JP2018055231A - Biometric authentication device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a biometric authentication device with high system operability, high speed and high accuracy.SOLUTION: A biometric authentication device for authenticating an individual using features of a living body 1 comprises: an imaging device 9 for imaging the living body; and an authentication processing unit 10 that extracts biometric features of the living body by processing an image imaged by the imaging device 9 to perform authentication by comparing similarities of the biometric features, wherein the authentication processing unit 10 includes a living body determination unit acquiring temporal changes in luminance values of the living body appearing in the image and extracting texture features of the living body imaged in the image to determine that the living body is the true living body based on the temporal changes of these luminance values and the texture features.SELECTED DRAWING: Figure 1

Description

The present invention relates to an authentication system that authenticates an individual using biometric information, and more particularly to a high-speed and high-precision authentication technique that is excellent in forgery resistance and confidentiality of biometric information.

  Among various biometric authentication techniques, finger vein authentication is known as a technique that can realize highly accurate authentication. Since finger vein authentication uses a blood vessel pattern inside the finger, it achieves excellent authentication accuracy, and it is difficult to forge and falsify compared to fingerprint authentication, so that high security can be realized.

  In recent years, there have been an increasing number of cases in which biometric authentication devices are installed in devices such as mobile phones, notebook PCs (Personal Computers), mobile terminals such as smartphones and tablet terminals, lockers, safes, and printers to ensure the security of each device. ing. The fields to which biometric authentication is applied include entrance / exit management, attendance management, login to a computer, etc. In recent years, biometric authentication systems have also been introduced for payments and identity verification at event venues. In the future, for example, there may be a case where biometric authentication is performed when an event ticket is purchased at home, registration data is transferred to a server via the Internet, and identity verification is performed at the venue.

  In this way, as the usage scenes of biometric authentication devices diversify, it is considered more important than ever to maintain the reliability of the authentication system, and in particular to ensure the safe management of registered data and the registration of correct data. It is done. For example, in a situation where personal identification is performed using a personal smart device, unlike an ordinary office entry / exit management system, the biometric registration work cannot be performed by the administrator of the authentication system. It is necessary to perform registration work by the user himself. At this time, it is necessary to prevent a user from illegally registering a foreign object or a forged biological body, or an outsider from illegally intercepting or falsifying registered data. Therefore, a technique for determining that the presented living body is a true living body and a technique for realizing registration management of biological information with high confidentiality and tamper resistance are desired. In addition, when the registration terminal and the authentication terminal cooperate with each other via the network, or when different authentication systems cooperate with each other, many authentication terminals access the registration data in the authentication server. Therefore, it is essential to apply a data protection technique that converts registered data into information that can be safely disclosed.

  Non-patent document 1 as a technique for determining whether or not a presented face image is counterfeit is a technique for safely publishing registered biometric information as prior art of attack avoidance and data protection against a biometric authentication system. There exists patent document 2. FIG.

Jukka Maatta, et al., "Face Spoofing Detection from Single Images Using Micro-Texture Analysis," IJCB 2011.

Patent Publication No. 2015-226323

  In order to enhance the reliability of the input biometric information and to carry out highly practical authentication processing while maintaining the secrecy of the biometric information, a technique for determining that the presented subject is a true living body, and Therefore, a technique for collating encrypted biometric information with high speed and high accuracy is required.

  In the biometric authentication technique described in Non-Patent Document 1, in order to determine whether the face photographed by the camera is an actual face or a printed face, LBP (Local Binary) that is a local texture feature of the face image is used. (Pattern) The feature quantity is extracted, and this histogram is learned by SVM (Support Vector Machine) to determine whether it is a real face or a printed face image. However, the attack target is limited to printing on paper, and there is no suggestion that it can respond to other threats. In addition, there is no technical disclosure specifically for detecting forgery of biometric authentication using a fingertip.

  In the biometric authentication technique described in Patent Document 2, there is a description of a technique related to biometric authentication that protects privacy based on an error correction code. However, a technical disclosure for collating secret biometric features with high speed and high accuracy. There is no.

  Therefore, in the present invention, in order to realize a highly reliable authentication system that can be used in various situations and high-speed and high-precision biometric authentication, it is determined with high accuracy that the presented subject is a true living body, An object of the present invention is to realize a biometric authentication device that can collate biometric information with high speed and high accuracy.

  In order to achieve the above object, in the present invention, an imaging unit that captures a living body and an authentication processing unit that processes an image captured by the imaging unit and authenticates the living body are provided. A living body that acquires a temporal change in the brightness value of a living body that is reflected, extracts a texture feature of the living body captured in the image, and determines that the living body is a true living body based on the temporal change in the brightness value and the texture feature A biometric authentication apparatus having a determination unit is provided.

  In order to achieve the above object, in the present invention, an imaging unit that captures a living body and an authentication processing unit that processes an image captured by the imaging unit and performs biometric authentication, the authentication processing unit includes: A plurality of biological determination units with different determination criteria, a feature extraction unit that extracts biological features of a biological from an image, and a matching unit that compares the similarities of biological features And when the plurality of biometric determination units determine that the biometric is a real biometric, the feature extraction unit extracts the biometric feature of the biometric and the collation unit extracts the biometric feature extracted and the registration data A biometric authentication apparatus configured to compare similarities is provided.

  Furthermore, in order to achieve the above object, the present invention includes an imaging unit that images a living body and an authentication processing unit that processes an image captured by the imaging unit and performs biometric authentication. A feature extraction unit that extracts a biometric feature of a living body from an image, and the extracted biometric feature is divided into a position correction feature amount and a collation feature amount composed of mutually independent biometric feature amounts. Provided is a biometric authentication device configured to include a collation unit that obtains a deviation amount and compares the similarity between a biometric feature and registered data using a collation feature amount using the positional deviation amount.

  ADVANTAGE OF THE INVENTION According to this invention, it can detect that the biological body input into a biometrics authentication apparatus is an actual biological body.

1 is a diagram illustrating an overall configuration of a biometric authentication system according to Embodiment 1. FIG. It is a figure explaining the apparatus structure of the biometrics authentication system based on Example 1. FIG. It is a figure which shows an example of the processing flow of a pulse detection based on Example 1. FIG. It is explanatory drawing of one method which detects a fingertip area | region based on Example 1. FIG. It is explanatory drawing of one method which detects a pulse from a pulsation signal based on Example 1. FIG. It is a figure which shows an example of the schematic processing flow which performs foreign material determination based on Example 1. FIG. It is explanatory drawing of one method which extracts the local texture characteristic of the image based on Example 1. FIG. It is explanatory drawing of one method based on Example 1 which acquires the histogram of a texture feature from the local texture feature of an image. It is a figure which shows an example of the processing flow of the learning of machine learning and the determination for performing a foreign material determination from the histogram of the texture characteristic based on Example 1. FIG. It is explanatory drawing of one method based on Example 1 which extracts the local color feature-value for classifying a finger area | region and a background area | region. It is explanatory drawing of the authentication system using the registration data disclosed by the server based on Example 2. FIG. It is a figure which shows an example of the processing flow of the authentication process using the biometric feature which can be disclosed safely based on Example 2. FIG. It is a figure which shows an example of the apparatus structure of the finger vein authentication apparatus based on infrared transmission imaging | photography based on Example 3. FIG.

  Various embodiments which are modes for carrying out the present invention will be described with reference to the drawings. In this specification, the biometric features are anatomically different biometric features such as finger veins, fingerprints, joint patterns, skin patterns, finger outlines, fat prints, finger length ratios, finger widths, etc. Means.

  Example 1 is an example of a biometric authentication system using a finger blood vessel. That is, the image processing apparatus includes an imaging unit that captures a living body and an authentication processing unit that processes an image captured by the imaging unit and performs biometric authentication, and the authentication processing unit obtains a temporal change in the luminance value of the living body that appears in the image. Implementation of a biometric authentication system having a biometric determination unit that extracts a texture feature of a living body photographed in an image and determines that the living body is a true living body based on a temporal change in luminance value and a texture feature It is an example.

