JPH10124669A - Fingerprint collation device - Google Patents

Abstract

(57) [Summary] [Problem] To provide a fingerprint matching device that is resistant to various changes in a collected fingerprint image and has a high personal identification rate. SOLUTION: A fingerprint collating apparatus A includes a fingerprint image capturing means 1 for optically collecting a fingerprint and outputting an image signal, and an A / D converting means 2 for converting the image signal into fingerprint image data.
And a feature extracting means 3 for extracting a fingerprint feature from the fingerprint image data, and outputting a code label of a code vector closest to one frame of the extracted fingerprint feature, thereby obtaining a sequence of fingerprint features. Means 4 for obtaining one sequence of quantized code sequences, a statistical probability model estimating means 5 for estimating a statistical probability model using the quantized code sequence, and a statistical pattern model for each registered finger. A registration means 6 for registering the identification target, and a determination means 7 for determining the acceptance probability of the symbol sequence of the fingerprint to be identified by the statistical probability model of the registered finger, and determining that the identification target is a registrant if the reception probability is equal to or greater than a set threshold. Is provided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a fingerprint collating apparatus for collating a fingerprint.

[0002]

2. Description of the Related Art In general, a fingerprint collating apparatus includes a fingerprint image capturing means for optically capturing a fingerprint by pressing a fingerprint of a finger against a prism or the like, and an A / A converter for converting an image signal into fingerprint image data.
D conversion means, feature parameter extraction means for extracting feature parameters from the converted fingerprint image data, memory for storing the feature parameters, registration means for storing the feature parameters in the memory means, Collation means for performing a registrant collation operation for collating the registrant characteristic parameters with the registrant characteristic parameters stored in the storage means.

[0003]

However, in the above fingerprint collating apparatus, if various variations such as missing or distortion occur in the collected fingerprint image, the collation accuracy is significantly reduced. This is because variations in the fingerprint sampled image input at the time of matching cannot be absorbed by the feature extraction method or matching algorithm extracted from the fingerprint image at the time of registration. It is said that fingerprints are lifelong, but actually, fingerprint images are collected according to season, environment,
It changes under the influence of physical condition.

[0004] Fingerprint feature points (end points and branch points of fingerprint ridges)
In the case of a method of performing pattern matching on the basis of, a fingerprint image that is tilted (rotated) or distorted can be flexibly collated by improving preprocessing and generating many rules and networks. However, if there is a wrinkle or blurred image in the collected fingerprint image at the time of registration or verification due to finger injury or poor collection, etc., the matching points are not sufficiently extracted, and the verification accuracy is reduced. .

On the other hand, in the case of a method in which pattern matching is performed at the level of the original image of the fingerprint or in the frequency domain, the emphasis is not placed on extracting the characteristic points of the fingerprint. On the other hand, the accuracy of collation decreases with a distorted or skewed fingerprint image. An object of the present invention is to provide a fingerprint matching device that is resistant to various changes in a collected fingerprint image and has a high personal identification rate.

[0006]

In order to solve the above-mentioned problems,
The present invention employs the following configuration. (1) The fingerprint collating device includes: a fingerprint image capturing unit that optically collects a fingerprint of a finger and outputs an image signal; an A / D conversion unit that converts the image signal into digital data; Extracting a fingerprint feature quantity from the extracted fingerprint feature quantity, and outputting a code label of a code vector closest to one frame of the extracted fingerprint feature quantity, thereby performing one-sequence quantization on one-sequence fingerprint feature quantity. A quantizing means for obtaining a code sequence; a statistical probability model estimating means for estimating a statistical probability model using the quantized code sequence obtained by the quantizing means as a symbol sequence; Registration means for registering the statistical probability model for each registered finger, and the acceptance probability of the symbol sequence of the fingerprint to be identified based on the statistical probability model of the registered finger. If more than a set threshold identification target person and a determination means that the registrant.

(2) The fingerprint collation device optically collects a fingerprint of a finger and outputs an image signal, a fingerprint image capturing device, and an A / D converter that converts the image signal into digital data. A feature extracting means for extracting a fingerprint feature from the digital data obtained, and a distance (D _min ) between each frame of the extracted fingerprint feature and a closest code vector is obtained. The first candidate normalized distance (S) obtained by normalizing the inter-vector distance by each distance (D _min ) finds up to n code labels of code vectors whose thresholds are smaller than or equal to the threshold (γ). a quantizing means for obtaining an n-sequence symbol sequence; a statistical probability model estimating means for estimating a statistical probability model using the quantization code sequence obtained by the quantizing means; A registration means for registering a statistical probability model for each registered finger estimated by the determination means, and an acceptance probability of a symbol sequence of the fingerprint to be identified by the statistical probability model of the registered finger. Determining means for determining that the person is a registrant.

