JPH10261083A - Personal identification device and personal identification method - Google Patents

Abstract

(57) [Summary] [Problem] The calculation of the similarity of the biological information of a human individual is caused by the fact that the biological information changes over time, and the difficulty greatly changes due to noise, changes in the surrounding environment, and the like. It was difficult to obtain an accurate value. With respect to a plurality of elements of personal identification including at least one individual's biometric information, for each element, an input unit for inputting information and converting it to a feature quantity if necessary, a conversion based on the input information or information A feature storage unit that stores the obtained feature amount, a matching unit that performs matching of the input information or the feature amount converted from the information, and calculates a similarity;
Input information or a feature amount converted from the information,
A reliability calculating unit that calculates and outputs the reliability of the information from the information or the feature amount stored in the storage unit, and further includes a plurality of matching units that determine the overall matching result based on the output similarity and reliability. An identification result integration unit having an integration rule for determination.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a personal identification device and a personal identification method for fixing an individual using characteristic information of a human individual.

[0002]

2. Description of the Related Art Heretofore, identification of individuals at financial institutions or government offices has been performed using a registered seal.
Also, in a system using a cash card, usually,
The identification of an individual is performed using a password registered in advance. However, in such a conventional method, at present, accurate identification of an individual cannot be performed due to loss or theft of a seal, forgetting or leaking a personal identification number. Therefore, recently, there has been proposed a method in which individual biometric information such as a voice, a face image, and a fingerprint is registered in advance, and the individual is identified based on the coincidence determination.

FIG. 1 is a block diagram of a system for performing individual identification using voice information which is one of biological information. Information about the voice of the person is stored in the voice recording unit 2 in advance. At the time of identification, the voice information input from the voice input unit 1 and the voice information of the voice recording unit 2 are compared by the voice matching unit 3, and the similarity is calculated. If the similarity exceeds a certain value, the user is identified. There have been proposed a number of techniques for identifying the

FIGS. 2 and 3 are block diagrams of a personal identification system using a face image and a fingerprint. In FIG. 2, the face image information of the person is stored in the face image storage unit 2 in advance. Then, the face image information input from the face image input unit 1 is compared with the face image information in the face image storage unit 2 by the face image matching unit 3 to calculate the similarity, and the similarity exceeds a certain value. In that case, identify yourself.

[0005] In FIG. 3, the fingerprint information storage unit 2 of the user is stored in advance. Then, the fingerprint information input from the fingerprint input unit 1 and the fingerprint information of the fingerprint information unit 2 are compared by the fingerprint matching unit 3 to calculate the similarity, and when the similarity exceeds a certain value, the fingerprint is identified as the person. . The identification is performed using the same processing as in the case of voice. Further, in order to obtain higher identification accuracy, a system using a combination of a personal identification system using biological information and a conventional system using a seal or a password has been proposed. FIG. 4 is a block diagram of a system using voice information and personal identification number information. In FIG. 4, the voice information of the user is stored in the voice storage unit 2 in advance, and the password information set by the user is stored in the password number storage unit 5 in advance. Then, a voice is input from the voice input unit 1 and a password is input from the password input unit 4, and these are compared with the voice information of the voice storage unit 2 and the password of the password storage unit 5, and the similarity is calculated. If the similarity exceeds a certain value, the person is identified as the person. Matching using this system is performed when the identification result obtained from the biometric information and the identification results of the voice collation unit 3 and the password identification unit 6 are both identified as the principal only when it is determined that the individual is the principal. There are many. Even when face image information or fingerprint information is used instead of voice information, it can be easily realized with the same system configuration as in FIG.

[0006]

As described above, in the conventional personal matching method using the current personal biometric information, the similarity between the registered biometric information and the current information is calculated, and the personal identification is performed. Is going. When calculating the similarity of biometric information, always output an accurate value due to the fact that the biometric information changes over time, or the difficulty greatly changes due to the presence of noise or changes in the surrounding environment. Is extremely difficult. That is, in these methods, there is a problem that it is difficult to constantly calculate a highly accurate similarity. For voice information, a method of improving recognition accuracy under noise conditions and a method of improving recognition accuracy by adapting to time change of voice have been proposed, but none of them has no noise or no time change of voice. Inferior to the recognition accuracy in the situation. The same applies to a method using a face image or a fingerprint. Also, in a system using a plurality of pieces of information as shown in FIG. 4, when the accuracy of individual identification results is low, it is difficult to perform high-precision identification only by simple combination.

