KR100577745B1 - Facial authentication method - Google Patents

Abstract

The present invention relates to a face authentication method, comprising: a first step of extracting an image including a front face image by photographing a specific face; A second step of detecting a normalized face image by finding two eyes and mouth regions of the face in the image including the extracted face image; The inclination angle of the face is found from the vector direction connecting the two eye regions of the normalized face image, and the straight face image is obtained by rotating the face image through an affine transform. Performing a third step; The facial image obtained through the third step is cut out into a quadrangle formed by connecting points spaced apart from the eye region and the mouth region by a predetermined distance, and matched to an aligned image size of a pre-built database. A fourth step of acquiring an image; A fifth step of determining and authenticating the similarity between the sorted face images of the pre-established database and the acquired specific face image is performed.

Therefore, the present invention has the effect of correcting the pose change of the individual face and performing accurate authentication.

Face, authentication, similarity, eyes, mouth

Description

How to authenticate your face {Method for authenticating human face}             

1 is a flowchart of a face authentication method according to the present invention

2 is a schematic perspective view of an active sliding camera applied to the present invention;

Figure 3 is a schematic perspective view of a pan camera applied to the present invention

4 is a conceptual diagram illustrating a method of detecting a normalized face image in the face authentication method according to the present invention.

FIG. 5 is a view showing candidate regions capable of binocular and mouth regions in an image including a face according to the present invention.

FIG. 6 is a diagram for describing determining two eye regions and a mouth region in an image including a face according to the present invention. FIG.

7 is a view for explaining the inclination angle of the face to the two eye areas in accordance with the present invention

8 is a view for explaining a method of acquiring a specific face image aligned according to the present invention.

As illustrated in FIG. 9, a diagram for describing determining a similarity degree between each of the face images arranged in the database in advance and the acquired face image.

<Description of Symbols for Main Parts of Drawings>

100: Pan camera 110: Slide axis

120: transfer device 201,202: eye area

210: mouth area 250: square shape

300: video

501,511,512,513,514,515: face area

The present invention relates to a face authentication method, and more particularly, to a face authentication method capable of correcting a pose change of an individual face and performing accurate authentication.

Recently, security products using biometrics have become a bot.

This type of biometrics is a method of extracting a database of unique features such as face, fingerprint, iris, retina, voice, and vein to make a database, and verifying the identity of the user when compared to the established database.

In order to extract a feature from a user, fingerprints are acquired by a contact method, and in the case of an iris, it is troublesome to acquire an image of an eye in close proximity to a camera, and in the case of a voice, a user must wear a microphone.

However, in the case of face recognition, when acquiring the face information of the user, the feature can be acquired without the user's consciousness.

Face recognition has become an important research field in the future as an essential function for interactive services between users and systems in intelligent surveillance systems or mobile robots, and many methods for increasing recognition rates have been proposed.

However, in the case of face recognition, the recognition rate is sensitively reacted according to the surrounding environment, and the main factors that degrade the performance of face recognition include a change in lighting environment and a change in face pose.

On the other hand, the face authentication method has been recently developed in various ways, the first method is to build a database by inputting the face image of the front pose without tilting, and then acquired in the authentication process and the face image built in the database The face images were compared and authenticated, but the face images obtained in the authentication process should be aligned in the front without tilting to have the best authentication rate.

In such a face authentication method, in order to obtain a face image without tilting from the front face, the user has to pose a front face toward the camera, and it is difficult to realize a perfect front face. .

The second method assumes that the face has an ellipse shape to correct the two-dimensional face pose change, and estimates the inclination angle of the face by analyzing the principal axis of the image pixels forming the face. The image was corrected by rotating (Affine transform).

In the face alignment process of the personal face image through the rotation of the principal face and the affine transform, it is difficult to accurately extract the face region.

In the third method, facial features are detected for fine face alignment, and face images are corrected through a two-dimensional alignment process.

However, this method consists of eye and mouth templates to extract facial features such as eyes, nose and mouth, but it is not easy to extract facial features because eyes and mouth are easily changed. have.

Finally, we designed a 3D face template for 3D face image correction, and based on this, the face was transformed and aligned through texture mapping of the 2D face image.

In the case of such a 3D face template, various face deformations are possible, but texture mapping is very long, and thus real-time implementation is difficult. When face deformation is large, an error occurs in the actual face authentication process. have.

Accordingly, an object of the present invention is to provide a face authentication method capable of correcting a pose change of an individual face and performing accurate authentication.

According to a preferred aspect of the present invention, there is provided a method including: a first step of capturing a specific face to extract an image including a front face image;

A second step of detecting a normalized face image by finding two eyes and mouth regions of the face in the image including the extracted face image;

The inclination angle of the face is found from the vector direction connecting the two eye regions of the normalized face image, and the straight face image is obtained by rotating the face image through an affine transform. Performing a third step;

The facial image obtained through the third step is cut out into a quadrangle formed by connecting points spaced apart from the eye region and the mouth region by a predetermined distance, and matched to an aligned image size of a pre-built database. A fourth step of acquiring an image;

There is provided a face authentication method comprising a fifth step of determining and authenticating a similarity between the sorted face images of the pre-established database and the acquired specific face image.