(3) The fingerprint matching device has the configuration of (2) above, and changes the threshold value (γ) of the quantizing means depending on the existence area in the fingerprint image of each frame, or determines whether or not the threshold value (γ) for the second candidate and thereafter is different. Limit the total number of output labels of all frames to L.

(4) The fingerprint collation device optically collects a fingerprint of a finger to output an image signal, and an A / D converter for converting the image signal into digital data. Means for extracting a fingerprint feature from the extracted digital data, and outputting a code label of a code vector closest to one frame of the extracted fingerprint feature, thereby obtaining one series of fingerprint features for one series. , And randomly select a plurality of frames from a region where the amount of variation in the obtained symbol sequence is large. The symbols of these frames are identified by a first candidate normalized distance (S) having a threshold (γ). Quantizing means for obtaining a plurality of symbol sequences replaced with the symbols of the following code vector code labels, and a quantized code sequence obtained by the quantizing means Statistical stochastic model estimating means for estimating a statistical stochastic model by using as a register, a registering means for registering a statistical stochastic model for each registered finger estimated by the statistical stochastic model estimating means, and a statistical stochastic model of the registered finger And determining means for determining the acceptance probability of the symbol sequence of the fingerprint to be identified by the above-described method, and determining that the identification target is a registrant if the acceptance probability is equal to or greater than a set threshold.

[0010] (5) The fingerprint collation device is characterized in that
(4) The above-mentioned feature extracting means includes:
Any of group delay spectrum conversion, analysis methods such as spectrum and cepstrum, feature point matching, direction code matching, and projection matching are used.

[0011]

[Action and effect of the invention]

[Claim 1] At the time of fingerprint registration or fingerprint collation, the fingerprint image capturing means optically collects the fingerprint of the registrant (or identification target) finger and outputs an image signal. A / D
The conversion means extracts a fingerprint feature from the converted digital data. The feature extracting unit extracts a fingerprint feature from the converted digital data.

[0012] The quantizing means calculates one of the extracted fingerprint features.
By outputting the code label of the code vector closest to the frame, one sequence of quantized code sequences is obtained for one sequence of fingerprint feature amounts. For example, when five fingerprints are collected at the time of registration, the number k of symbol sequences is five.
As a result, learning data required by the statistical probability model estimation unit is generated.

The statistical probability model estimating means estimates a statistical probability model using the quantized code sequence obtained by the quantizing means as a symbol sequence. In other words, since the transition of the time-series data can be captured by a statistical probability model, a statistical probability model capable of recognizing a registered finger can be estimated from a fingerprint image that has been distorted or displaced. Rate is obtained.

At the time of fingerprint registration, the registration means registers the statistical probability model for each registered finger estimated by the statistical probability model estimation means. Thus, the data size can be made substantially constant regardless of the number of input fingerprints at the time of registration.

At the time of fingerprint collation, the determination means obtains the acceptance probability of the symbol sequence of the fingerprint to be identified based on the statistical probability model of the registered finger, and determines that the identification target is a registrant if the acceptance probability is equal to or greater than a set threshold. I do.

The quantizing means obtains the distance (D _min ) between the closest code vector for each frame of the fingerprint feature quantity extracted by the feature extracting means, and calculates the distance between each frame and the code book. The first candidate normalized distance S obtained by normalizing each inter-vector distance by each distance (D _min ) finds up to n code labels of code vectors whose threshold value is equal to or less than the threshold γ. Is obtained. This makes it possible to make the number of learning patterns necessary for estimating the statistical probability model larger than the number of registered fingerprints. As a result, the same effect as when a large number of registered fingerprints are collected at the time of registration can be obtained.

[Claim 3] The threshold value γ of the quantization means
Is changed depending on the existence area of the fingerprint image of each frame, or the total number of output labels of all frames after the second candidate is limited to L. As a result, the influence of the outer peripheral portion (up, down, left, and right) of the fingerprint image having a large fluctuation amount is suppressed, and a decrease in the matching accuracy can be avoided.