The present invention solves such a problem of the conventional individual identification method. When an individual is identified using a plurality of pieces of information, the reliability and the similarity of the identification are determined for each piece of information. The purpose of the present invention is to perform a more accurate individual identification by using an integrated rule that determines the entire individual identification result based on the information. By updating the integration rule by learning, it is possible to perform individual identification with higher accuracy. In the case of personal identification using biological information, low reliability is used when more accurate identification such as time lapse or surrounding conditions cannot be performed, and high reliability when the surrounding conditions are almost ideal without time lapse. Output, and the integration unit considers the reliability and outputs the final identification result with emphasis on the identification result with high reliability, so that more accurate individual identification can be performed compared to the conventional method. it can.

[0008]

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention relates to a personal identification apparatus and method for identifying an individual by using individual biological information, which includes at least one individual biological information. Regarding multiple personal information of personal identification,
An input unit for inputting personal information and converting it into a feature amount if necessary, and a feature amount that can be collated and compared with the input personal information are stored. A storage unit that performs a comparison between the input personal information and the stored feature amount and outputs a similarity information, and calculates a reliability information from the input information and the stored feature amount A reliability calculation unit, and an identification result integration unit having an integration rule for determining an overall comparison result based on the similarity information and the reliability information output from the plurality of matching units are provided. Further, the invention is characterized in that an integration rule is learned from learning data.

According to the present invention having such features,
After inputting multiple pieces of information from each input unit and converting them to features if necessary, perform a match determination for each piece of information and output a degree of similarity for each piece of information and a degree of reliability that indicates how much similarity is reliable at the same time I do. Then, the identification result integration unit calculates the overall collation result based on the integration rules based on the plurality of similarities and the degrees of reliability, and identifies the individual. The integration rules are learned and updated using learning data.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION

Embodiment 1 FIG. 5 is a block diagram showing an overall configuration of a personal verification method according to an embodiment of the present invention. The personal verification method shown in the figure uses a plurality of input means for personal verification and a plurality of verification means. In this embodiment, three pieces of biometric information of voice information, face image information, and fingerprint information are used as a plurality of pieces of information.

As for voice information, a voice input unit 1 for inputting voice, a voice storage unit 2 for storing voice information registered in advance, a voice matching unit for comparing input voice information with stored voice information and calculating similarity 3. The voice reliability calculation unit 4 for calculating the reliability of the input voice information is provided with the face image information
As for the fingerprint information, the face image input unit 5, the face image storage unit 6, the face image matching unit 7, the face image reliability calculation unit 8,
A fingerprint input unit 9, a fingerprint storage unit 10, a fingerprint collation unit 11, and a fingerprint reliability calculation unit 12 are used. The identification result integration unit 13
The similarity calculated by each of the collating units 3, 7, 11;
The reliability calculated by each of the reliability calculation units 8 and 12 is integrated according to the integration rule, and finally a determination as to whether the user is the principal is output.

In this embodiment, the speech information is stored in the speech storage unit in a form in which a word determined for the speech input unit is uttered several times in advance and features are extracted. The face image information is obtained by inputting a face image facing the front to the image input unit in advance,
The image is stored in the face image storage unit in a form in which features are extracted. The fingerprint information is stored in the fingerprint storage unit in a form in which a fingerprint is input from the input unit in advance and features are extracted from the entire fingerprint. At the time of personal identification, the input information is subjected to feature conversion, and the similarity is calculated by comparing the input information with personal information stored in each storage unit.

