WO2006025129A1

WO2006025129A1 - Personal authentication system

Info

Publication number: WO2006025129A1
Application number: PCT/JP2005/004214
Authority: WO
Inventors: Kiyomi Nakamura; Hironobu Takano
Original assignee: Toyama-Prefecture
Priority date: 2004-08-30
Filing date: 2005-03-10
Publication date: 2006-03-09

Abstract

A personal authentication system capable of identifying and authenticating a person accurately with high precision at a high speed by recognizing the iris image of the person, and applicable to an information system or other management business. The personal authentication system comprises a means for imaging an eye (10) with a dynamically imaging camera (14) and detecting the iris (12) or pupil (27) part, a means for capturing a detected iris pattern, and a means for converting the rotational orientation and shape of the iris pattern by a rotary diffusion neural network and learning/storing as vector information. The personal authentication system further comprises an iris pattern judging means for comparing vector information obtained by converting an arbitrary imaged iris pattern through the storing/converting means by rotary diffusion neural network with the vector information of iris pattern learnt/stored by being converted by the storing/converting means. Personal identification is carried out by recognizing and correcting the orientation of each vector information being compared, determining at least one of the inner product or the minimum distance of each vector information and then judging match/mismatch of each vector information.

Description

Specification

Personal authentication device

Technical field

TECHNICAL FIELD [0001] The present invention relates to a personal authentication device that uses an iris network to acquire an iris pattern of a human eye to identify and authenticate an individual.

Background art

[0002] Conventionally, with the development of computer technology and the Internet, computers are widely used not only in companies but also in general households, and the demand for security is increasing. In addition, in order to increase international crime and to take safety measures at various facilities, research on identifying individuals using physical characteristics such as fingerprints and faces has been actively conducted. In recent years, the iris has attracted attention as information for personal identification (biometrics authentication) using physical features. The iris begins to be made in the 7th and 8th month of pregnancy and stabilizes in the second year after birth. The iris is composed of complex random patterns depending on pigments and muscles, and the iris pattern is different between the left and right eyes of identical twins or the same person. The advantage of iris personal identification is that it has been stable for two years after birth, and since then it has not changed, so counterfeiting that does not require re-registration is difficult. Another point is that there is a lower possibility of injury than a finger or face. Therefore, iris recognition can be expected as a security system for personal computer and mobile phone passwords, gate management for entrance and exit.

[0003] As a technique used for conventional iris recognition, there is an iris pattern matching method as disclosed in Patent Documents 16-16. Patent Document 1 discloses a method of identifying an eye position by detecting the center position and diameter of a pupil and an iris in iris recognition. Patent Document 2 discloses a method of determining impersonation by extracting the density of a specific region of a human eye in iris recognition. Patent Document 3 discloses an iris authentication device that detects whether or not an iris is a living iris by eye movement or the like in iris recognition. Patent Document 4 discloses an iris imaging device that enables iris imaging even if the identified person moves after the position of the identified person is measured in iris recognition. Patent Document 5 discloses a personal identification device that makes it possible to identify and identify the iris of a person to be identified! Patent Document 6 discloses an iris image acquisition device that enables the imaging of an iris image by searching for the position of the eye from the silhouette of the person to be identified in iris recognition.

[0004] However, the conventional pattern matching method used for iris recognition is vulnerable to rotation of the recognized iris because the iris is circular, and it takes time to recognize because it matches the registered image during recognition. There was a problem of power. Also, when using the system, the subject had to adjust his eyes to the designated position on the system. This is very problematic when the iris recognition system is miniaturized and used as a password for a PC or mobile phone.

[0005] On the other hand, focusing on the information processing process of the spatial recognition 'memory system (parietal association area) of the brain, as disclosed in Non-Patent Documents 1 and 2, the rotation that can recognize the rotation direction and shape of the object A diffusion type neural network has been announced by the present inventors. This rotation-diffusion neural network performs polar coordinate conversion and is suitable for recognizing the shape and rotational orientation of concentric patterns such as irises. In addition, orientation-invariant shape recognition and shape-invariant orientation recognition are possible at the same time. Furthermore, since the memory matrix is created during learning (registration), the recognition time is very short.

Patent Document 1: Japanese Patent Application Laid-Open No. 2004-21406

Patent Document 2: JP 2002-312772 A

Patent Document 3: Japanese Patent Laid-Open No. 2001-34754

Patent Document 4: Japanese Unexamined Patent Publication No. 2000-11163

Patent Document 5: JP-A-10-137220

Patent Document 6: JP-A-10-137219

Non-patent literature 1: IEICE Transactions

, D-II, VOL.J81-D-II, No.6, pp.ll94-1204,1998

Non-Patent Document 2: IEICE Technical Report, NC2002-207, pp.25-30,2003 Disclosure of Invention

Problems to be solved by the invention

[0006] However, the conventional rotational diffusion type-Ural network recognizes characters and faces using still images, and its use is limited for recognition in offline processing using still images. It was inferior in practicality. In practice, it is impossible to capture the rotation direction of the iris pattern in the same orientation as the learning (registered) image during recognition. Therefore, the iris image actually used for recognition is rotated and corrected! / Wow! Therefore, the recognition rate of practical accuracy was not reached. Furthermore, the identification (rejection) rate for unregistered iris patterns has not been investigated and verified, and is insufficient for application to personal authentication devices.