According to a fourth aspect of the present invention, the quantizing means outputs a code label of a code vector closest to one frame of the extracted fingerprint feature amount, so that one series of fingerprint feature amount is output. Find a quantization code sequence. Then, a plurality of frames are randomly selected from the region where the amount of variation in the obtained symbol sequence is large, and the symbols of these frames are replaced by a code of a code vector whose first candidate normalized distance (S) is equal to or smaller than a threshold (γ). A plurality of symbol sequences replaced with label symbols are obtained.

The number of learning patterns required for the statistical probability model estimating means can be made larger than the number of registered fingerprints. As a result, the same effect as when a large number of registered fingerprints are collected can be obtained. In the learning of the statistical probability model, as the number of learning patterns increases, the estimation accuracy of the model improves, and the matching rate also improves.

[Feature 5] The feature extracting means extracts a fingerprint feature from the converted digital data.
It is preferable to use any one of the following methods as the feature extraction method. (1) Group delay spectrum conversion A method of suppressing a gentle envelope component of a frequency spectrum in a fingerprint ridge and separating and enhancing individual spectral peaks.

(2) Spectrum analysis method, cepstrum analysis method LPC spectrum analysis method and L
PC cepstrum analysis method. (3) Feature point matching (minutiae matching) focusing on the network structure of the feature points of fingerprints and the number of ridges between feature points. Direction code matching that codes the global flow direction of fingerprint ridges. Projection matching in which the projection of density values in the horizontal direction orthogonal to the longitudinal direction of the finger is a feature amount.

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention is shown in FIGS.
It will be described based on. As shown in FIG. 1, the fingerprint collation device A includes a fingerprint image capturing unit 1 that optically collects a fingerprint and outputs an image signal, an A / D conversion unit 2 that converts an image signal into digital data, Feature extracting means 3 for extracting a fingerprint feature quantity from the obtained digital data, quantizing means 4, statistical probability model estimating means 5 for estimating a statistical probability model, and registration for registering the statistical probability model as a registration pattern. And a determination unit for determining whether the identification target person is a registrant.

The fingerprint image capturing means 1 is installed near the entrance door, and has a right-angle prism in which the fingerprint surface of the finger is pressed against the upper surface of the prism; a light source for projecting illumination light on the prism standing surface; The CCD elements are arranged in parallel and take in the reflected light corresponding to the fingerprint ridge peak as a bright image and the reflected light corresponding to the fingerprint ridge valley as a dark image (none of them are shown). And every predetermined time (several hundred ms)
A video signal for one screen (a grayscale image of a fingerprint ridge) is transmitted to the A / D converter 2 via a coaxial cable.

The A / D converter 2 has an A / D converter (not shown), converts a video signal transmitted from the fingerprint image capturing unit 1 into digital data, and stores the digital data in a built-in memory. The obtained fingerprint original image has 128 vertical pixels × 128 horizontal pixels and 256 gradations.

The feature extracting means 3 has an arithmetic unit (not shown), and performs group delay spectrum conversion on the original fingerprint image 30 in the horizontal direction (x-axis direction) and the vertical direction (x-axis direction) as shown in FIG. (y-axis direction), and the obtained group delay spectrum GDSX in the x-axis direction and the obtained group delay spectrum GDSY in the y-axis direction
Is the fingerprint feature amount.

The group delay spectrum GDSX in the x-axis direction is
The image grayscale value (256 gradations) for one line (128 points) in the horizontal direction (x-axis direction) of the fingerprint original image is regarded as waveform data for one frame, and a hamming window is applied to each frame of 128 frames, and a 14th order LPC analysis is performed, and the frequency band channel number is obtained by converting into a group delay spectrum of 32.

As a result, the group delay spectrum GDSX is
The vector sequence is composed of vectors (128 dimensions) for 128 frames. Further, the group delay spectrum GDSY in the y-axis direction is obtained in the same manner as the group delay spectrum GDSX by setting the analysis direction to vertical (y-axis direction).

The feature extracting means 3 may extract a fingerprint feature by the following configuration. The fingerprint original image is extracted from either the horizontal or vertical (x-axis and y-axis directions). Only the frame where the fingerprint exists may be used as the feature amount. The order of analysis and the number of channels may be set to different values suitable for fingerprint collation.

The quantization means 4 has a vector quantizer 41 and a vector quantization codebook 42. The vector quantizer 41 converts the fingerprint feature value from a feature value sequence to a symbol (code label) sequence, and uses it as learning data for statistical probability model estimation. The vector quantizer 41 is
As shown in FIG. 3, with reference to the vector quantization codebook 42, a code label corresponding to a code vector closest to the obtained fingerprint feature amount (vector) is output (the vector is quantized).