In the present embodiment, the degree of reliability is as follows.
It is defined by the difference between the N ratios and the number of days elapsed from the time of creating the dictionary to the time of voice input. With respect to the SN ratio, higher reliability is calculated as the difference is smaller, and as to the elapsed days, the smaller the number of days, the higher the reliability. The image information is defined by the direction of the face at the time of inputting the image and the number of days elapsed from the time of creating the dictionary to the time of inputting the voice. As for the direction of the face, a higher reliability is calculated when the face is closer to the front than when the face is horizontal, and when the number of elapsed days is smaller, the reliability is calculated. Fingerprint information is defined using the area of the input fingerprint portion. The higher the area, the higher the reliability is calculated. Regarding the image, if a face portion cannot be detected, a re-input is requested. Regarding the fingerprint, if the area is too small to be used for collation, a re-input is required before the integration of the similarity and the reliability.

FIGS. 6, 7 and 8 show specific rules for calculating the reliability. FIG. 6 shows a reliability calculation rule for audio information. FIG. 6A shows a reliability calculation rule relating to the SN ratio.
(B) shows the reliability calculation rule regarding the elapsed days. The voice reliability calculation unit calculates the reliability of the voice information based on these rules. Here, the average of the reliability obtained by the rule of FIG. 6A and the reliability obtained by the rule of FIG. 6B is defined as the reliability of the audio information. FIG.
7A shows a reliability calculation rule for face image information. FIG. 7A shows a reliability calculation rule for the face direction, and FIG. 7B shows a reliability calculation rule for the elapsed days. The face image information reliability calculation unit calculates the reliability of the face image information based on these rules. Here, the average of the reliability calculated in FIG. 7A and the reliability calculated in FIG. 7B defines the reliability of the face image information. FIG. 8 shows a rule for calculating the reliability of fingerprint information. The fingerprint information reliability calculation unit calculates the reliability using this rule. The reliability of each information is 1.
0, the minimum value is 0.2, and the voice information and the face image information are .0.
The fingerprint information is output in increments of 0.2 in increments of 05.

In this embodiment, the calculation of the similarity in the collation unit for the voice information, image information and fingerprint information of 3, 7, and 11 is performed in the same manner as the conventional method. FIG. 9 shows an integration rule used in the identification result integration unit. This integration rule specifies the similarities output from the voice matching unit, the face image matching unit, and the fingerprint matching unit, and the reliability values output from the voice reliability calculating unit, the face image reliability calculating unit, and the fingerprint reliability calculating unit. This is a rule for integrating the degrees and calculating the final similarity. This rule selects information whose reliability of each information is equal to or greater than a threshold value (step 20), averages the similarities of all the selected personal information, and (step 21) determines whether or not the obtained value exceeds a threshold value. If the person is determined to be the person (step 22), the final similarity is calculated. Here, the threshold is set to 0.6. The identification result integration unit calculates the final similarity using the integration rule, and determines that the user is the principal if the similarity is equal to or greater than a predetermined threshold, and otherwise determines that the user is not the principal. Here, the threshold is set to 0.8, and when the final similarity is 0.8 or more, it is determined that the user is the user, and when the final similarity is less than 0.8, it is determined that the user is not the user.

Next, the operation of this embodiment will be described.
For voice, image, and fingerprint, feature amounts were extracted from data created 14 days ago and stored in each storage unit. The SN ratio at the time of voice data collection was 30 dB.

First, the person himself / herself entered three pieces of information of voice, image and fingerprint to identify the individual. The SN ratio at the time of inputting the information was 5 dB, and the face was recognized as facing the front with respect to the face image input unit. As for the fingerprint information, 90% of the data determined by the system was obtained.

As a result of calculating the similarity based on the conventional method in each information collating unit, the similarity of voice information is 0.4, the similarity of face image information is 0.9, and the similarity of fingerprint information is 0. 0.9. Next, the reliability calculation rule section of each information calculated the reliability based on each reliability calculation rule. From FIG. 6A, the audio information reliability based on the SN ratio is calculated as 0.2, and from FIG. 6B, the audio information reliability based on the elapsed days is calculated as 0.6. .4 was output. As the face image information reliability, an average 0.8 of the reliability 1.0 based on the face orientation in FIG. 7A and the reliability 0.6 based on the number of days elapsed in FIG. 7B was output. As the fingerprint information reliability, 1.0 was output according to the rule shown in FIG.