  FIG. 1 is a diagram illustrating an example of the overall configuration of a biometric authentication system using a finger blood vessel according to the present embodiment. Needless to say, the configuration of the present embodiment may be configured not as a system but as a device in which all or a part of the configuration is mounted on a housing. The apparatus may be a personal authentication apparatus including an authentication process, or the authentication process may be performed outside the apparatus, and may be a blood vessel image acquisition apparatus or a blood vessel image extraction apparatus specialized for acquiring a blood vessel image. Further, this configuration may be regarded as a blood vessel image acquisition device using a general-purpose color camera mounted on a smartphone or tablet. Moreover, the embodiment as a terminal may be used as described later. A configuration including at least an imaging unit that captures a biometric image and an authentication processing unit that processes the captured image and performs biometric authentication is referred to as a biometric authentication device.

  The biometric authentication system of the present embodiment illustrated in FIG. 1 includes an input device 2 that is an imaging unit, an authentication processing unit 10, a storage device 14, a display unit 15, an input unit 16, a speaker 17, and an image input unit 18. The input device 2 includes a light source 3 installed in the housing and an imaging device 9 installed in the housing. The authentication processing unit 10 has an image processing function.

The light source 3 is a light emitting element such as an LED (Light Emitting Diode), for example, and irradiates light on the finger 1 presented on the upper part of the input device 2. The light source 3 may be capable of irradiating various wavelengths depending on the embodiment, or may be installed on the upper part of the finger 1 to irradiate the transmitted light of the living body. Moreover, it is good also as a structure by which the light source 3 is not mounted. The imaging device 9 captures an image of the finger 1 presented on the input device 2. The imaging device 9 may be a color camera or an infrared camera. A distance camera that can measure the distance of the subject may be separately installed. Furthermore, a plurality of fingers 1 may be used.
The image input unit 18 acquires an image captured by the imaging device 9 of the input device 2 and inputs the acquired image to the authentication processing unit 10.

  The authentication processing unit 10 includes a central processing unit (CPU) 11, a memory 12, and various interfaces (IF) 13. The CPU 11 extracts biometric features during image processing on an input image by executing a program stored in the memory 12, biometric determination for determining that the living body is a true living body, registration processing, and authentication processing. Various processing such as feature extraction to be performed and collation for comparing similarities of biometric features. These various processes will be described in detail later. The function memory 12 stores a program executed by the CPU 11. The memory 12 temporarily stores an image input from the image input unit 18.

  The interface 13 connects the authentication processing unit 10 and an external device. Specifically, the interface 13 is connected to the input device 2, the storage device 14, the display unit 15, the input unit 16, the speaker 17, the image input unit 18, and the like.

  The storage device 14 stores user registration data and the like. The registration data is information for collating users obtained during the registration process, and is image data such as a finger vein pattern, for example. Usually, a finger vein pattern image is an image obtained by capturing a finger vein, which is a blood vessel mainly distributed under the skin on the palm side of a finger, as a dark shadow pattern.

  The display unit 15 is, for example, a liquid crystal display, and is an output device that displays information received from the authentication processing unit 10. The input unit 16 is, for example, a keyboard, and transmits information input from the user to the authentication processing unit 10. The speaker 17 is an output device that transmits information received from the authentication processing unit 10 as an acoustic signal such as voice.

  FIG. 2 is a diagram showing an example of a schematic flow of biometric authentication registration processing and authentication processing using a finger blood vessel described in the present embodiment. The registration process and the authentication process are realized by a program executed by the CPU 11 of the authentication processing unit 10 described above, for example.

  First, the flow of registration processing in the upper part of FIG. 2 will be described. First, the authentication processing unit 10 displays a guide for prompting the user to present a finger on the display unit 15, and the user holds the finger according to the guide (S201). If the smart device is used as an authentication device, the user holds his / her finger in the air, and if the device is equipped with a finger rest, the user places the finger rest on the finger rest.

  Next, the authentication process part 10 performs the biometric determination process which determines with a biometrics determination part that a biological body is a true biological body. First, pulse detection is performed on a finger that is a living body (S202). The flow rate of blood flowing inside the fingers varies with the heartbeat. Hemoglobin, which is a tissue in blood, has the property of absorbing light, and when the blood flow varies, the light reflectance and absorption rate of the entire finger changes. Therefore, when a finger that is a living body is photographed, the luminance value of the infrared image or the luminance value of each color of the color image changes in synchronization with the pulse, and the magnitude of the change varies depending on each color. On the other hand, when a foreign object or a fake finger is held over, no change in the luminance value associated with the pulse is observed. That is, it is possible to determine whether the presented subject is a real finger or a foreign object based on whether or not a pulse change can be detected based on the temporal change in the luminance value of the living body.

  Next, as a continuation of the biometric determination process, the biometric determination unit determines a real finger (real finger), a foreign object, and a forged finger using the texture characteristics of the subject (S203). A real living body such as a real finger and a non-living foreign body, a forged living body, a living body photographed with a moving image, and the like have different characteristics of captured images, particularly fine texture features. Therefore, a large number of fingers taken with real fingers, foreign objects, forged living bodies and videos are pre-photographed, and machine learning is used to learn what kind of differences are in the texture characteristics of the images. It is determined whether or not it is a real living body. That is, the texture feature of the photographed living body is extracted, and it is determined based on this texture feature that the living body is a true living body.

  At the end of the biometric determination process, a determination process that combines the above pulse detection and foreign object determination based on texture features is performed (S204). Here, when it is determined that both the pulse detection and the result of the foreign matter determination based on the texture feature are real living bodies, the input subject is determined to be a real living body. When it is determined that the input biological body is a true biological body, the authentication processing unit 10 performs a feature extraction process for extracting a biological feature from the captured biological image (S205), and the acquired biological feature is displayed. Registration data is registered in the database (S206). If not, registration fails and the process ends (S207).

  Next, the flow of authentication processing in the lower part of FIG. 2 will be described. First, the authentication processing unit 10 displays a guide for prompting the user to present a finger on the display unit 15, and the user holds the finger according to the guide (S211). Next, as a living body determination process for determining that the living body is a true living body, determination of a real finger and a foreign object / fake finger based on the texture characteristics of the image is performed (S212). Then, from this result, it is determined whether or not the finger is a real finger (S213). If it is determined to be a real finger, a feature extraction process for extracting a biometric feature is performed (S214). The verification process with the registered data is performed (S215), and finally the authentication result is determined (S216). If the actual finger determination is unsuccessful, the authentication process is terminated as an authentication failure (S217).

  As described above, in the authentication process, unlike the registration process, pulse detection is not performed. This is because, as will be described later, in pulse detection, it is necessary to measure a change in luminance value for a certain period of time, for example, about 5 seconds, and in general, it takes a long calculation time. Therefore, in the authentication process of the biometric authentication system of the present embodiment, forgery resistance is not reduced by performing only the foreign object determination based on the texture characteristic of the subject as the biometric determination for determining that the living body is a true living body. An effect of shortening the authentication processing time can be obtained.

  According to the biometric authentication system of the present embodiment, at the time of registration, guide display for prompting a finger, pulse detection, and foreign object detection based on texture features are performed, so that the user can perform registration work himself / herself. At the same time, unauthorized registration can be rejected. As a result, the registrant can perform the registration process without supervising and supporting the registration work as an operator, thereby improving the operability of the authentication system.

  FIG. 3 is an example of a processing flow for detecting a pulse of a living body held over in the pulse detection (S202) during the registration process of FIG. As described above, one of the necessary conditions for determining whether or not the subject is a true living body is that the pulse can be detected correctly in the registration process of the present embodiment. A processing procedure for detecting a pulse will be described with reference to FIG.