The vector quantization codebook 42 stores all feature vectors of the codebook creation fingerprint, for example, 25
It is composed of representative spectra when clustering 40 into 6 classes. For creating a codebook, an algorithm such as a Splitting (two-split) LBG algorithm whose learning result is not affected by initial conditions at the time of creating a codebook is used.

For vector quantization, Full Search (brute force search) vector quantization is used in terms of quantization accuracy and collation rate. However, in codebook design, the dependency on the learning sequence is suppressed, and general-purpose search is performed. To enhance the nature,
There is also a method using mean value separated normalized vector quantization.
Also, multi-stage vector quantization may be used to reduce the amount of calculation (the amount of calculation).

The statistical probability model estimating means 5 uses the learning data obtained by the feature extracting means 3 for estimating a hidden Markov model as shown in FIG. Learning of the Hidden Markov Model is performed by repeatedly calculating and correcting each parameter of the output probability and the transition probability of the Hidden Markov Model in the registered finger so that the acceptance probability of the learning data is maximized (individual model learning). .

For the above parameter estimation calculation, Forw
Since the ard-Backward algorithm is used, many rules and network generation are not required.
Furthermore, if the smoothing of the probability value or the process of raising the probability value is performed on each estimated parameter, the collation rate decreases due to the decrease in the parameter estimation accuracy, which occurs when the number of input fingerprints is small, and the vector The effect of a quantization error at the time of quantization can be reduced.

As the number of fingerprint inputs at the time of registration increases, the estimation accuracy of the hidden Markov model improves, and as a result, the matching accuracy improves. Even if the number of fingerprint inputs increases, the data size does not increase because the statistical probability model for each registered finger estimated by the statistical probability model estimating means 5 is registered in the memory.

The registration means 6 stores each estimation parameter of the personal model obtained by the statistical probability model estimation means 5 in a memory as registration data of a registration finger.

The judging means 7 determines the acceptance probability of the symbol sequence of the person to be identified converted by the quantizing means 4 and the statistical probability model for each registered finger stored in the memory as a Viter.
It is determined using a bi-algorithm, and if the acceptance probability is equal to or greater than a set threshold, the identification target finger is determined to be a registered finger. In the calculation for obtaining the above-mentioned Viterbi algorithm, a log is taken, and the multiplication can be replaced by an addition operation, so that the calculation efficiency is high.

Further, a registered finger corresponding to the model that gives the maximum acceptance probability among all registered finger models for the fingerprint of the identification target may be used as the recognition result of the identification target fingerprint.

Next, second to fourth embodiments of the present invention will be described with reference to FIGS. The fingerprint matching device of each embodiment includes a fingerprint image capturing unit, an A / D conversion unit, a feature extraction unit, a quantization unit, a statistical probability model estimation unit, a registration unit, and a determination unit. Note that the fingerprint image capturing unit and the A / D conversion unit have the same structure as the fingerprint collation device A.

The feature extracting means has an arithmetic unit (not shown), and performs a group delay spectrum conversion on the original fingerprint image 30 in the horizontal direction (x-axis direction) and the vertical direction (y-axis direction) as shown in FIG. (In the axial direction), and the obtained group delay spectrum GDSX in the x-axis direction and the obtained group delay spectrum GDSY in the y-axis direction are defined as fingerprint feature amounts.

The group delay spectrum GDSX in the x-axis direction is
The image grayscale value (256 gradations) for one line (128 points) in the horizontal direction (x-axis direction) of the fingerprint original image is regarded as waveform data for one frame, and a hamming window is applied to each frame of 128 frames, and a 14th order LPC analysis is performed, and the frequency band channel number is obtained by converting into a group delay spectrum of 32.

As a result, the group delay spectrum GDSX is
The vector sequence is composed of vectors (128 dimensions) for 128 frames. Further, the group delay spectrum GDSY in the y-axis direction is obtained in the same manner as the group delay spectrum GDSX by setting the analysis direction to vertical (y-axis direction).

Incidentally, the feature extracting means may extract the fingerprint feature quantity by the following configuration. The fingerprint original image is extracted from either the horizontal or vertical (x-axis and y-axis directions). Only the frame where the fingerprint exists may be used as the feature amount. The order of analysis and the number of channels may be set to different values suitable for fingerprint collation.

The quantization means has a multi-label output type vector quantizer and a vector quantization codebook.