The information result integration unit integrates the similarity and reliability of each information. Integration is performed using the integration rule of FIG. Since the reliability of the audio information is 0.4 and less than 0.8, the similarity of the audio information is not used for the final similarity calculation. Since the reliability 0.8 of the face image information and the reliability 1.0 of the fingerprint information are each 0.8 or more, the final similarity is obtained by averaging these similarities according to the rule of FIG. . Since the similarity of the face image information is 0.9 and the similarity of the fingerprint information is 0.9, the final similarity is calculated as 0.9 and the threshold value is 0.8 or more. Then she was determined to be herself.

If the matching is performed by the conventional method, the reliability is not calculated, so that only the voice information similarity 0.4, the face image information similarity 0.9, and the fingerprint image information similarity 0.9 are output. . In order to integrate the matching results, a method of averaging all the similarities or a method of determining that the individual is the only person when all the similarities exceed the threshold value are generally used. Is less than the threshold value 0.8 and the similarity of the voice information is a low value of 0.4, so that the user is not determined to be a person by either method. In this system,
Since the reliability of the voice information is low, the identity of the person can be correctly identified by ignoring the similarity of the voice information. This result indicates that highly accurate individual identification can be performed by calculating the reliability.

Next, consider a case in which another person inputs three pieces of information of a voice, a face image, and a fingerprint in a similar system. The voice information was input using the noised person, and the face image information was input using the photograph of the person photographed from the front. The SN ratio with respect to the surrounding environment was 30 dB, and the number of elapsed days was 14 days. As a result of calculating the similarity based on the conventional method in each information collating unit, the similarity of the voice information is 0.9, the similarity of the face image information is 0.9, and the similarity of the fingerprint information is 0.4. became.
The reliability of each piece of information is 0.8, which is an average of 1.0 of FIG. 6A and 0.6 of the rule of FIG.
Is calculated. As for the face image information, 0.8 is calculated by averaging 1.0 in FIG. 7A and 0.6 in FIG. 7B. For fingerprints, the contact area should be sufficient
Therefore, 0.8 was calculated. Since all the reliability values exceed 0.8 when integrated by the integration rule, the final similarity 0.73 is calculated by averaging the three similarities. Since it is less than 0.8, it is possible to output a correct result indicating that the user is not the user.

In the example of the first person, the similarity is 0.4, 0.9, 0.9 in the order of voice, face image, and fingerprint. In the example of the next person, the similarity is 0. 9, 0.9, 0.
4. If only the numbers are seen, the values are the same, and it is impossible to output a judgment that the former is the principal and the latter is another person only by the similarity. Calculate reliability along with similarity,
It can be seen that by performing the integration according to the integration rule, a highly accurate determination can be made as compared with the conventional method.

Embodiment 2 In this embodiment, a system having the same system configuration as that of Embodiment 1 but changing the information held in each storage unit is used. In the first embodiment, the voice information stores the feature information of the voice information when a predetermined word is uttered, but in the present embodiment, the five vowels “A”, “I”, “U”, “E” are stored. The feature extraction of the utterance of “O” is stored in the voice storage unit. The storage method is performed by uttering a word specified by the system, performing voice recognition, and retaining data of a corresponding vowel portion and extracting features. In the first embodiment, the front image is used as the face image information. However, in the present embodiment, information obtained by extracting a plurality of images such as upward and sideways in addition to the front is stored. In the first embodiment, the entire fingerprint is used as the fingerprint information, but in the present embodiment, features are extracted and stored in a form in which the entire fingerprint is divided into a plurality of locations.

When recognizing speech, the system asks the user to utter a word specified by the system, extracts features and extracts vowel parts by speech recognition, and then stores information stored for each vowel. The similarity is obtained by comparing with. As for the face image, the similarity is calculated by comparing the input image with an image in the direction of the face closest to the face of the input image among the stored images. As for the fingerprint, the information inputted in a plurality of locations stored fragmentarily is compared with the information in the storage unit, and the similarity is output.