  First, a finger area for a held finger is detected (S301). Although there are various methods for detecting a finger region, in this embodiment, in a method that will be described later, attention is paid to each pixel in the image, and a luminance difference from the surrounding pixels is calculated for each color to obtain a feature amount. Based on this, it is determined whether each pixel in the image is a finger region or a background. In addition, the boundary between the finger and the background is acquired as the contour line of the finger region.

  Next, the fingertip region for finding the fingertip region portion in the detected finger region is detected (S302). The change in luminance due to the pulse is particularly observed at the fingertip portion that is physically at the end, and a signal with a high S / N can be obtained. Therefore, in this embodiment, a fingertip region is detected from the input finger image, and the pulse of the presented finger is detected with higher accuracy by observing the luminance change.

  Here, an embodiment of the fingertip region detection method in S302 will be described. First, as shown to (a) of FIG. 4, the fingertip position of the finger outline 60 is detected. Since the fingertip 61 is a portion where the contour line 60 of the finger bends greatly, the region on the curved direction side is the finger internal region 62, and the curvature of the contour line 60 exceeds a certain value, and A position where the curvature is large, that is, a position where the bending is the most is determined as the fingertip 61.

  Subsequently, as shown in the enlarged view of the fingertip 61 in FIG. 5B, a plurality of straight lines 63 that are orthogonal to the contour line near the fingertip are detected, and the number of straight lines passing through the finger inner region 62 is the largest. Many positions are detected to obtain an initial position 64 at the center of the fingertip region. By using the intersection of a large number of straight lines 63 orthogonal to the contour line, it is possible to stably detect the approximate center position near the finger contour having a high curvature.

  Thereafter, as shown in FIG. 6C, a plurality of distances between each pixel constituting the finger outline 60 closer to the fingertip 61 than the initial position 64 at the center of the fingertip region and the initial position 64 at the center of the fingertip region are set. Calculation is performed to calculate the variance of the distance from each point, and the center 65 of the fingertip region having the smallest variance is detected. This method can be implemented by repeatedly searching by the steepest descent method based on the amount of change in dispersion when the center initial position 64 is slightly moved in the horizontal and vertical directions. Thereby, the center 65 of the fingertip region can be set at a substantially uniform position from the contour line near the finger contour. Subsequently, a circle having the maximum radius that does not protrude from the finger area is obtained with the center 65 of the acquired fingertip area as a center point, and this is defined as the fingertip area 66. The fingertip region can be obtained by the above processing.

  As another determination method of the fingertip region, after the fingertip is detected by the above-described method, the center axis of the finger is acquired so as to pass through the fingertip, and the finger located between the fingertip and a certain distance on the finger central axis is obtained. An area may be acquired as a fingertip area, or a finger joint pattern may be detected, and an area surrounded by the finger joint pattern from the fingertip may be defined as a fingertip area. According to this method, since a plurality of candidates are likely to appear in the result when the position where the contour line is most bent is detected as a fingertip, stable detection is difficult.

  On the other hand, the method for obtaining the center of the fingertip area of the above-described embodiment determines one point of the contour line because the fingertip area is determined based on not only the fingertip position detection result but also information on a number of orthogonal lines. There is an advantage that detection with high stability is possible as compared with fingertip detection for obtaining. In addition, by defining the above circular region, a region having the same position and the same area can always be detected even if a temporal variation occurs in the magnification rate of the finger.

  Next, in the flow of FIG. 3, a pulsation signal of the fingertip region is detected based on the detected luminance value inside the fingertip region (S303). Here, the method for detecting the pulsation signal of this embodiment will be described. Here, an embodiment in the case where the imaging device 9 is a color camera will be described. First, for the luminance values of the three colors of red (R) / green (G) / blue (B), the average value inside the fingertip region obtained above is obtained. Next, the values of R−B and R−G are acquired as the difference between the respective average values and the average value of each color. Finally, a result obtained by summing the average luminance of R and the difference signals of RB and RG is obtained. This technique uses the fact that the amount of change in luminance value when the blood flow varies varies depending on the color, and that image noise components due to ambient light tend to be affected simultaneously in each color.

  That is, when the difference signal is calculated, a pulsation component is left, but since the noise component is removed by the difference, only the pulsation component can be acquired. For example, since the amount of change in luminance of red (R) accompanying the change in blood volume is larger than the amount of change in blue (B), pulsation is also seen in the difference signal of RB. At this time, for example, if noise similar to red (R) and blue (B) is applied simultaneously due to the influence of external light, only the noise component is canceled by the difference calculation. In this way, it is possible to acquire a pulsation signal resistant to noise by using the differential signal.

  Note that when the finger is moving, the detection of the pulsation signal described above becomes unstable. Therefore, guidance is provided to stop the finger, or pulse detection itself is skipped while the finger is moving. Also good. That is, the recognition processing unit 10 detects finger movement using the output signal of the imaging device 9 such as a color camera, and controls to not start pulse detection or interrupt pulse detection while the finger is moving. be able to.

  Further, when it can be assumed that the finger is completely stationary, the pulsation signal may be acquired as follows. First, for each pixel in the fingertip region, a difference in luminance value between pixels at the same position in two temporally continuous frames is calculated. Next, it is determined whether the luminance difference of each pixel is positive or negative. Next, the difference between the number of pixels determined to be positive and the number of pixels determined to be negative is divided by the total number of pixels that have changed positively and negatively, and this is obtained as a pulsation signal. According to this method, the number of pixels that have changed from light to dark or simultaneously from dark to light is extracted as a signal, so when there is no change in blood flow, the change to positive and negative is affected by image noise. Since the number of generated pixels is almost the same, the pulsation signal changes near zero, but when blood flows instantaneously, the luminance difference of many pixels changes to a positive value. This has the effect of increasing the S / N of the pulsation signal.

  Next, in the flow of FIG. 3, the periodicity of the detected pulsation signal is determined (S304). As shown in FIG. 5, when an actual finger is presented, a peak appears periodically in the pulsation signal 50 indicating the pulse signal intensity. In order to detect this, first, a time derivative 51 is calculated for the pulsation signal 50. If it is defined that the pulsation signal increases in the positive direction when the blood volume increases, the time differential value increases in the positive direction in the portion where the inflow of blood flow occurs, and in the negative direction in the outflow state of the blood flow It changes to zero and becomes zero in a portion where there is no change in blood flow. Here, it is determined that the value of the time derivative 51 has exceeded a certain threshold value, and is set as pulsation. If the threshold is not exceeded, it is considered as noise and the period is not counted.

  In the example of FIG. 5, a slight pulsation signal is also seen between pulsations # 5 and # 6, but the time differentiation 51 does not exceed the threshold value and is not counted as a pulsation. Finally, this is carried out in time series to obtain the time interval between the time when the previous periodic peak was detected and the current peak detection, and this is taken as the current pulsation signal cycle.

  Finally, it is determined whether the pulsation signal is like a pulse (S305). First, the validity of the period of the pulsation signal is determined. Here, it is determined whether the period of the currently detected pulsation signal is within a preset time range, and if it is within the range, the count value of the number of pulsations is incremented by one. It is determined that the pulsation is not caused and the count value of the number of pulsations is reduced by one. However, the initial value of this count value is assumed to be zero. The count value is shown at the bottom of FIG. In addition, the range of the preset time regarding the period of the pulsation signal described above is a normal range of a normal human pulse, for example, when a range of 40 to 120 times per minute is defined as normal, the time range. Can be determined to be between 0.5 seconds and 1.5 seconds.

  Next, the number of pulsations is continuously measured to determine whether the pulse is present. When the number of pulsations is continuously counted and the value becomes, for example, 5 times, this pulsation signal can be regarded as a highly reliable pulsation due to the living body. In the example of FIG. 5, the cycle deviates from the normal range on the way, so the count value is decreased at the corresponding location. Subsequently, it is determined whether or not the count value of the pulsation exceeds a certain threshold value, that is, whether or not the pulse rate is high (S306). If so, it is determined that the pulse detection is successful (S307). On the other hand, if the pulse detection is not successful within a certain time, that is, if a preset time-out time has passed, the pulse detection fails (S308). The periodicity of the peak value of the pulsation signal may be detected based on frequency analysis such as FFT (First Fourier Transform) or MUSIC method (Multiple Signal Classification).