[0007] The present invention has been made in view of the above-described problems of the prior art, and can recognize a human iris image to accurately identify and authenticate a person with high accuracy and high speed. The purpose is to provide a personal authentication device that can be widely used in information systems and other management tasks.

Means for solving the problem

[0008] The present invention includes a camera capable of capturing a moving image and a storage device that stores images captured by the camera in a predetermined cycle, and the image information of a person's face captured by the camera is used as an iris of a human eye. Or, compare with a template that falls within the size range of the pupil, and scan the iris or pupil to detect the iris or pupil part of the eye, and the iris pattern that captures the iris pattern detected by this iris' pupil detection means The rotation direction and shape of the iris pattern is obtained by transforming the rotation direction and shape of the iris pattern by a rotational diffusion type-Ural network that creates a diffusion pattern by multiplying a polar coordinate transformed image by a periodic Gaussian function. The memory conversion means for learning and storing the shape as vector information, and learning by converting by the memory conversion means. Similarly, an iris pattern shape determining means for comparing an arbitrary acquired iris pattern with the vector information converted by the rotational diffusion type neural network by the storage conversion means in the same manner as described above, and by the iris pattern shape determining means The shape determination is performed by recognizing and correcting the direction of each vector information to be compared with each other, forming an orientation memory matrix and a shape memory matrix in the vector information, and recognizing the orientation that is an output of the rotational diffusion type-Ural network. -Eron and shape recognition -Eron is a personal authentication device that performs shape recognition by correlating each of the recognized iris shapes with each other. [0009] During learning by the iris pattern acquisition unit, an original image of a predetermined 0 ° azimuth of the iris or a polar coordinate conversion image thereof is registered as vector information, and at the time of recognition, the iris pattern shape determination unit can perform arbitrary orientation. The input iris image is corrected using the recognition orientation obtained by the rotation diffusion type neural network, and the rotation diffusion type-Ural network output is obtained using the original image or its polar coordinate conversion image as the vector information, This is to be compared with a preset threshold value.

[0010] Further, the iris pattern acquisition means includes pupil center position detection means by labeling, center correction by least square method, iris edge detection means using Laplacian filter, and iris size standard using line correction. It has an eaves means.

[0011] At least one of the inner product and the minimum distance of each vector information is obtained, and the identity / inconsistency of each vector information is judged by comparing with a preset threshold value to perform personal identification.

[0012] The iris pattern acquisition means includes a flash light emitting device that causes a pupil reaction, an infrared light source for imaging, and an infrared transmission filter attached to the lens of the camera, and the size of the pupil is substantially constant. In this state, an iris image is acquired. Furthermore, the iris pattern acquisition means continuously measures the time change of the pupil diameter, and can cope with impersonation using a still image such as a photograph.

In addition, the correction of the azimuth in the shape determination by the iris pattern shape determination means replaces the correction of the azimuth recognized by the rotational diffusion type-Eural network, and learns the diffusion pattern of the iris used for recognition. By using at least one of the inner product and the minimum distance of the respective vector information using the plurality of azimuth diffusion patterns, the input azimuth is recognized and the azimuth is corrected by vector synthesis.

[0014] The iris pattern acquisition means searches for the position of the person's face and eyes to be recognized from a low-resolution image by thinning out, identifies the position of the eyes, and then detects an iris region from the high-resolution image. It's okay.

[0015] In addition, each pixel value in a certain area of an image including a human eye is measured to obtain an average luminance, and this is standardized to a certain value to set the luminance of the iris. Furthermore, the luminance average and standard deviation of the pupil and iris of the measurer are obtained, and the binary threshold value determined by the ratio of the standard deviation is used. It is a personal authentication device that determines the pupil and iris.

[0016] Further, the pupil diameter change at the time of light reflection caused by flash light irradiation is measured, and the biological reaction is detected from the difference or ratio between the maximum pupil diameter and the minimum pupil diameter, and this difference or ratio is the reference value. If the following, the measured iris image is a personal authentication device that determines that the image is impersonated.

The invention's effect

[0017] The personal authentication device according to the present invention can accurately and quickly identify an individual using an iris pattern, and reliably prevent impersonation due to iris counterfeiting that reduces the burden on the subject at the time of authentication. Can be stopped. In particular, since the rotation and displacement of the iris pattern can be corrected, it is not limited by the usage pattern such as the installation status of the authentication apparatus.

Brief Description of Drawings

FIG. 1 is a schematic diagram showing a device configuration of a personal authentication device according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram showing a rotational diffusion type-Ural network used in the personal authentication device of this embodiment.

FIG. 3 is a conceptual explanatory view showing image conversion of a rotation diffusion type-Ural network used in the personal authentication device of this embodiment.

FIG. 4 is a schematic flowchart showing iris image acquisition using a rotation diffusion type-Ural network used in the personal authentication device of this embodiment.

FIG. 5 is a graph showing changes in pupil diameter due to flash irradiation of the personal authentication device of this embodiment.

FIG. 6 is a front view showing an eye and a template in iris image acquisition by the personal authentication device of this embodiment.

FIG. 7 is a schematic diagram showing labeling of the personal authentication device of this embodiment.