The calculation method of the reliability and the integration rule are the same as those in the first embodiment. In this system,
Consider a case where another person has input three pieces of information: voice, face image, and fingerprint. An image of the fingerprint information was obtained and input. For the voice information and the face image information, the voice information was input in the situation of the S / N ratio of 5 dB and the face image information was input in the horizontal direction so as to reduce the reliability as much as possible. The similarity was output as 0.4 for voice information, 0.4 for face image information, and 1.0 for fingerprint information. Regarding the reliability, the voice information is 0.4 and the face image information is 0.4.
8. 0.8 was output as fingerprint information. As a result of the judgment by the integration rule, the average similarity of the face image information and the fingerprint information with the reliability of 0.8 or more is 0.7, which is lower than the threshold value 0.8. Was done.

In the above example, in the case of the system of the first embodiment, since a landscape image is input as a face image, the reliability of the face image is reduced, and as a result, the similarity of the fingerprint information is finally reduced. The degree of similarity is determined, and the person is determined to be the person. On the other hand, in the present system, when the face image is stored, not only the front image but also the horizontal image is stored, so the reliability of the face image is increased, and as a result, it is possible to output a correct result that the user is not himself / herself. Was. When another person obtains some information of the person, a method of intentionally lowering the similarity of the other information and impersonating the person may be considered.However, as in this embodiment, the part used for recognition is limited or plural. By preparing the user and not informing the user of the part to be used for recognition, it is virtually impossible to intentionally lower the reliability and input information. Therefore, highly accurate individual identification can be performed even for a malicious person.

Embodiment 3 In this embodiment, a system configuration as shown in FIG. 10 is employed, and an individual is identified using password information in addition to three pieces of biometric information of voice information, face image information, and fingerprint information. . In FIG. 10, the voice input unit 23, the face image input unit 27, the fingerprint input unit 3
1. The voice information, the face image information, the fingerprint information, and the personal identification number information respectively input from the personal identification number input unit 35 are used as the voice information of the voice storage unit 24, the face image information of the face image storage unit 28, and the fingerprint storage unit 33. The voice information, together with the fingerprint information and the password information in the password number storage unit 37, are stored in the voice matching unit 25,
The face image information is input to the face image matching section 29 and the face image reliability calculating section 30, and the fingerprint information is input to the fingerprint matching section 32 and the fingerprint reliability calculating section 34 and the password is input to the voice reliability calculating section 26. The number information is input to a password collating unit 36 and a password reliability calculating unit 38. The similarity information obtained by each matching unit and the reliability information obtained by each reliability calculation unit are input to the identification result integration unit 39, and the individual is identified based on the integrated result. . The calculation of the similarity and the reliability of the three pieces of biometric information is performed in the same manner as in the second embodiment. Regarding the password information, the password collating unit includes:
0 or 0.0, that is, two types of similarity, correct or incorrect, are output. Here, the password reliability output unit always outputs a reliability of 1.0. The integration rule shown in FIG. 11 is used. In the integration rule FIG. 9 of the second embodiment, a reliability of 0.8 or more is output. The integration rule shown in FIG. 11 is used.
In the integration rule of FIG. 9 according to the second embodiment, the average of similarity of information having a reliability of 0.8 or more is compared with the threshold.
Then, only when all the similarities of the information having the reliability of 0.8 or more exceed the threshold value 0.8, it is determined that the user is the principal. (Steps 40 and 41) In this system, it is assumed that another person has input four pieces of information of a voice, a face image, a fingerprint, and a personal identification number. If the recorded information is prepared for the voice information, face image information and fingerprint information of the person, high similarity and high reliability can be obtained for any information of voice, face image and fingerprint It is. However, in the present system, the reliability of the personal identification number is always 1.0, and unless the similarity of the personal identification number becomes 1.0, that is, unless the correct personal identification number is input, the personal identification is not determined. It has become. As described above, it is a great advantage of the present invention that the robustness of excluding others can be easily changed by changing the input information and the integration rules.