Hereinafter, exemplary embodiments of the present invention will be described with reference to the accompanying drawings.

1 is a flowchart of a face authentication method according to an embodiment of the present invention, in which an image including a front face image is extracted by photographing a specific face (step S10).

The step S10 may include: performing a pan operation to rotate left and right at an angle with respect to one axis, and preparing an active sliding camera capable of linear movement; After performing the linear motion and the pan operation, it is preferable to perform a step of extracting an image including a front face image by photographing a specific face with the active sliding camera.

In operation S20, normalized face images are detected by finding two eyes and mouth regions of the face in the extracted image including the face image.

Here, in order to perform face authentication, a normalized face image having a certain size in a certain face area of a specific person must be detected in the image. In order to detect the normalized face image, the positions of eyes and mouth must be regularly normalized. .

For this normalization process, instead of the template-based method, the eyes and mouth are searched for by pixels with darker brightness than the surrounding pixels.

Subsequently, the inclination angle of the face is found from the vector direction connecting the two eye regions, and the face image is rotated by an affine transform to obtain the straight face image. (Step S30)

As shown in FIG. 7, the inclination angle 'α' of the face can be known from the vector direction connecting the two eye regions 201 and 202.

Here, the affine transform is a conventional method of compensating a tilted face image, and a description thereof will be omitted.

Subsequently, the face image acquired through the step S30 is cut into a quadrangle formed by connecting the points separated by a predetermined distance from the two eye regions and the mouth region, and aligned to match the aligned image size of the pre-established database. Acquire a specific face image (step S40).

That is, as shown in FIG. 8, in the image 300, the face image is cut out into a quadrangular shape 250 formed by connecting points spaced a predetermined distance from the two eye regions 201 and 202 and the mouth region 210.

Here, it is preferable to cut out the face image in the shape of a rectangle formed by connecting points spaced apart from the center of the two eye regions 201 and 202 and the mouth region 210 by a predetermined distance.

Thereafter, the similarity between the face images arranged in the pre-built database and the acquired specific face image is determined and authenticated (step S50).

At this time, the judgment of the similarity uses following formula (1).

-----(1)

-----(One)

는 추출된 영상 벡터의 표준편차이고,

는 데이터 베이스 영상 벡터의 표준편차이다.Where E is the expected value, Ifd is the extracted image vector, Itmpl is the database image vector,

Is the standard deviation of the extracted image vector,

Is the standard deviation of the database image vector.

That is, as shown in FIG. 9, when the similarity between each of the face images arranged in the database in advance and the acquired specific face image is calculated using Equation (1), the highest similarity value is obtained. You will be authenticated.

In FIG. 9, among the face images 511, 512, 513, 514, and 515, the acquired specific face image 501 is most similar to the face image of 514, and thus the authentication is performed.

FIG. 2 is a schematic perspective view of an active sliding camera applied to the present invention. The active sliding camera fixes a pan camera 100 and the pan camera 100 capable of capturing a pan operation and an image, and the fan camera. It consists of a transfer device 120 that can transfer the 100 in a straight line.

At this time, the transfer device 120 is preferably a transfer device capable of performing a left, right linear movement using the sliding shaft (110).

3 is a schematic perspective view of a fan camera according to the present invention, and includes a body 2 on which a fan motor 1 is mounted, and one end of a fan shaft 4 driven by a timing belt connected to the fan motor 1. A fan case 5 rotatably supporting the fan; A case 6 fixedly supported at the other end of the fan shaft 4; The bracket 10 is fixed to the case and the camera 8 is installed.

4 is a conceptual view illustrating a method of detecting a normalized face image in the face authentication method according to the present invention. In step S30 of the face authentication method of FIG. 1, two eyes of a face in an image including the extracted face image are illustrated. The normalization of the facial image by finding the intrusion region will be described in detail with reference to FIG. 4.

First, for all the pixels in the image, the average brightness values of the pixels in the image in the vertical, horizontal, positive diagonal and negative diagonal directions about each pixel are calculated and centered over the brightness average values. When the brightness value of the pixel becomes dark, a weight is assigned to each direction.

For example, as shown in FIG. 4, the brightness value of one pixel A of the image 300 is the brightness average value of the pixels in the vertical direction (a) and the brightness of the positive diagonal (c) pixels around the pixel A. When it becomes darker than the average value and becomes brighter than the brightness average value of the pixels in the horizontal direction b and the brightness average value of the negative diagonal direction d pixels, the weight A is assigned to the pixel A.

In addition, when the brightness value of the pixel becomes darker than the brightness average value of the pixels in all directions, the weight is assigned to the pixel.

When weighting is applied to all pixels of the image, if the central pixel brightness value is dark in all directions, the weight becomes +4, and since the eye has the darkest characteristic relatively in the image, the weight corresponding to +4 Assign a set of pixels to candidate areas of eyes and mouth.