FIG. 8 is a front view showing a display screen when an iris image is acquired by the personal authentication device of this embodiment.

FIG. 9 is a front view showing an image for obtaining an average luminance for acquiring an iris image by the personal authentication device of this embodiment.

[Fig. 10] Brightness of iris and pupil in iris image acquisition by personal authentication device of this embodiment It is a graph which shows a frequency value and a cumulative pixel number.

FIG. 11 is a schematic flowchart showing processing of the personal authentication device of this embodiment.

12] A vector diagram showing the relationship between the inner product used for processing by the personal authentication device of this embodiment and the minimum distance.

13] A flowchart showing the recognition process of the personal authentication device of this embodiment.

14] This is a flow chart of impersonation determination in the recognition processing of the personal authentication device of this embodiment.

[15] An eye image for recognizing an iris pattern used in an embodiment of the personal authentication device of the present invention.

FIG. 16 is a graph showing the orientation recognition characteristic (a) and the shape recognition characteristic (b) of the recognition experiment result by the example of the personal authentication device of the present invention.

17] A graph showing a direction recognition characteristic (a) and a shape recognition characteristic (b) when a real-time recognition experiment is performed by five persons according to an embodiment of the personal authentication device of the present invention.

[18] FIG. 18 is a graph showing an identity rejection rate and an acceptance rate by others according to an embodiment of the personal authentication device of the present invention.

19] This is a graph showing the rejection rate and the rejection rate using the inner product as a criterion, according to one embodiment of the personal authentication device of the present invention.

20] This is a graph showing the rejection rate and the acceptance rate of others using the minimum distance as a criterion, according to an embodiment of the personal authentication device of the present invention.

Explanation of symbols

10 huge

12 Iris

14 Camera

15 Lens

16 computers

18 display

20 Near-infrared projector

22 Flash light emitting device 24 Infrared transmission filter

26 Template

27 pupil

BEST MODE FOR CARRYING OUT THE INVENTION

Hereinafter, a first embodiment of the personal authentication device of the present invention will be described with reference to FIG. 1 and FIG. The personal authentication device of this embodiment uses a rotation diffusion type-Eural network. First, the rotation diffusion type-Eural network will be briefly described. Rotation-Diffusion-Eural Network is a diffusion pattern created by multiplying a polar coordinate transformed image by a periodic Gaussian function in the rotation direction, and recognizes the direction of the object and the shape of the object. It consists of a shape recognition system. Figure 2 shows a conceptual diagram of the rotation diffusion type-Eural network of this embodiment. The network's orientation recognition memory system-Euron (azimuth recognition-Euron) is arranged, for example, 30 per 12 ° on the circumference, and the shape recognition memory system-Euron (shape recognition-Euron) is an appropriate number, For example, 10 pieces correspond to the shape of each object.

[0021] This rotational diffusion type-Eural network inputs the converted image on the polar coordinates generated by the original image force to the diffusion layer, and diffuses the rotation information in the surrounding space. The object orientation and shape are recognized using the diffusion pattern that is the output of the diffusion layer. In this invention, the non-rotation state used conventionally and the rotation reference point in the general xy plane of the mathematical Cartesian coordinate system (for example, (X, y) = (l, 0) from the origin) The coordinate system was rotated 90 ° counterclockwise to match the position vector. As shown in the explanatory diagram of FIG. 3, the point (X, y) on the Cartesian coordinates corresponds to the point (R, Θ) on the polar coordinates, and x = Rcos Θ and y = Rsin Θ. The object direction is defined as the rotation angle from the non-rotation state to the counterclockwise direction, with the rotation angle being 0 °. In addition, object orientation is a problem in object orientation recognition, but measures against this can be done by another method. Therefore, in object orientation recognition, the figure (object) is positioned at the center of the image, and the center of rotation of the object coincides with the origin on the xy coordinates of the original image.

In the explanatory diagram of FIG. 3, it is assumed that the original image used for learning and recall (Arabic numeral 1) is a binary image of 480 X 4 80 dots, and the original image is displayed in polar coordinates with a certain radius and angle. Split and The created image is used as a converted image. Figure 3 shows an example of the Arabic numeral 1 with a rotation angle of 0 ° and its converted image. Object orientation recognition is performed on the premise that object position recognition has already been performed, and it is considered that there is no effect of object position deviation. The converted image is generated using Equation (1).

[Number 1]

9

Γ _{ι θ} ^ ∑ ∑ I (x _{i jt} Y ij) (l)

/ = 0

[0023] Here, I (x, y) is a point on the xy coordinate (x = Rcos Θ, y = Rsin Θ) with the origin at the center of the original image

ij ϋ ϋ ϋ

), T is the pixel value at coordinates (r, Θ) on the converted image. Formula (1) divides the original image by 20 with a radius of 200 dots and an angle every 3 °, and further divides the small area surrounded by the boundary into 10 x 10, and calculates the value of each point. It is shown that the sum of them is used as the value of one element of the transformed image. For this reason, one element of the converted image has a value of 0–100, r takes a value from 1 to 20, and Θ takes a range of integer values from 1 to 120. Figure 3 shows how to calculate the pixel value T at the coordinates (r, Θ) of the transformed image by dividing a small area with a radius of 10 dots X angle 3 ° on the corresponding original image into 10 X 10 Each point is represented by (X, y), and the total pixel value I (x, y) is calculated.