Embodiment 4 In Embodiment 2, similarity and reliability are determined using a neural network instead of the integration rule and threshold shown in FIG. The neural network has an input layer (units 42 to 45) and an intermediate layer unit 46 to 46 shown in FIG.
49. A fault structure composed of the output layer units 50 to 51 is adopted, and the connection is only a forward connection of the input layer → intermediate layer → output layer, and there is no connection in each layer. Figure 1 shows the structure of each unit.
This is a modeling of brain neurons as shown in FIG. Weights indicating the strength of the connection are given to the inputs of the units 52 and 53. The summation of the inputs is performed by the adder 54 in consideration of the weights, and the value is calculated by the conversion function unit 55 using a certain conversion rule. Is converted and output. At the time of learning, the weight and conversion rule of each unit are changed.

In this embodiment, six input layers and the intermediate layer 21
And a neural network having one output layer unit. At the time of learning, a vector having three elements of three similarities and three degrees of reliability obtained by inputting voice information, face image information, and fingerprint information by the subject or another person in advance is input, and only the case of the subject is supported. Learn enough to activate the units in the output layer.

At the time of recognition, the user inputs voice information, face image information, and fingerprint information in the same manner as in the second embodiment. Similar to the second embodiment, the similarity and the reliability of each information are calculated. The calculated similarity and reliability are input to a learned neural network. If the output of the output layer exceeds a predetermined threshold value, it is determined that the user is the user, and if not, that the user is another person.

The integration rules in the first and second embodiments are as follows:
There was a problem that it was difficult to decide the best one in advance, but by learning rules using training data,
Appropriate integration rules can be easily set.

[0032]

According to the present invention described in detail above, personal identification is performed using a plurality of personal information, and reliability information is output for each personal information. Since the identification result used can be prioritized,
It is possible to obtain higher individual identification accuracy than before. Since the integration rule is updated by learning, it is possible to obtain robust personal identification accuracy even when the surrounding situation changes.

[Brief description of the drawings]

FIG. 1 is a block diagram of a system for performing individual identification using voice information.

FIG. 2 is a system block diagram for performing individual identification using a face image.

FIG. 3 is a system block diagram for performing individual identification using a fingerprint.

FIG. 4 is a system block diagram for performing personal identification using voice information and personal identification number information.

FIG. 5 shows an embodiment of the present invention.

FIG. 6 is a diagram showing a first reliability calculation rule;

FIG. 7 is a diagram showing a second reliability calculation rule;

FIG. 8 is a diagram showing a third reliability calculation rule;

FIG. 9 is a diagram showing an integration rule used in the identification result integration unit of the present invention.

FIG. 10 shows an embodiment of the present invention.

FIG. 11 is a diagram showing a method of determining an integration rule used in the identification result integration unit of the present invention.

FIG. 12 is a diagram showing one configuration in a case where determination of similarity and reliability of the present invention is performed using a neural network.

FIG. 13 is a diagram showing one configuration of each unit.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1: Voice input part 2: Voice collation part 3: Voice storage part

Claims (4)

[Claims] 1. A personal identification device which collects a plurality of personal information including at least one or more biometric information and identifies an individual by using the collected personal information. A storage unit for storing a feature amount converted from information; an input for inputting the personal information and converting it to a feature amount if necessary; a personal information or feature amount in the storage unit and the input personal information or feature amount A matching unit that collates and outputs similarity information; a reliability calculation unit that obtains reliability information from the personal information or feature amount in the storage unit and the input personal information or feature amount; An individual identification device, comprising: an identification result integration unit that outputs an identification result from similarity information and reliability information using an integration rule. 2. The personal identification apparatus according to claim 1, wherein an integration rule in the identification result integration unit is learned using learning data created based on personal information input to the input unit. 3. A personal identification method for collecting a plurality of personal information including at least one or more biometric information and identifying the individual based on the collected personal information. The personal information is input, and if necessary, the personal information is used as a feature amount. A step of converting; a step of storing personal information registered in advance for each personal information or a feature amount converted from this personal information; a step of storing the stored personal information or feature amount and the input personal information or feature amount Collating and outputting similarity information; obtaining reliability information from the stored personal information or feature quantity and the input personal information or feature quantity; and obtaining the plurality of pieces of similarity information and Obtaining an identification result from the reliability information using an integration rule. 4. The personal identification method according to claim 3, wherein the integration rule is learned and updated at any time using learning data created based on the input personal information.