The multi-label output type vector quantizer is:
The fingerprint feature parameters extracted by the feature extraction means 3 are converted from one feature parameter sequence into a plurality of symbol (code label) sequences, and used as learning data for statistical probability model estimation. That is, the multi-label output type vector quantizer refers to the code book as shown in FIG. 6 and calculates a code label corresponding to the code vector closest to the characteristic parameter (vector) of the input fingerprint and a code vector in the vicinity thereof. Output the corresponding code label.

The multi-label output type vector quantizer is as follows.
Distance D between input vector and each vector in codebook
Based on _min , the first candidate normalized distance S shown below is output up to n code labels with the threshold γ or less at the same time.
In the case of a frame in which a code vector equal to or smaller than the threshold value γ does not exist in the series of the second and subsequent candidates (the blank area in which the label is suppressed as shown in FIG. 7), the label corresponding to the same frame as the optimal candidate is applied. I do.

In the fingerprint collating apparatus of the second embodiment, the threshold value γ is constant in all the frames of the fingerprint image. In the fingerprint matching device of the third embodiment, the threshold value γ is changed depending on the existence area of the fingerprint image of each frame. Further, in the fingerprint matching device of the fourth embodiment, as shown in FIG. 7, the number of output labels for the second and subsequent candidates is limited to L.

By using the method of the second or third embodiment, it is possible to suppress the adverse effect on the collation accuracy of a part (upper, lower, left and right ends) in the fingerprint image area where the amount of fluctuation is large.

(Equation 1)

D ₁ is the input vector and each code vector i
, And D _min indicates the distance of the optimal code vector. The vector quantization codebook is composed of representative spectra when all feature vectors of the fingerprint for creating a codebook are clustered into, for example, 256 classes. To create a codebook, for example, Split
An algorithm such as a ting (binary) LBG algorithm whose learning result is not affected by initial conditions for codebook learning (creation) is used.

In the above embodiment, FullSearch (brute force search) vector quantization is used for vector quantization in terms of quantization accuracy and collation rate. In order to suppress the noise and increase the versatility, mean value separated normalized vector quantization may be used. Also, multi-stage vector quantization may be used to reduce the amount of calculation (the amount of calculation).

The statistical probability model estimating means repeatedly calculates and corrects each parameter of the output probability and the transition probability in the hidden Markov model from the symbol sequence so that the acceptance probability of the symbol sequence is maximized (individual model). Learning). For parameter estimation, Forward-B
The ackward algorithm is used. In addition, averaging based on the weighted sum for each estimated parameter, raising the probability value, and the like are also effective.

The registration means stores in the memory the estimated parameters of the statistical probability model for each registered finger obtained by the statistical probability model estimating means as registered finger registration data.
Thus, the data size is substantially constant regardless of the number of input fingerprints at the time of registration.

The determining means obtains, using the Viterbi algorithm, an acceptance probability between the symbol sequence of the person to be identified converted by the quantization means and the statistical probability model for each registered finger stored in the memory. If the identification target is not less than the set threshold, it is determined that the identification target person is a registrant, and if the reception probability is less than the set threshold, the identification target is not a registrant.

In each model of the registered finger, the registered finger corresponding to the model giving the maximum acceptance probability may be used as the recognition result. In the calculation for obtaining the above-mentioned Viterbi algorithm, the log is taken and the multiplication can be replaced by an addition operation, so that the calculation efficiency is high.

The present invention includes the following embodiments in addition to the above embodiments. In the first embodiment, the code label of the code vector closest to one frame of the extracted fingerprint feature is output, so that one fingerprint feature for one series is output.
In this configuration, the quantization means 4 obtains a sequence of quantization code sequences. As a fifth embodiment, the quantizing means 4 determines whether the obtained symbol sequence has a large variation in the symbol sequence {near the end point,
In the case of GDSX, a plurality of frames are randomly selected from the vicinity of the upper and lower ends, and in the case of GDSY, the vicinity of the right and left ends. A configuration may be used in which a plurality of symbol sequences replaced with a symbol are obtained.

In the learning of the statistical probability model, generally, as the number of learning patterns increases, the estimation accuracy of the model improves and the matching rate also improves. However, in the fifth embodiment, the statistical probability model estimating means 5 is required. The number of learning patterns can be made larger than the number of registered fingerprints. As a result, even with the normal number of fingerprints collected, the same effect as when many registered fingerprints are collected can be obtained, and a high matching rate can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram of a fingerprint matching device according to a first embodiment of the present invention.