On the other hand, in the case of the rotational diffusion type-Eural network of this embodiment shown in FIG. 2, first, an input pattern of 300 X 300 pixels is polar coordinates of 25 X 120 pixels excluding the pupil (center) portion. Convert to the image above. Next, a polar coordinate conversion image is input to the diffusion layer to obtain a diffusion pattern that is vector information. By applying an azimuth memory matrix and a shape memory matrix to this diffusion pattern, an azimuth recognition-Euron output and a shape recognition-Euron output are obtained. The resulting 30 orientation recognition-Euron output forces are also recognized using the population vector method. In addition, shape recognition associates each recognition object with a different shape recognition-Euron on a one-to-one basis, and shape recognition is performed using the maximum output of these -Eurons.

[0025] The operation of this rotational diffusion type neural network is performed by learning using an orthogonal learning method in the rotational diffusion type dual network in the inscription process. The number of learning patterns is determined by the number of learning irises X the number of learning orientations. For example, in the case of 10 irises, 10 (iris) X 6 (way Place) = 60 (pattern). In the memorization process that forms the memory, learning is performed between the diffusion pattern V of the transformed image for learning V and the orientation recognition-Euron teacher signal TO and the shape recognition neuron teacher signal TF according to equations (2)-(7). , Orientation memory matrix M and shape memory

o

Generate a matrix M.

F

[Equation 2]

`` V (1) V (2) (60)

[Equation 3]

[Equation 4]

Μ = ΤΟΧ

[Equation 5]

TO = [TO ^(,) , TO (2) To (60)] ₍₅

[Equation 6]

[Equation 7]

TF-[TF ⁽¹⁾ TF (2) TF (60)

] (7) Next, in the recall (recognition) process that recalls memory, the transformed image of the input iris image in an arbitrary orientation is input to the diffusion layer, and the output is the product of the diffusion pattern V and the orientation memory matrix M. Orientation recognition-Euron output YO, product of diffusion pattern V and shape memory matrix M

OF The shape recognition-Euron output YF is obtained by (Equations (8) and (9)).

[Equation 8]

YO = M. V (8)

[Equation 9]

YF = M _F V (9)

Next, FIG. 1 shows a system configuration diagram of one embodiment of the present invention. This system has a small camera 14 and a lens 15 capable of capturing a moving image for capturing an iris 12 of a human eye 10, a computer 16 for capturing the captured iris image, and a display 18. The main body of the computer 16 includes an image input board for capturing image data into the CPU, a DLL (Dynamic Link Library) for manipulating and processing iris images, and other storage devices. The small camera 14 also cuts the visible light noise reflected in the iris 12, near-infrared projector 20 which is an infrared light source for clearly capturing iris patterns, flash light emitting device 22 for causing pupil reflection, and so on. A plastic infrared transmission filter 24 is installed.

The light emitting device 22 can emit light at an arbitrary timing (in frame units) in synchronization with an external trigger output signal from the image input board of the computer 16. The input image is a grayscale image with 256 gradations of 640 x 480 pixels. This system is capable of real-time image capture at approximately 13 frames Z s. The computer 16 and its operating system used a commercially available personal computer.

[0029] Next, the processing flow of the iris recognition system according to the embodiment of the present invention is shown in the flow chart of FIG. First, the eye 10 of the person to be recognized is photographed with the small camera 14 (slO). Next, when the number of imaging frames is 19 frames or 50n + 19 frames (n = 1, 2,...), The imaging iris diameter is initialized (si 1). Next, the flash light is applied to the eye 10 by the flash light emitting device 22 while the number of frames is 20 frames, or 50n + 20 frames and 50n + 20 + n frames (n = l, 2, ...) (sl2) . In this system, as shown in Figure 5, The image of iris 12 when the pupil diameter is constant pupil size (2.9 mm-3. Omm) is acquired. In order to obtain an iris image, the average luminance of a certain region including the eye 10 as shown in FIG. 9 is used to normalize the image, and the one-eye partial template 26 as shown in FIG. The pupil 27 is detected from (sl3). Then, whether or not the portion detected by the one-eye partial template is the pupil 27 is determined using labeling which is an existing determination method (sl4). As shown in Fig. 7, labeling detects specific parts such as irises by attaching the same label (number) to all connected pixels (connected components) and assigning different numbers to different connected components. It is a method to do.

[0030] When it is determined that the detected part is the pupil, the pupil center, the pupil diameter, and the pupil area are simultaneously measured, and the pupil detection is completed (sl5). After pupil detection, the pupil center measured by the above labeling is corrected using the least squares method. After correcting the pupil center, if the iris diameter is initialized, the iris diameter is measured. If the iris diameter is initialized, the previously measured value is used as it is (si 6). In addition, Laplacian processing is performed in the measurement of iris diameter. At each radius of the pupil, the right iris is 0 ° above the center of the pupil, the counterclockwise direction is +, the angle is 75 ° —— 135 °, and the left iris is 1 ° to the angle 75 ° — 135 °. Add the pixel values of each. The parts with the maximum cumulative pixel value are the right edge of the iris and the left edge of the iris, respectively.

