplary embodiment may be any terminal device or a server capable of executing the facial recognition method described with reference to FIGS. 1 to 9, and the processor 510 as illustrated. , Memory 520, and database 530, and storage 540.

The database 530 may store data regarding face images stored in advance for face recognition. In an embodiment, the database 530 may store information about a feature vector of each face image.

The storage device 540 is a storage medium capable of semi-permanently storing data, such as a hard disk drive or a flash memory, and may store one or more of various algorithms and data such as the first to third deep learning algorithms described above. For example, the storage device 540 may include a preprocessing algorithm 541 for the preprocessor 100 to perform a preprocessing operation, a pose estimation algorithm 542 for the pose estimator 200 to perform a pose estimation operation, and pose normalization. The unit 300 stores a pose normalization algorithm 543 for performing a pose normalization operation, a face recognition algorithm 544 for the face recognition unit 400 to perform a face recognition operation, and various other software or programs. Can be. In this case, the preprocessing algorithm 541 may include the first deep learning algorithm described above, the pose estimation algorithm 542 may include the second deep learning algorithm described above, and the face recognition algorithm 544 may include a feature extraction deep. It will be appreciated that the algorithm may include a running algorithm 410, an FC layer 420, and a face recognition algorithm 430.

In this configuration, such various programs or algorithms may be stored in the storage device 540 and loaded and executed in the memory 520 under the control of the processor 510. Alternatively, some programs or algorithms may reside on an external device or server that is separate from face recognition system 500, and when the system 500 transmits data or variables to that external device or server, the external device or The server may execute a program or algorithm and then pass the data to system 500.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited to the above embodiments. Those skilled in the art will understand that various modifications and variations are possible from the above description. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

100: pretreatment unit
200: pose estimator
300: pose normalization unit
400: face recognition unit

Claims (11)

As a face recognition method using a computer,
(a) detecting a face region in the original image by using a first deep learning algorithm and calculating detection region information including coordinates of the detected face region;
(b) calculating pose information about a face pose in the detected face region by using a second deep learning algorithm;
(c) calculating a pose normalized face image from the original image based on at least some of the pose information and the detection region information; And
and (d) recognizing a face by comparing the pose normalized face image with a face image pre-stored in a database by using a third deep learning algorithm. The method of claim 1,
And the pose information includes information on a rotation angle of the face of the detected face pose, a yaw direction, and a yaw difficulty according to the yaw direction. The method of claim 2,
The third deep learning algorithm includes a layer that adds different weights to the feature vector of the pose normalized face image according to the degree of difficulty, thereby providing a plurality of pose normalized faces having different yaw directions of the same person. And mapping at least some feature components of the feature vector to the same feature value in a predetermined vector space irrespective of the degree of difficulty. The method of claim 2,
The rotation angle of the face has a value in the range of -90 degrees to +90 degrees,
The yaw direction has a value indicating one of a left side, a front side, and a right side;
The difficulty level of the detected face pose (i) when both eyes are visible, (ii) one eye and partially visible, and (iii) only one eye, respectively, Face recognition method characterized in that it is set to represent the image. The method of claim 2,
(C) calculating the pose normalized face image comprises calculating the pose normalized face image based on the rotation angle and yaw direction of the face and the detection area information. The method of claim 3, wherein the third deep learning algorithm,
A feature extraction algorithm for extracting a feature vector from the pose normalized face image;
A fully connected (FC) layer 420 for adding different weights to the extracted feature vectors according to the degree of difficulty; And
And a face recognition algorithm for recognizing a face by comparing the feature vector output from the FC layer (420) with a face image previously stored in a database. The method of claim 6,
The feature extraction algorithm (410) comprises one or more convolutional layers, one or more active functions, and one or more pooling layers. The first FC layer 421 and the second FC layer of claim 6, wherein the FC layer 420 adds first to third weights different from each other according to the lower, middle, and upper levels of the difficulty level. 422), and a third FC layer (423). The method of claim 8,
The feature vector extracted by the feature extraction algorithm 410 is input to the first FC layer 421 when the yaw difficulty is lower, and the feature vector is input to the second FC layer 422 when the difficulty is difficult. If the difficulty level is different, the feature vector is configured to be input to the third FC layer (423). The method of claim 9,
For at least one feature component of the feature vector extracted by the feature extraction algorithm, if the feature component is processed in each of the first to third FC layers 421, 422, 423, it is mapped to the same feature value in a predetermined vector space regardless of the difficulty level. The first to the third weight is set so that the face recognition method, respectively. A computer-readable recording medium having recorded thereon a program for executing the face recognition method according to any one of claims 1 to 10 on a computer.

Images