  Also, in order to prevent blinks such as outside light from being mistakenly determined as pulses, for example, an R / G / B color image is regarded as three images, and independently by a method such as Independent Component Analysis (ICA). It is also possible to acquire a pulsation signal from the changed signal component.

  FIG. 6 shows an example of a processing flow for determining a real finger, a foreign object, and a forged finger based on the texture feature of the image described in S203 and S212 of FIG. Even if the pulse can be observed by the registration process of the above-described method, the photographed image is not necessarily a real living body. Therefore, in the registration process of the configuration of the present embodiment, in order to further improve the accuracy of foreign object detection, a process for determining whether the finger is a real finger, a foreign object, or a fake finger is used together with the texture characteristics of the image.

  First, as a texture feature of an image, an LBP feature amount in a finger inner region obtained by the above-described method is obtained (S601). The region from which the LBP feature value is extracted may be the entire finger inner region, or may be limited to the region from which the authentication feature value is extracted in order to speed up and stabilize the processing. In this embodiment, a texture feature quantity based on LBP (Local Binary Pattern) generally known as a texture feature is used. The LBP is information obtained by coding the magnitude of the luminance value of the pixel of interest and the neighboring pixels.

  As shown in FIG. 7A, the most basic LBP acquisition method is to encode the magnitude of the luminance value of the pixel of interest and its eight neighboring pixels 120. Specifically, the following method is used. To win. First, the luminance values of the pixel of interest and the neighboring pixel at the upper left thereof are compared, and 0 is given if the luminance value of the pixel of interest is larger, and 1 is given if the luminance value is the same or smaller. Next, the target pixel is compared with the neighboring pixel immediately above it, and 0 or 1 is similarly given. Next, the pixel of interest is compared with the neighboring pixel on the upper right side, and 0 or 1 is similarly given. Thus, when comparing with the target pixel in the clockwise direction starting from the upper left neighboring pixel, eight values of 0 or 1 are obtained. If a code of 0 or 1 is obtained and arranged from the upper bit to the lower bit in the order obtained, an 8-bit numerical value is obtained for one pixel of interest. This is the basic LBP code 121 for the pixel of interest.

In the example of FIG. 7A, a value of (01010011) ₂ in binary notation and 83 in decimal notation are obtained. When the LBP value is acquired, it is possible to roughly estimate what kind of texture is in the vicinity of the target pixel. For example, if LBP is 0, i.e., consisting of binary representation and (00000000) _2, the pixel of interest represents the higher luminance value than its 8 neighboring all pixels, i.e. bright bright spot to the position of the pixel of interest is observed Means that Similarly, a maximum of 256 patterns of texture information can be expressed when the luminance value decreases from left to right or when the luminance value decreases from diagonally upper right to diagonally lower left.

  Since the basic LBP described above is coded clockwise from the upper left pixel, different results are obtained when the image is rotated. In order to solve this, the above-described LBP can be expanded to obtain a rotation invariant LBP. The rotation-invariant LBP is obtained by performing 8 types of clockwise coding from all the pixels in the vicinity of 8 and obtaining an LBP code that has the minimum value. The upper and lower pixel values 120 illustrated in FIG. 7B are obtained by rotating the lower image 90 degrees clockwise with respect to the upper image. Since the basic LBP is coded in order from the upper left position, the result for the rotated image is different in the upper and lower stages, but on the other hand, if coding is performed from the position where the coded result is the smallest, The same code. This is the rotation invariant LBP 122.

  However, even if the rotation invariant LBP is used, the encoding result differs if the magnification of the subject is different. Therefore, as shown in FIG. 7 (c), not only the comparison with eight neighboring pixels (distance 1, pixel indicated by P1 in the figure) with respect to the target pixel, but also eight points on the circumference at positions separated by distance 2 ( Comparison is also made for P2) in the figure and 8 points (not shown) on the circumference at a distance of 4 to obtain three 8-bit codes. The LBP obtained in this way is referred to as an enlarged rotation invariant LBP here. In the present embodiment, it is assumed that this enlarged rotation invariant LBP is acquired and used for foreign object determination. Note that the distance value and the number of codes to be acquired are 8 points (P3 in the figure) on the circumference at a position separated by a distance 3 that was not used in the above, and the discrimination accuracy is maximized. It can be arbitrarily set such as setting or setting the processing speed to be high.

  Next, in the flow of FIG. 6, an LBP histogram is acquired (S602). The enlarged rotation invariant LBP extracted as described above is calculated for one pixel of interest, but since the finger region portion shown in the image is composed of a large number of pixels, a large number of LBPs are obtained from one finger image. Can be extracted. A highly reliable feature quantity can be obtained by using a plurality of LBPs for determination. In this embodiment, a normalized histogram of LBP is created as one technique. Normalized histogram is a frequency distribution that holds how many times the number of pixels calculated for the acquired LBP value is calculated, and since the number of pixels for calculating LBP differs depending on the area of the finger inner region, When the number of pixels is normalized to 1, it is converted into a value indicating how much ratio of each LBP appears in the whole.

Since the expanded rotation invariant LBP in the present embodiment is composed of three 8-bit codes, there are 256 possible values for each code, from 0 to 255. As shown in FIG. 8, the histogram for the value of each code is 256. It becomes a dimension. Here, since LBP ₁ , LBP ₂ , and LBP ₄ are obtained for three distances from the target pixel, a 256-dimensional histogram can be obtained for each of the three distances. Here, the histogram of the enlarged rotation invariant LBP acquired from the distance _i is represented as HLBP _i . In this embodiment, the histograms of HLBP ₁ , HLBP _2, and HLBP ₄ are arranged in order to form a 768-dimensional histogram. In FIG. 8, this LBP histogram is shown as HLBP _1,2,4 .

  At the end of the flow of FIG. 6, real finger / foreign / fake finger determination is performed from the above-described histogram (S603). In the distribution of the acquired histogram, characteristics such as a high frequency of a specific dimension can be seen according to the presence or absence of fine sesame salt noise or smooth luminance change. That is, it is possible to determine whether the object is a foreign object or an actual finger from the histogram. An example of a method for determining whether a finger is a real finger or a foreign object / fake finger from a histogram feature is to collect a large number of real finger images and a foreign object / fake finger image, and use the histogram feature value to determine whether the finger is a real finger. In order to make a precise determination, there is a method of obtaining a determination parameter by machine learning and using the obtained determination parameter. In the present embodiment, a method based on a random forest that is one of machine learning will be described as an example.

  FIG. 9 is an example of a learning process flow for performing parameter learning based on an enlarged rotation-invariant LBP and a random forest, and a determination process flow. First, the learning process flow shown in the upper part of FIG. 9 will be described. First, an image database in which various foreign objects are photographed and an image database in which many real fingers are photographed are prepared (S901). At this time, a label for identifying whether the subject is a foreign object or a real finger is attached to all images.

  Next, a finger region is detected from all the prepared image data, and a histogram feature amount that is the above-described texture feature in the finger region is extracted (S902). In other words, as described above, the enlarged rotation invariant LBP is extracted for the pixel determined to be the finger region, and the information of all the pixels is converted into a 768-dimensional histogram.