[0031] Next, using the fact that the iris diameter is almost constant regardless of the person, the relative ratio force between the measurement size (pixels) on the image of the iris 12 and the pupil is 2.9 mm— 3. Acquire the Omm image as the reference image (si 7). Here, the pupil diameter is set to 2.9 mm-3.0 mm in order to allow a slight size error in order to facilitate the acquisition of the iris image. Fig. 8 shows the screen of display 18 when an iris image is acquired. From the obtained image, a 300 x 300 pixel image centered on the pupil is cut out (sl8). Since the measurement size on the image of the iris and pupil changes depending on the distance from the camera 14 and the zoom, the size is standardized using a known linear interpolation method in order to make the iris size constant. In addition, in order to reduce the influence of the surrounding light environment, the average luminance of multiple images is obtained and a correction coefficient is set to normalize the luminance (sl9). This standardized image is defined as a reference iris pattern (rotation-diffused-input image of the neural network) (s20). To learn and memorize the reference iris pattern.

Here, a method for determining a luminance standard for detecting pupils and irises will be described. Here, as shown in FIG. 9, each pixel value in a certain area A of the image including the eyes is measured to obtain an average luminance, and this is standardized to a constant value for each measurement image, thereby allowing each measurer to The variation in luminance of each acquired image is eliminated. The average brightness of area A was measured in order to ensure that the sclera, iris, and pupil are clear even after the brightness standard is reached. This is to set the value to the middle of the 256-level display. Range B surrounded by the inner line is the pupil detection range.

[0033] Further, a method for determining a binary threshold for pupil detection will be described. After the luminance standard, measure the luminance of the iris and pupil. The measurement method is the same as when the luminance standard is determined. The optimum binary threshold is determined from the ratio of the standard deviations obtained by calculating the average brightness and standard deviation of the pupils and irises of multiple measurers. The threshold Y is determined by the following formula.

Y = (AV -AV)-SD / (SD + SD) + AV

2 1 1 1 2 1

Where AV is the average pupil brightness, AV is the average iris brightness, and SD is the pupil brightness standard.

1 2 1

Semi-deviation, SD

2 is the standard deviation of iris brightness.

[0034] Fig. 10 shows a graph of the cumulative number of pixels for each luminance value using the data of all subjects after the luminance specification. According to Fig. 10, it can be seen that the iris luminance and pupil luminance are clearly separated by the binary key threshold. As a result, the center of the pupil is detected by the one-eye template, and the pupil end 'iris end' is accurately detected.

On the other hand, an iris input pattern at the time of recognizing the iris of another person is also obtained by the same procedure as that for acquiring the reference iris pattern. Using the obtained image, personal identification is performed by shape recognition. In this system, the regular diffusion pattern is used to perform learning and recognition using the above-described rotational diffusion type-Eural network.

[0036] Further, in the present invention, a new shape recognition criterion is added in order to improve the discrimination accuracy of the unlearned iris. The new shape recognition criteria used the inner product and the minimum distance, which are often used to investigate vector similarity. However, these methods are known to be vulnerable to pattern variations. In the case of iris recognition, there are various pattern variations, but one of them is misorientation. Image from camera 14 It is almost impossible to capture the learning image and the recognition image in exactly the same direction when performing recognition with. Therefore, it is possible to introduce the inner product and the minimum distance as shape recognition criteria by correcting the direction using orientation recognition, which is a feature of the rotational diffusion type-Eural network.

[0037] Next, a description will be given of the azimuth correction when performing personal authentication by the rotational diffusion type-Eural network. If the orientation of the learning image is defined as 0 °, the orientation of the input image at the time of recognition is not necessarily 0 °. Unlike other people. Therefore, in the authentication using the rotational diffusion type neural network, the orientation is first recognized and the orientation of the input image is corrected.

[0038] Fig. 11 shows the flow of personal authentication using a rotation diffusion type-Eural network. First, the direction correction range, step angle, and recognition method are selected. Let us consider the case where the lower limit of azimuth correction is set to 3 °, the upper limit is set to 3 °, and the step angle is set to 1 °. Rotation-diffusion type-If the result of orientation recognition by the Yural network is 10 °, the orientation of the input image used for recognition is rotated 10 ° counterclockwise with respect to the learning image (registered person's iris pattern). Will be. Therefore, in order to make the orientation of the input image coincide with the orientation of the learning image, the step is rotated by 10 ° ± 3 °. In this case, the correction azimuth is 13 ° to 7 °, and an iris pattern with seven different azimuth corrections is obtained. The reason why the correction azimuth is not limited to 10 ° is to take into account the resolution and error of the recognition azimuth obtained by the rotational diffusion type-Eural network. Using the iris pattern with azimuth correction, individual iris pattern (shape) authentication is performed by the specified recognition method (inner product, minimum distance, shape recognition-Euron output). In recognition by inner product and minimum distance, the vector calculation of the input image and the learning image (inner product, minimum distance) corrected for rotation within the specified range based on the rotation orientation recognized by the rotational diffusion type-Eural network. I do . In this example, the inner product and minimum distance between the input image and the learning image that have been subjected to seven types of rotation correction (correction direction: –13 ° —— 7 °) are calculated.