FIG. 2 is a conceptual diagram of a fingerprint feature extraction method and vector quantization in the fingerprint matching device according to the first embodiment.

FIG. 3 is an explanatory diagram showing an operation of a vector quantizer in the fingerprint matching device according to the first embodiment.

FIG. 4 is an explanatory diagram of a hidden Markov model.

FIG. 5 is a conceptual diagram of a fingerprint feature extracting method and a multi-label output type vector quantizer in the fingerprint matching device according to the second to fourth embodiments of the present invention.

FIG. 6 is an explanatory diagram showing the operation of a vector quantizer in the fingerprint matching device according to the second to fourth embodiments.

FIG. 7 is an explanatory diagram showing the operation of a multi-label output type vector quantizer in the fingerprint matching device according to the second to fourth embodiments.

[Description of Signs] A Fingerprint collation device 1 Fingerprint image capturing means 2 A / D conversion means 3 Feature extraction means 4 Quantization means 5 Statistical probability model estimation means 6 Registration means 7 Judgment means

Claims (5)

[Claims] 1. A fingerprint image capturing means for optically sampling a fingerprint of a finger to output an image signal, an A / D conversion means for converting the image signal into digital data, and a fingerprint feature based on the converted digital data. A feature extraction unit for extracting a quantity, and outputting a code label of a code vector closest to one frame of the extracted fingerprint feature quantity, thereby forming one series of quantized code sequences for one series of fingerprint feature quantities. Quantizing means to be obtained, statistical probability model estimating means for estimating a statistical probability model using the quantization code sequence obtained by the quantizing means as a symbol sequence, and a registration finger estimated by the statistical probability model estimating means. A registration means for registering a statistical probability model for each, a reception probability of the symbol sequence of the fingerprint to be identified by the statistical probability model of the registered finger, and the reception probability is set. A fingerprint collation device comprising: a determination unit configured to determine that the identification target is a registrant if the identification target is equal to or greater than the threshold. 2. A fingerprint image capturing means for optically collecting a fingerprint of a finger to output an image signal, an A / D conversion means for converting the image signal into digital data, and a fingerprint feature based on the converted digital data. A feature extracting means for extracting a quantity, a distance (D _min ) between each frame of the extracted fingerprint feature quantity and a closest code vector is obtained, and a distance between each frame and each vector of the codebook is determined by each distance. Up to n code labels of code vectors whose first candidate normalized distance (S) normalized by (D _min ) is equal to or smaller than the threshold (γ) are determined, and a maximum of n symbol sequences are determined for one feature amount. Quantization means to be determined, statistical probability model estimating means for estimating a statistical probability model using the quantization code sequence obtained by the quantization means, and estimation by the statistical probability model estimating means A registration means for registering the statistical probability model of each registered finger, and an acceptance probability of the symbol sequence of the fingerprint to be identified based on the statistical probability model of the registered finger. And a determination unit for determining that the fingerprint is the same. 3. The method according to claim 2, wherein the threshold value (γ) of the quantizing means is changed depending on the existence area of each frame in the fingerprint image, or the total number of output labels of all frames after the second candidate is limited to L. Fingerprint collation device. 4. A fingerprint image capturing means for optically collecting a fingerprint of a finger to output an image signal, an A / D conversion means for converting the image signal into digital data, and a fingerprint feature based on the converted digital data. A feature extraction unit for extracting a quantity, and outputting a code label of a code vector closest to one frame of the extracted fingerprint feature quantity, thereby forming one series of quantized code sequences for one series of fingerprint feature quantities. Then, a plurality of frames are randomly selected from an area having a large amount of variation in the obtained symbol series, and the symbols of these frames are replaced by a code of a code vector whose first candidate normalized distance (S) is equal to or smaller than a threshold (γ). Quantizing means for obtaining a plurality of symbol sequences replaced with label symbols, and statistically using the quantized code sequences obtained by the quantizing means as symbol sequences. Statistical probability model estimating means for estimating a probability model; registration means for registering a statistical probability model for each registered finger estimated by the statistical probability model estimating means; A fingerprint collation device comprising: a determination unit that determines an acceptance probability of a symbol sequence and determines that the identification target is a registrant if the acceptance probability is equal to or greater than a set threshold. 5. The feature extracting means includes group delay spectrum conversion, an analysis method such as spectrum or cepstrum,
5. The method according to claim 1, wherein any one of a feature point matching, a direction code matching, and a projection matching is used.
The fingerprint matching device according to any one of the above.