  Next, the texture histogram feature amount extracted from the image is input to the random forest as learning data for learning (S903). The random forest is a classifier for classifying an input into a specific class by a plurality of decision trees, and obtains a final classification result by a majority vote or a weighted evaluation value for the classification result obtained by each decision tree. In each decision tree, the input data is distributed to either the left or right node of the binary tree according to the magnitude of the value of a specific dimension among the feature values extracted from the image and moved to its child nodes. Depending on the size of the value of a specific dimension, it is divided into left and right, and the classification result when it finally reaches the leaf node is regarded as the classification result of that input. At this time, for each node of the tree, the correctness rate of the classification result is the highest for which dimension is selected and which threshold value is assigned to the left and right according to the value of that dimension. Need to be determined. In the random forest learning phase, these parameters are determined by random numbers, but this is a general technique and will not be described in detail.

  Thereafter, in order to sufficiently converge the learning result, learning is repeatedly performed, and it is determined whether the number of times satisfies a predetermined number (S904). When the learning has been repeated a predetermined number of times, the parameter obtained as the learning result is determined as the final result of the random forest for the determination process (S905), and the learning process is terminated.

  Next, the determination process flow shown in the lower part of FIG. 9 will be described. First, a finger region is detected from the input unknown image, and an enlarged rotation-invariant LBP histogram, which is a histogram feature amount of the texture in the finger region, is extracted (S911). Subsequently, the histogram feature quantity of the texture of the unknown input image is input to each decision tree of the determination random forest (S912), and the determination result is obtained (S913). If the texture feature is similar to a real finger, it is determined to be a real finger, and if it is similar to a foreign object, it is determined to be a foreign object. As described above, it is possible to determine whether the finger is a real finger based on the texture feature.

  Further, in the real finger determination using the random forest of the present embodiment, it is possible to realize a more accurate real finger determination by configuring the random forest in multiple stages and performing the foreign object determination. That is, the biometric authentication system according to the present embodiment includes an imaging unit that captures a living body and an authentication processing unit that processes an image captured by the imaging unit and performs biometric authentication. A plurality of biological determination units having different determination criteria, a feature extraction unit that extracts biological features of the biological from the image, and a matching unit that compares the similarities of the biological features. When the plurality of biological determination units determine that the biological body is a true biological body, the feature extraction unit extracts the biological feature of the biological body, and compares the similarity between the extracted biological feature and the registered data. And authenticate.

  As described with reference to FIG. 2, in the authentication process according to the present embodiment, pulse detection is not performed, and only foreign object / counterfeit detection based on texture features is performed. For this reason, there is a possibility that the frequency of erroneously inputting a foreign object or a fake finger that should have been correctly discarded if there is pulse detection may be increased. Therefore, in the real finger determination in the authentication process, a biometric determination unit that performs determination by acquiring a determination parameter by machine learning from a histogram feature amount of a texture feature uses a random forest connected in multiple stages to machine learning. If a pulse is detected, learning is performed specifically for a data set including foreign objects and forged fingers that can be correctly discarded.

  As a result, the random forest outputs a determination result close to pulse detection, that is, one stage of the random forest, which is a multi-stage biometric determination unit, functions as an alternative to pulse detection. Accordingly, foreign object acceptance errors can be reduced even in the authentication process in which pulse detection is not performed. Furthermore, by constructing a random forest that specializes in rejecting foreign objects that cannot be rejected correctly by pulse detection, foreign objects that could not be rejected by pulse detection can be correctly rejected in this random forest. It is also possible to further improve the overall foreign matter detection accuracy. Hereinafter, a specific example of a configuration in which the random forest that is the biometric determination unit is configured in multiple stages such as three stages will be described.

  First, a first-stage random forest is learned using an image database of many foreign objects that tend to be determined as real fingers by pulse detection and an image database of many real fingers. In the present specification, this is called a pulsating foreign object check random forest. Since this random forest for checking pulsating foreign matter can discriminate between a foreign matter that tends to be erroneously determined to be a real finger in pulse detection and a real finger with high accuracy, it is possible to complement a pulse detection error. Next, a second-stage random forest is learned using a general many foreign object image databases and many real finger image databases. In the present specification, this is usually called a random forest for foreign matter check. This normal foreign matter checking random forest can discriminate widely general foreign matter and real fingers with average accuracy.

  Finally, a third-stage random forest is learned using a large number of finger image databases and a large number of real finger image databases photographed as moving images. Here, this is called a moving image finger check random forest. This moving finger check random forest can be rejected with high accuracy against an attack in which an image of a finger taken with a moving image is presented. In this way, a three-stage random forest is created.

  In the determination process, first, an LBP histogram in the finger region of the input subject is acquired, and it is first determined whether or not the object is a foreign object by the first pulsating foreign object check random forest. If it is determined that the finger is a real finger, it is next determined whether or not it is a foreign object by the second-stage normal foreign object check random forest. Similarly, when it is determined that the finger is a real finger, it is finally determined whether or not it is a foreign object by the third-stage moving image finger checking random forest. Here, if it is determined as a real finger, it is determined as a real finger in all the determinations of the multistage random forest, and it is truly regarded as a real finger. If any random forest is determined to be a foreign object, the input is regarded as a foreign object, and the authentication process fails.

  As described above, when the original foreign matter and counterfeit features can be grasped, multiple random forests are constructed from multiple learning data to achieve integrated and highly accurate foreign matter determination specialized for foreign matter features. it can.

  In the above-described embodiment, the random forest is artificially formed in multiple stages. However, the boosting technique represented by AdaBoost, which is known as a machine learning technique, is combined with a random forest to automatically perform multistage learning. Also good. Specifically, learning is performed in the first-stage random forest so that the classification accuracy of a large number of real fingers and foreign object image data is improved. As a result, in order to correctly classify the actual fingers and foreign objects in error, the next random forest learns them mainly. This is repeated in multiple stages, and finally the determination is made by weighting the results of all random forests. As a result, a large number of random forests function in a complementary manner, and the overall error rate can be reduced.

  In the above-described embodiment, LBP is selected as an image feature amount. However, feature amount extraction and foreign matter determination may be performed by machine learning. For example, a foreign object can be determined by a convolutional neural network (CNN) generally known as deep learning. Even in this case, as described above, while convolving a plurality of filters with respect to the inside of the finger region, the filter value and the neural network coupling coefficient are repeatedly learned so as to increase the accuracy rate. Thereby, a foreign object and a real finger can be distinguished with high accuracy.

  Further, in the determination of the real finger, the foreign object, and the fake finger based on the texture feature, it is possible to determine whether or not the living body is a print on paper or a video image by the following method. First, a finger region is detected by an arbitrary method, and then the authentication processing unit 10 controls the reflected light source attached to the biometric authentication device to blink. At this time, if an actual living body is presented, only the finger region blinks and the background region does not blink. However, when printing on paper or presenting a moving image, since the paper or screen is flat, the finger area and the background area blink simultaneously. By blinking the light source using this texture feature, it is possible to determine that the moving image is presented. If the authentication device is a smart device, an attached flash light or a liquid crystal display may be blinked instead of the reflected light source.

  Further, if the irradiation direction of the light source can be controlled, it is also possible to irradiate the subject sequentially from a plurality of directions and detect a change in shadow at that time as a texture feature. If there is no change in shadow, it is considered a foreign object. Further, if the color of the light source can be controlled, it may be determined whether or not the finger is unique by determining whether or not the reflectance for each color is unique to the finger. Further, the authentication processing unit 10 may request a specific finger bending / extension or finger gesture before the registration and authentication operations using the speaker 17 or the like, and may determine that the object is a foreign object if it cannot be performed.

  In the above-described pulse detection process and foreign object detection process using texture features, it is necessary to accurately separate the internal area and the external area of the finger. Furthermore, the detection of the finger region is an important process for performing the authentication process with high accuracy. In particular, a living body photographed with a general-purpose color camera of a smart device needs a highly accurate background separation technique because various subjects are reflected in the background. Here, as an example, a method for detecting a finger region using a feature amount obtained by calculating a luminance difference between each pixel and its surrounding pixels for each color and a random forest will be described.