[0039] The maximum value is used for the inner product, and the minimum value is used for the minimum distance. In the case of an inner product, if the maximum value is greater than the determination threshold, the person is identified, and if the maximum is less than the threshold, the person is determined to be another person. Minimum if minimum distance If the value is smaller than the determination threshold value, the person is identified, and if the value is larger than the threshold value! These vector operations are performed for all registered learning images to identify individuals. If none of the learning images is determined to be the person, it is determined that they are not registered. When shape recognition-Euron output is used for shape determination (personal authentication), the orientation-corrected image is input as an input image again to the rotational diffusion type -Eural network, and it is determined by shape recognition -Euron output. If the shape recognition neuron output representing the registered person is larger than the preset judgment threshold, it is judged that the input iris image is the iris of the person representing -Yelon. If any shape recognition-Euron output does not exceed the judgment threshold, it is registered and judged as a bad one.

[0040] Here, the defining formula for the inner product is

[Equation 10]-= (1 0)

It is expressed. V is a vector representing the normal diffusion pattern of the learning iris image, V is for recognition

L R

This is a vector representing a normal diffusion pattern of an iris image to be used. Also, | v |,

L I V R

I is the absolute value of V and V, and represents the length of each vector. here,

L R I V L I

= 1, I V

R I = 1 (normalized). Therefore,

[Equation 11] _co ^V L; ^V R _{= V} V (1 1)

| VL || V _R | ^L "and cos Θ represents the similarity between the two vectors V and V. That is, when V = V, cos

L R L R

Θ = 1, indicating the highest similarity.

[0041] The minimum distance is the vector difference (distance) of I V— V

L R I, as shown in Figure 12,

Show. When V = V, the distance becomes zero, indicating the highest similarity. V, large of V

L R L R

If the standard is set to 1, the square of the minimum distance is expressed as an inner product.

[Equation 12] v _L -v _R | ² = (v _L -v _R )-(v _L -v _R )

= | v _L | + | V _R | -2V _L V _R

= 2-2V _L -V _R (12) Therefore, the relationship between the minimum distance and the inner product is expressed as follows.

[Equation 13]

[0042] From this equation, when the similarity between two vectors is the highest (V = V), the inner product is the maximum value V ·

L R L

Since V = 1, the minimum distance is I V-V I = 0 and takes the minimum value. Therefore, V and V

R L R L R

When the size is set to 1, the result of shape identification (personal authentication) by the minimum distance and inner product is the same.

FIG. 13 shows a flowchart of the recognition of the rotational diffusion type-Eural network in the above-described iris pattern recognition according to this embodiment. Orientation correction is performed in the recognition process when the inner product and the minimum distance are used, since the iris image of 0 ° is used for the learning (registered) image, the recognition target must be in the 0 ° orientation. Because there is.

According to the personal authentication device of this embodiment, the orientation of the iris pattern can be recognized by the rotational diffusion type-Ural network, so that it is possible to cope with the rotational change by correcting the orientation. Furthermore, by introducing the inner product and the minimum distance as the shape recognition criterion, it is possible to achieve a 0% acceptance rate by using an iris-corrected iris pattern. Moreover, it is possible to automatically detect the pupil center, pupil edge, and iris edge by detecting the pupil center position by labeling, center correction by the least square method, and edge detection using a Laplacian filter. The iris and pupil size on the image will change depending on the distance from the camera 14 and zoom, but it can be scaled up or down using linear interpolation and the iris size can be standardized to accommodate the size change. it can. Furthermore, by imposing a flash light on the eye 10 to induce a pupil response and measuring a temporal change in the pupil diameter, it is possible to reject impersonation using an iris photograph or the like.

[0045] The personal authentication device of the present invention creates a memory matrix that characterizes each learned iris. Therefore, the recognition time is short because the amount of calculation is small. In addition, since the pupil center position, pupil edge, and iris edge are automatically detected, eye alignment is not required and it can be used in a wide range of applications.

[0046] It should be noted that the personal authentication device of the present invention is not limited to the above-described embodiment, but has a camera power capable of capturing moving images. If an iris image can be directly captured by the CPU, the image input board can operate the iris image. A DLL for processing is not always necessary. In addition, the personal authentication device of the present invention can be used without using a flash light emitting device. In this case, an iris image converted into a fixed relative pupil size by the linear interpolation is captured.

[0047] Further, the vector information used for recognition is not limited to the original image captured as described above or its polar coordinate conversion image, but may be information obtained by performing image processing on the original image using a diffusion pattern or a Laplacian filter. .

[0048] Further, as a method for preventing impersonation, the following processing is performed to further prevent impersonation. In this case, it is determined that the pupil's light reflection should occur, and in the flowchart of FIG. 4, the processing below sl6 and sl7 is performed as shown in FIG. In this process, the LED is flashed to obtain a constant pupil size, and the impersonation is determined by comparing the relative pupil diameter before and after the light emission. In the process of Fig. 14, the relative pupil diameter from when the first LED emits light during recognition until the light reaction of the pupil occurs is used for impersonation determination. In this embodiment, the first flash light is emitted at the 20th frame of the measurement image, so the light response occurs by 30th frame. Therefore, the relative pupil diameter of 20 to 29 frames where flash light is emitted and light reflection occurs is preserved. The maximum or minimum is calculated from the stored relative pupil diameters, and the difference or ratio is calculated. If the difference or ratio is equal to or less than the reference value, it is determined to be impersonation.