Here, in order to be more precise, since the mouth also has a relatively dark characteristic but slightly lighter than the eye, the set of pixels corresponding to the weight +2 to +4 is assigned as the candidate region of the mouth.

As described above, all assigned candidate regions 200 are displayed in the image as shown in FIG. 5, and the regions having potential for both eyes and mouth are found in these candidate regions.

At this time, the present invention finds a pair which satisfies the characteristics of the following formulas (a) and (b) among the possible areas of the eyes and the mouth and assigns them to the areas of the eyes.

Dmin <distance between two eye candidate regions <Dmax ---- (a)

Rmin <area ratio of two eye candidate regions <Rmax ---- (b)

Here, Dmin and Dmax are the minimum and maximum values of the distance D between the two eye regions 201 and 202 according to the face area, and Rmin and Rmax are the area ratios of the two eye regions 201 and 202, as shown in FIG. The minimum and maximum values of.

In other words, Rmin and Rmax are generally the minimum and maximum values of the area ratio of two human eye areas. In the present invention, Rmin is 0.95 and Rmax is 1.05.

If only one of the above formulas (a) and (b) is satisfied, it can be assigned to the area of the eyes.

In addition, since the area occupied by the face in the captured image may change, the Dmin and Dmax are proportionally reduced or increased with the face area.

As such, when two eye regions satisfying the above-described characteristics, Equations (a) and (b) are found, the width of the region is equal to Eq. If the same characteristic is satisfied, it is determined as the mouth area.

Lmin <width of the mouth candidate area <Lmax ---- (c)

Here, Lmin and Lmax are minimum and maximum values of the width L of the mouth region 210 according to the face area, as shown in FIG. 6.

 As described above, the present invention has the effect of correcting the pose change of the individual face and performing accurate authentication.

Although the invention has been described in detail only with respect to specific examples, it will be apparent to those skilled in the art that various modifications and variations are possible within the spirit of the invention, and such modifications and variations belong to the appended claims.

Claims (6)

A first step of extracting an image including a front face image by photographing a specific face; A second step of detecting a normalized face image by finding two eyes and mouth regions of the face in the image including the extracted face image; The inclination angle of the face is found from the vector direction connecting the two eye regions of the normalized face image, and the straight face image is obtained by rotating the face image through an affine transform. Performing a third step; The facial image obtained through the third step is cut out into a quadrangle formed by connecting points spaced apart from the eye region and the mouth region by a predetermined distance, and matched to an aligned image size of a pre-built database. A fourth step of acquiring an image; And a fifth step of determining and authenticating a similarity between the sorted face images of the pre-established database and the acquired specific face image. The method of claim 1, The fifth step, And calculating the similarity between each of the face images arranged in the database in advance and the acquired face image using Equation (1), and authenticating when the most similarity value is high.

-----(One) Where E is the expected value, Ifd is the extracted image vector, Itmpl is the database image vector,

Is the standard deviation of the extracted image vector,

Is the standard deviation of the database image vector. The method of claim 1, The first step is, Preparing an active sliding camera capable of linear movement and performing a pan operation capable of rotating left and right at a predetermined angle with respect to one axis; And performing a linear motion and a pan operation, and extracting an image including a front face image by photographing a specific face with the active sliding camera. The method of claim 1, The second step, Setting pixels having brightness values darker than surrounding pixels among the pixels of the image extracted in the first step as candidate regions of two eyes and a mouth; Finding a pair of regions satisfying Equation (a) among the candidate regions of the two eyes and the mouth and assigning the pair of regions to the two eyes; Dmin <distance between two eye candidate regions <Dmax ---- (a) Here, Dmin and Dmax are the minimum and maximum values of the distance D between two eye regions according to the face area. And identifying an area whose area width satisfies the following formula (b) among the candidate areas in the vertical direction from the center between the two eye areas and assigning the area to the mouth area. Lmin <width of the mouth candidate area <Lmax ---- (b) Here, Lmin and Lmax are the minimum and maximum values of the width L of the mouth area according to the face area. The method of claim 4, wherein Assigning to the areas of the two eyes, And detecting and allocating a pair of regions satisfying the following equation (c) among the candidate regions of the two eyes and the mouth.  Rmin <area ratio of two eye candidate regions <Rmax ---- (c) Here, Rmin and Rmax are the minimum and maximum values of the area ratio of the two eye regions. The method according to claim 4 or 5, Among the pixels of the extracted image, setting pixels having darker brightness values compared to surrounding pixels as candidate regions of both eyes and mouth may include: Calculating brightness average values of pixels in the image in a vertical direction, a horizontal direction, a positive diagonal direction and a negative diagonal direction with respect to all pixels of the image extracted in the first step; ; When the brightness value of the center pixel becomes darker than the brightness average values, weighting each direction for each pixel; And setting the set of the highest summed weight pixels as candidate regions of both eyes and mouth.

Images