  First, with respect to 8 neighborhoods of a distance D pixel from a certain pixel of interest, a red plane neighborhood pixel 181 and a green plane neighborhood for three colors of a red plane interest pixel 180, a plane attention pixel 182 and a blue plane attention pixel 184 The difference is obtained by a combination of the three colors of the pixel 183 and the neighboring pixel 185 of the blue plane. As a result, 9 × 8 = 72 dimensional features are obtained. P0r, P0g, and P0b shown in FIG. 10 are pixels of interest of three colors R, G, and B, respectively. Similarly, PDr, PDg, and PDb are neighboring pixels of a distance D of three colors of R, G, and B, respectively. It is. As an example of the color difference, for example, the difference between the target pixel 180 on the red plane and the neighboring pixel 183 on the green plane is P0r-PDg, and since there are eight PDg, eight P0r-PDg are obtained. Thus, a value of 72 is obtained.

  Furthermore, by changing the distance D to N ways, it becomes a 72N-dimensional feature. If the luminance values of the three colors R / G / B of the target pixel are included as a three-dimensional feature, 72N + three-dimensional features are obtained for each pixel. Similar to the method of analyzing texture features in the random forest described above, prepare many finger images labeled with correct answers between the finger area and the background area, and correctly correct the background pixels and the pixels without the finger area. Learn random forest parameters so that you can. As a result, when an arbitrary finger image is input, the above-described color feature is obtained for each pixel, and if it is determined for each pixel whether it is the background or the finger region, only the portion of the finger region can be extracted. Although an unnatural finger region portion may appear due to a determination error, it may be processed into a smooth region by expansion / contraction processing in morphology. Further, it is possible to speed up the processing by performing learning and determination of the finger region in a state where the original image is reduced to a small size and enlarging the final result to the original size.

  Further, as labeling of the learning data, a label may be assigned to each kind of finger such as the index finger and the middle finger, and joint parts such as the base clause, the middle clause, and the last clause, and each may be determined in detail. Furthermore, learning data is prepared for each pose of various fingers such as Janken Goo, Choki, and Par, and as a result, the posture of the finger can be learned. In this way, finger gesture determination can be realized, for example, to detect foreign objects or counterfeiting, or to use as authentication information to improve authentication accuracy, or to narrow down user data from a large amount of registration data By utilizing it, it is possible to speed up the matching process.

  An embodiment of an authentication system using the above-described finger gesture determination will be described below. For example, when using various Internet banking services using a smartphone, in the past, after logging in using a personal identification number or one-time password in addition to a personal ID such as an account number or contractor number, you can use `` Balance inquiry '' or `` Operations such as pressing buttons indicating various services such as “transfer” are complicated until the target service is used. On the other hand, when finger gesture determination and finger-based biometric authentication are combined, for example, when “Balance Inquiry” is performed, the “Janken” pose with a finger is used. If the observed gesture is choki, use the shape of the gesture and the biometric features of the finger to log in and change mode to “Balance inquiry”. Perform simultaneously and automatically. Similarly, when the user presents a par, login and mode transition to “transfer” are performed.

  As described above, since the configuration of the present embodiment enables the switching of various functions of various business systems and the identity verification at the same time, the identity verification can be surely performed without performing complicated operations as in the past. It is possible to use various services after being performed. Note that as the number of fingers held up increases, more biological features can be observed.For services that require stronger security, assigning more gestures to hold up the fingers to increase the security strength. It becomes possible to raise.

  In addition, the authentication system randomly determines a finger gesture, and requests the user to hold the finger according to the instruction, thereby performing foreign object / counterfeit check of the held finger. For example, a general foreign object / counterfeit organism is known to hold daily items or a living body printed on paper, but when such a foreign object or counterfeit finger is held over, it is specified on the spot. It is not easy to imitate finger gestures. Therefore, the act of performing the designated gesture itself becomes evidence that the hand of the living body is held up, and the possibility of eliminating foreign objects and forgery can be increased. Furthermore, a plurality of gestures may be presented, or the temporal change of the gestures at that time may be confirmed in a moving image. As a result, it is possible to improve the accuracy of foreign object / counterfeiting determination against an attack particularly when a stationary subject is held over.

  Also, in a server authentication apparatus in which a large number of users' biometric information is registered, when performing so-called 1: N authentication in which only biometric information is authenticated without inputting personal IDs, a finger gesture is used. Thus, it is possible to limit the registered data by performing the gesture from a large number of biological databases. Therefore, the authentication process is speeded up and the acceptance rate of others can be improved by reducing the number of verification targets. Moreover, you may further limit registration data by showing a some gesture continuously. In this case, since the biometric information is presented multiple times, the matching rate with the registered data is increased, and the shape and order of the gestures to be presented themselves function as a password. It can be realized only by operation. Therefore, convenience is enhanced and authentication accuracy can be dramatically improved.

  According to the present embodiment, it is possible to provide an authentication system with high operability and high accuracy that can detect that an input living body is an actual living body and can perform registration work by the user himself / herself.

  Example 2 is an example of an authentication system using biometric information that can be disclosed safely. FIG. 11 shows a configuration example of the authentication system according to the second embodiment. When a smartphone 201 owned by the user 200 performs purchase of a ticket and photographing and registration of biometric information, a key already purchased from the smartphone to the authentication server 202 is shown. At the same time, the encrypted biometric registration data 204 is transferred via the network 205. At this time, the registration data 204 is stored in the database 206 in the server.

  At the time of authentication at the event venue, when the user holds the finger 1 over the authentication terminal 207 installed at the venue, encrypted input data 208 generated from the presented biometric information is generated, and the authentication server 202 Forwarded to In the authentication server 202, the encrypted registration data 204 is disclosed, and this is compared with the encrypted input data 208 to perform personal authentication. If it is determined that the authentication has succeeded, a signal indicating the success of the authentication is sent to the authentication terminal 207 at the event venue, and the user can enter the venue by opening the gate door, for example. Note that the authentication terminal 207 may download registration data from the authentication server 202 in advance, and the authentication terminal 207 may perform authentication processing by verification.

  As described above, in the configuration of the present embodiment, registration data needs to be communicated via a network, but data protection is essential in order to prevent data leakage. At this time, it is effective to invalidate threats such as data leakage and tampering by ensuring that it is safe to publish the template. For example, PBI (Public Biometric Infrastructure) technology can be used to secure the template safely. Publish. In general PBI technology, verification processing is performed with encryption, but verification processing in an encrypted state takes more processing time than verification processing in a normal non-encrypted state. In particular, when the matching process is performed by template matching based on a feature quantity that is not position-invariant, the patterns are compared while correcting the positional deviation between the patterns to be compared. Becomes larger and requires a large number of comparisons.

  Therefore, in the collation processing with a preferred configuration of the present embodiment, the registration data is divided into biological position correction information and collation information, and information for which no collation information is estimated is selected from the position correction information. Collation is performed in an encrypted state, and the amount of misalignment obtained from the result is used, and collation information is collated at high speed in an encrypted state, thereby realizing safe and high-speed collation. That is, the biometric authentication system according to the present embodiment includes an imaging unit that captures a living body and an authentication processing unit that processes an image captured by the imaging unit and performs biometric authentication. A feature extraction unit that extracts biometric features and the extracted biometric features are divided into position correction feature amounts and collation feature amounts that are composed of biometric features that are independent of each other. And a collation unit that compares the similarity between the biometric feature and the registered data by using the collation feature amount by using the positional deviation amount.

  FIG. 12 is a diagram illustrating an example of an authentication processing flow according to the second embodiment in which a registration template corresponding to PBI is collated at high speed and with high accuracy. First, guidance for prompting the presentation of a living body is displayed (S1201), and when the user holds a plurality of fingers over the smartphone, the foreign object detection detailed in the above-described embodiment is performed (S1202). In addition, although the pulse detection mentioned above is good to implement here, it is good also as a structure which does not implement a pulse detection for shortening processing time according to the condition to authenticate.