Example 1

[0049] Hereinafter, the results of implementing the personal authentication device using the iris pattern of the present invention will be described.

First, the iris images used for learning and recognition were changed to 3, 5, and 10, and a recognition experiment was performed. Fig. 15 shows the iris image (300 X 300 pixels) used for learning and recognition. Learning, All recognition was performed with the iris image of the subject's right eye. The learning orientation was 0 ° and 360 ° with 6 orientations every 60 °. The number of learning patterns is given by (number of recognized irises) X (number of learning orientations), and is 18 patterns, 30 patterns, and 60 patterns for each number of subjects.

[0050] Fig. 16 (a) shows the orientation recognition characteristics and Fig. 16 (b) shows the shape recognition characteristics when the recognition experiment was performed by 10 people. (a) In the orientation recognition characteristics of the figure, the horizontal axis is the input rotation orientation of the iris, and the vertical axis is the recognition orientation. As a result, very good linearity was seen in the input rotation direction and recognition direction of the iris, and it was possible to recognize the direction correctly. The azimuth error was a mean standard deviation = -0.83 ± 0.75 °. From this, it can be said that the rotation direction of the iris of 10 people can be recognized in all directions. Also, in the shape recognition characteristics in (b), the horizontal axis is the input rotation direction of the iris, the vertical axis is the shape recognition-Euron output, ◯ is the average value of the target neuron output, and X is the average value of the non-target neuron output Represents a value. The vertical line in each input direction represents the standard deviation. The target neuron output was 0.67—1.15 (mean standard deviation = 0.94 ± 0.11), and the non-target neuron output was −0.86—0.51 (0.02 ± 0.18). 0. Power 360. The average value of the target neuron is approximately 1.0 and the average value of the non-target neuron is approximately 0.0, and the target neuron output is higher than the non-target neuron output. Always big. From this, it can be said that the iris recognition ability of 10 subjects is possible in all directions. The results of the recognition experiment with three and five people were also able to recognize both the direction and shape correctly, as in the case of ten people.

[0051] Next, the iris images used for learning and recognition were changed to three and five people, and after learning in advance, a real-time recognition experiment was performed. The iris image (300 X 300 pixels) used for learning was selected from Fig. 15. Learning'recognition was performed with the same subject's right eye iris. The number of learning notes is given by (number of recognized irises) X (number of learning orientations), 18 and 30 patterns for 3 and 5 subjects, respectively. The images used for learning were taken in advance on the day of the experiment.

[0052] Fig. 17 (a) shows the azimuth recognition characteristics when a recognition experiment is performed by five persons, and Fig. 17 (b) shows the shape recognition characteristics. In the azimuth recognition characteristics of Fig. 17 (a), the horizontal axis is the input rotation direction of the iris, and the vertical axis is the recognition direction. Good linearity was seen in the input rotation direction and recognition direction of the iris, and it was found that the direction could be recognized almost correctly. Azimuth error is averaged standard deviation = -3.92 It was ± 0.22 °. From this, it can be said that the personal authentication apparatus of this embodiment can recognize the rotation directions of the five irises in real time in all directions. Also, in the shape recognition characteristics of Fig. 17 (b), the horizontal axis is the input rotation direction of the iris, the vertical axis is the shape recognition-Euron output, ◯ is the average value of the target neuron output, and X is the non-target neuron output. Represents an average value. The vertical line in each input azimuth represents the standard deviation. The target neuron output was 0.67—1.27 (mean standard deviation = 0.99 ± 0.20), and the non-target neuron output was −0.69—0.58 (0.03 ± 0.24). In the input rotation direction of the iris from 0 ° to 360 °, the average value of the target neuron is approximately 1.0, the average value of the non-target neuron is approximately 0.0, and the target neuron output is always higher than the non-target neuron output. large. As a result, it was found that real-time iris recognition by five subjects is possible in all directions. As a result of conducting real-time recognition experiments with three people, the orientation and shape were correctly recognized as in the case of five people.

Next, using the authentication system according to the flowchart shown in FIG. 13, a person authentication experiment was performed using images captured from the camera. The experiment was conducted with 10 subjects. Of the 10 people, 5 were studied and 5 were unlearned. The images used for learning were taken in advance on the day of the authentication experiment. Learning was performed in a total of 10 sets by replacing the iris images of five people used for learning one by one. As the learning iris image, the iris image when the input rotation direction of the iris is 0 ° was used. Since 10 subjects were recognized for 10 sets of learning, a total of 100 trials (50 trials for learning and 50 trials for unlearned persons) were obtained.