  It is determined whether or not the finger held up by the foreign object detection process is a true living body (S1203). If the finger is not a true living body, guidance for prompting the user to hold the finger is displayed again. Extraction is performed (S1204). In this embodiment, biometric features such as finger veins, fingerprints, and joint patterns are extracted from a plurality of fingers, and all the biometric features of all fingers are converted into binarized or multivalued feature values. The patterns are merged at the level and extracted as information to be collated by template matching. However, any biometric information that realizes PBI can be similarly applied to any feature extraction method and data format.

  Next, the extracted feature amount is decomposed into a linear sum of base feature amounts determined in advance (S1205). Here, the basis feature amount determined in advance is a feature amount in which each basis is independent, and is determined by the following method in this embodiment. First, feature values of a living body are extracted from a large number of hand images prepared in advance, and PCA basis features (PCA basis) are extracted by performing principal component analysis (Principle Component Analysis (PCA)) on these data. To do. At this time, M bases with higher M contribution ratios are acquired.

Assuming that the original biometric feature quantity is f, the PCA base is x _i , and the weight for the i-th PCA base is a _i , the feature quantity f before PCA conversion is expressed as a linear sum as shown in Equation 1 below. Note that x _i is arranged in ascending order of i in descending order of contribution obtained when PCA is performed, and in particular, x ₀ is an average value of a large number of biometric features prepared in advance.

f = a ₀ x ₀ + a ₁ x ₁ + a ₂ x ₂ +… + a _K x _K + a _{K + 1} x _{K + 1} +… + a _M x _M (Formula 1)

Subsequently, for the M PCA bases x _i , M ICA bases s _j and a mixing matrix A are obtained by independent component analysis (ICA). Here, assuming that a matrix in which M ICA bases s _i are arranged is s, and a row vector of the i-th row of the mixing matrix A is A i, the biometric feature is converted as follows.

f = a ₀ (A ₀ s) + a ₁ (A ₁ s) + a ₂ (A ₂ s) +… + a _M (A _M s) (Formula 2)

When the constant term is b _i and the feature quantity f is summarized using the ICA base s _i , the result is as follows. Note that i in the formula is an ICA base number.

f = b ₀ s ₀ + b ₁ s ₁ + b ₂ s ₂ +… + b _i s _i +… + b _M s _M (Equation 3)

Since the ICA basis s _i is a highly independent component selected from a large number of biometric features, the coefficient b _j (j ≠ i) of the other ICA basis is estimated from the coefficient b _i of the i-th ICA basis. Will be difficult.

  Next, the feature quantity f is separated into a position correction feature quantity and a collation feature quantity (S1206). In the present embodiment, an arbitrary K bases among the M bases are combined to form a matching feature amount, and a position correction feature amount is generated by the remaining (M−K) bases. At this time, the value of K ranges from 1 to M-1. That is, the number of bases M, the base number K of collation feature values, and the combination of extracting K bases from the M bases are adjustment parameters.

After that, hand images taken from a large number of subjects (including many images of the same subject) are prepared in advance, feature extraction is performed from these images, and the image is divided into the above-described position correction features and matching features. The process is performed for all the images, and the search results are searched for M, K, and a combination for extracting K bases from M, which result in the highest authentication accuracy. Then, a matching feature value and a position feature value are configured by a combination of ICA bases obtained from the result. When formulated, divided into the original feature f, as shown in the following equations 4 and 5 and the position correction feature value f _p and matching feature amount f _m. The subscripts pi and mi are ICA base numbers in the position correction feature value and the collation feature value, respectively.

f _p = b _p0 s _p0 + b _p1 s _p1 +… + b _{p (MK)} s _{p (MK)} (Formula 4)
f _m = bm ₀ s _m0 + b _m1 s _m1 +… + b _mK s _mK (Formula 5)

As described above, since the position correction feature quantity f _p and the matching feature quantity f _m are composed of independent bases, the matching feature quantity f _m is estimated even if the position correction feature quantity f _p is disclosed. It is not possible. Therefore, the position correction feature value f _p of the input feature value and the registered feature value is collated at high speed in the non-encrypted state to obtain the pattern displacement amount, and the matching feature value is similarly obtained using the result. If collation is performed in an encrypted state, safe and high-speed collation is possible. Even if the position correction feature value is misused, the data can be invalidated by changing the combination of the ICA bases described above to another combination. This means that the ICA base combination itself can be used as a key for data protection. However, since the matching feature value f _m needs to be encrypted, it is encrypted by an arbitrary encryption method and kept safe even if it is disclosed.

  If the ICA base is acquired as a base feature amount in advance by the procedure described above, the base feature amount can be defined.

Here, a method for obtaining the weight bi for expressing the original feature amount f by the linear sum of the ICA bases calculated as described above will be described. First, with respect to Equation 1 above, since the PCA base x _i is orthogonal, the variable a _i can be obtained by taking the inner product of the original feature quantity f and x _i . If we do this for M i, we get all variables ai. Also, since x _i , s _j , and the mixing matrix A are obtained in advance, b _i can be obtained from a _i and A. Therefore, to convert any biometric information f to f _p and f _m, i.e., it is able to obtain the coefficient b _pi and b _mi for a given biological information f.

Next, the position correction feature and the matching feature are converted into a collable format (S1207). For example, if the original biometric information f is vein template image ternary, it is converted into as described above and the position correction feature value f _p and matching feature amount f _m, they are converted to multi-value image by the quantum error The In that case, re-transformation into a ternary pattern is performed again by threshold processing, and correction is performed so that a ternary template image is obtained. Even if it is in another format, convert it to the original format.

  Next, registration and input position correction feature values are collated (S1208). Templates of position correction feature amounts are collated by a conventional method, and a position shift amount is obtained by searching for a position having the highest similarity.

  Subsequently, the matching feature amounts encrypted using the positional deviation amount are collated, and the similarity between both feature amounts is calculated (S1209). Basically, the similarity between both feature quantities is calculated without decoding, but if it is clearly safe, it can be decoded and verified. Finally, whether or not authentication is possible is determined based on whether or not the similarity between both feature quantities exceeds the authentication threshold (S1210). If they are similar beyond the threshold, it is determined that the authentication has succeeded, and otherwise, the authentication has failed, and the authentication process ends.

  Although it is not desirable to use the similarity between position correction features for authentication judgment from the viewpoint of data protection, if it can be determined that there is no practical problem, the similarity between position correction features is also used for authentication judgment. By doing so, the authentication accuracy can be improved.

  By adjusting the number M of bases acquired using PCA and ICA and the number K of bases used for position correction features by the method of the present embodiment described above, the amount of information used for position correction features and matching features Can be finely adjusted. For example, if the biometric feature to be used is not likely to be displaced, it is possible to adjust the balance according to the characteristics of the biometric feature, such as adjusting K to a large value and increasing the amount of information distributed to the matching feature. .

  In general, the simplest method is to extract multiple anatomically different biological features such as veins, fingerprints, and joint patterns from an image and divide them into position correction features and matching features for each biological feature. . However, the problem with this method is that there are at most several variations of biometric features that can be extracted, and there are few biometric feature options used for position correction features. Because it is necessary to divide it into features for position correction and features for collation, there is a problem that if any biological feature is over-spec to be used only for position correction, excess information cannot be used as collation information. is there. Furthermore, there is no guarantee that each biometric feature is necessarily independent information, and a collation feature may be inferred from the position correction feature.

  In response to these problems, according to the configuration of the present embodiment, by acquiring independent components from all the information of a plurality of biometric features based on ICA, position correction features and collations made up of biometric features independent of each other It is possible to adjust the balance by extracting features flexibly, and it is difficult to infer the matching feature from the non-encrypted position correction feature because it is assumed that the information is mathematically independent. As a result, the above-mentioned problems can be solved, and safer and more accurate authentication than before can be realized.