FIG. 18 shows the error rate using the shape recognition-Euron output, FIG. 19 the inner product, and FIG. 20 using the minimum distance as a criterion. The dotted line represents the rejection rate (the rate of rejecting the user by mistake), and the solid line represents the acceptance rate of others (the rate of accepting others by mistake). The vertical axis represents each error rate, and the horizontal axis represents the evaluation threshold. The method of calculating the rejection rate of the person is considered to have rejected the person if the output value of the shape corresponding to the person-the euron or inner product is smaller than the judgment threshold, or if the output value of the minimum distance is larger than the judgment threshold, The number of trials was counted and asked to determine the rejection rate. As a special case of rejecting the person, if the output value of the shape recognition neuron or inner product corresponding to the person exceeds the judgment threshold and the output value corresponding to the other person exceeds those, the output of the minimum distance corresponding to the person The value is smaller than the judgment threshold, but other It is conceivable that the output value corresponding to a person is lower than those. As a result of the recognition experiment, there was a special case of rejecting the person with shape recognition-Euron output 8%, inner product, minimum distance 2%. On the other hand, if the maximum value of the non-target neuron output and the inner product is larger than the judgment threshold, and if the minimum value of the minimum distance is smaller than the judgment threshold, it is assumed that the other person has been accepted (the rate of accepting others). The acceptance rate of others was calculated by counting the number of trials. From the experimental results, when the shape recognition neuron output was used as the criterion, the intersection of the rejection rate and the rejection rate was about 0.78, and the error rate was about 43%. When the inner product was used as the criterion, the intersection between the rejection rate and the acceptance rate was about 0.94, and the error rate was about 15%. However, if the decision threshold is 0.96, the rejection rate is 20%, but others can be completely rejected. When the judgment based on the minimum distance was used as the criterion, the intersection of the rejection rate and the acceptance rate of others was when the decision threshold was about 0.35 and the error rate was about 13%. However, when the decision threshold was 0.25, the rejection rate was 26%, but others could be completely rejected.

Claims

The scope of the claims

[1] A camera capable of capturing moving images and a storage device for storing images captured by the camera at a predetermined cycle,

Iris / pupil detection means for scanning the human face image information captured by the camera with a template that falls within the range of the iris or pupil size of the human eye and detecting the iris or pupil part of the eye When,

An iris pattern acquisition unit that captures an iris pattern detected by the iris / pupil detection unit, and a rotation diffusion type that creates a diffusion pattern by multiplying a rotation coordinate and shape of the iris pattern by a polar gauss function on a polar coordinate transformed image. -A memory conversion means for converting and learning as 'vector information' by means of a Yural network;

Learning by converting by the memory conversion means

And an iris pattern shape determining means for comparing an arbitrary iris pattern obtained in the same manner with the vector information converted by the memory conversion means by the rotational diffusion type-Ural network in the same manner as described above,

The shape determination by the iris pattern shape determination means recognizes and corrects the azimuth of each vector information to be compared with each other, forms an azimuth memory matrix and a shape memory matrix in the vector information, and the rotation diffusion type-Ural Recognize orientation recognition-Euron and shape recognition-Euron, which is the output of the network, and shape recognition of each recognition target iris-Euron, and shape recognition corresponding to each memorized shape recognition-Euron The personal authentication device is a feature.

[2] At the time of learning by the iris pattern acquisition means, an original image of a predetermined orientation of the iris or its polar coordinate conversion image is registered as vector information,

At the time of recognition, the iris pattern shape judging means corrects the iris image input in an arbitrary direction using the recognition direction obtained by the rotational diffusion type-Eural network, and the original image or its polar coordinate conversion image is vectorized. The personal authentication device according to claim 1, wherein the information is compared with a preset threshold value as information.

[3] The iris pattern acquisition means uses pupil center position detection means by labeling, center correction by least square method, iris edge detection means using Laplacian filter, and line correction. 2. The personal authentication device according to claim 1, further comprising iris size standard means.

[4] The method according to claim 1, wherein at least one of an inner product and a minimum distance of each vector information is obtained, and whether each vector information matches or does not match is determined by comparing with a preset threshold value.

1. The personal authentication device according to 1.

[5] The iris pattern acquisition means includes a flash light emitting device that causes a pupil reaction of the iris portion, an infrared light source for imaging, and an infrared transmission filter attached to a lens of the camera, and the iris The personal authentication device according to claim 1, wherein the iris image is acquired in a state where the size of the pupil of the pupil is substantially constant.

6. The personal authentication device according to claim 5, wherein the iris pattern acquisition means continuously measures a temporal change in pupil diameter.

[7] The direction correction in the shape determination by the iris pattern shape determination means is performed in place of the correction of the direction recognized by the rotation diffusion type-Ural network, and a plurality of learned iris diffusion patterns used for recognition. 3. The personal authentication device according to claim 2, wherein at least one of the inner product and the minimum distance of each vector information is obtained using a diffusion pattern of the azimuth and the vector synthesizes to recognize the input azimuth and correct the azimuth.

[8] The iris pattern acquisition means searches the position of the face and eyes of the person to be recognized from the low-resolution image by thinning out, specifies the position of the eyes, and then detects the iris region from the high-resolution image. Item 1. Personal authentication device.

[9] The personal authentication device according to [1], wherein each pixel value in a certain area of an image including a human eye is measured to obtain an average luminance, and the luminance of the iris is set by normalizing the average luminance.

10. The personal authentication device according to claim 1, wherein the brightness average and standard deviation of the pupil and iris of the measurer are obtained, and the pupil and iris are determined by a binary value threshold determined by a ratio of the standard deviation.

[11] Measure the pupil diameter change at the time of light reflection caused by flash light irradiation, detect the biological reaction from the difference or ratio between the maximum pupil diameter and the minimum pupil diameter, and if this difference or ratio is below the reference value The personal authentication device according to claim 1, wherein the measured iris image is determined to be spoofing.