  The above-mentioned multiple biometric features are further increased, and veins, finger joint patterns, fingerprints and skin patterns, finger outlines, fat prints, finger length ratios, finger widths, etc. are extracted and independent from these biometric features. If biometric features with high characteristics can be classified and some of them can be used for registration and the rest can be used as matching features, safety and The accuracy can be improved. Alternatively, ICA may be applied to the three RGB color images to extract independent image features. In this case, anatomically meaningful biological features may not be acquired, but independent components are acquired based on ICA even if independence cannot be guaranteed between multiple anatomical biological features. As a result, independence is ensured mathematically, and safer biometric feature separation can be realized.

  According to the present embodiment, it is possible to detect that an input living body is an actual living body, and it is possible to perform high-speed and high-accuracy collation while keeping biological information confidential, and high data confidentiality and high speed can be achieved. An accurate biometric authentication system can be provided.

  Example 3 is an example of another configuration of a biometric authentication system using a finger blood vessel. FIG. 13 is a diagram illustrating an overall configuration of a system according to the third embodiment. The user presents the finger 1 with the position of the finger aligned on the finger rest 240 provided in the input device 2. The light source 3 installed on the ceiling portion 242 is an infrared light source and irradiates infrared light toward the finger 1. The infrared light transmitted through the finger 1 reaches the imaging device 9 via the opening 241. The imaging device 9 is an infrared camera and captures infrared transmitted light from the finger 1. A finger vein pattern is photographed in this video, and biometric authentication based on the photographed finger vein pattern is performed by the authentication processing unit and the authentication terminal. The ceiling portion 242 can shield external ambient light from entering the imaging device 9.

  In this apparatus configuration, by applying the pulse detection and foreign object detection described in the first and second embodiments, the reliability of the input biological information can be improved. As the processing procedure, the method shown in FIGS. 2 and 3 can be applied.

  In the configuration of the biometric authentication system of the present embodiment, since the ceiling portion 242 is installed on the input device 2, the influence of external light can be reduced, and finger vein pattern imaging is based on infrared transmitted light. Since it is a method, it is possible to capture a clear image, so that it is possible to improve detection accuracy particularly in pulse detection.

  When performing pulse detection, the amount of movement of the finger in the image is measured so that the finger 1 does not move, and guidance is displayed to the user so as not to move the finger if movement is detected. While the finger 1 is stationary, the light amount of the light source 3 is continuously irradiated with a fixed value while the brightness value of the image of the fingertip region is adjusted to a brightness that does not cause whiteout (saturation). As a result, it is possible to more robustly observe the luminance change of the image due to the fluctuation of the blood flow.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. For example, the above-described embodiments have been described in detail for better understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

  Further, the above-described configuration, function, authentication processing unit, and the like have been described as an example of creating a program that realizes a part or all of them. Needless to say, it may be realized by hardware. That is, all or part of the functions of the authentication processing unit may be realized by an integrated circuit such as ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array) instead of the program.

  INDUSTRIAL APPLICABILITY The present invention can provide an authentication system with high operability that allows the user himself to perform registration work and a high-speed and high-accuracy biometric authentication device, and is useful as a personal authentication device.

1 finger 2 input device 3 light source 9 camera 10 authentication processing unit 11 central processing unit 12 memory 13 interface 14 storage device 15 display unit 16 input unit 17 speaker 18 image input unit 50 pulsation signal 51 time differentiation 60 finger outline 61 fingertip 62 finger Inner area 63 Straight line 64 perpendicular to finger contour line Initial position 65 of fingertip area center Fingertip area center 66 Fingertip area 120 Pixel of interest and its eight neighboring pixels 121 Basic LBP code 180 Red pixel of interest 181 Red plane of interest Neighboring pixel 182 Target pixel 183 of the green plane Neighboring pixel 184 of the green plane Target pixel 185 of the blue plane Neighboring pixel 200 of the blue plane 200 User 201 Smartphone 202 Authentication server 204 Encrypted biometric registration data 205 Net line 206 Authentication server Database 207 Authentication end of the venue End 208 Encrypted input data 240 Finger rest 241 Opening 242 Ceiling

Claims (15)

An imaging unit for imaging a living body;
An image processing unit that processes an image captured by the imaging unit and includes an authentication processing unit that performs authentication of the living body,
The authentication processing unit
Obtaining a temporal change in the brightness value of the living body reflected in the image, extracting a texture feature of the living body photographed in the image, and determining whether the living body is true based on the temporal change in the brightness value and the texture feature. A living body determination unit for determining that the living body is
A biometric authentication device. The biometric authentication device according to claim 1,
The temporal change in the brightness value of the living body is a change in the brightness value of the fingertip region of the living body.
A biometric authentication device. The biometric authentication device according to claim 2,
The authentication processing unit
Detecting the center of the fingertip region of the living body based on the number of passages of a plurality of straight lines orthogonal to the outline of the fingertip of the living body;
A biometric authentication device. The biometric authentication device according to claim 2,
The authentication processing unit
A pulsation signal is detected by a change in luminance value of the fingertip region of the living body, and the pulsation signal is time-differentiated to detect a pulse of the living body.
A biometric authentication device. The biometric authentication device according to claim 4,
The authentication processing unit
When the fingertip of the living body is moving, the pulse detection of the living body is not performed.
A biological recognition apparatus characterized by the above. The biometric authentication device according to claim 1,
A light source for irradiating the living body with light;
The authentication processing unit
Controlling the light source to blink when extracting the texture features;
A biometric authentication device. The biometric authentication device according to claim 1,
The authentication processing unit
Detecting the finger gesture of the living body when extracting the texture feature;
A biometric authentication device. The biometric authentication device according to claim 1,
The biological determination unit is
The determination is performed by acquiring a determination parameter by machine learning from the histogram feature amount of the texture feature.
A biometric authentication device. The biometric authentication device according to claim 8,
The biological determination unit is
Using a random forest connected in multiple stages to the machine learning,
A biometric authentication device. The biometric authentication device according to claim 9,
The random forest connected in multiple stages includes a random forest for pulsating foreign matter check and a random forest for normal foreign matter check,
A biometric authentication device. An imaging unit for imaging a living body;
An image processing unit that processes an image captured by the imaging unit and includes an authentication processing unit that performs authentication of the living body,
The authentication processing unit
A plurality of living body determination units having different determination criteria for determining that the living body reflected in the image is a true living body;
A feature extraction unit for extracting the biological features of the living body from the image;
A matching unit for comparing the similarity of the biometric features,
When the plurality of biological determination units determine that the biological is a true biological, the feature extraction unit extracts the biological features of the biological, and the biometric features extracted from the collation unit and the registered data Compare similarities,
A biometric authentication device. The biometric authentication device according to claim 11,
The plurality of biological determination units are
Extracting the texture feature of the living body photographed in the image, and determining that the living organism is a true living body based on the texture feature;
A biometric authentication device. The biometric authentication device according to claim 12,
The plurality of biological determination units are
Obtain a determination parameter by random forest from the histogram feature amount of the texture feature and perform determination,
The random forest includes a random forest for checking pulsating foreign matter and a normal forest for checking foreign matter connected in multiple stages,
A biometric authentication device. An imaging unit for imaging a living body;
An image processing unit that processes an image captured by the imaging unit and includes an authentication processing unit that performs authentication of the living body,
The authentication processing unit
A feature extraction unit for extracting the biological features of the living body from the image;
The extracted biometric feature is divided into a position correction feature amount and a collation feature amount constituted by mutually independent biometric feature amounts, and a positional deviation amount of the living body is obtained by the position correction feature amount, and the positional deviation amount is used. A matching unit that compares the biometric feature and the similarity of registered data by the matching feature amount,
A biometric authentication device. The biometric authentication device according to claim 14,
The authentication processing unit
When the living body determination unit determines that the living body is a true living body, and the living body determination unit determines that the living body is a true living body, the matching unit stores the biological feature and the registration data. Compare similarities,
A biological recognition apparatus characterized by the above.

Images