JPH06274628A - Personal identification estimating method - Google Patents

Abstract

PURPOSE:To estimate an accurate personal identification with the small number of data by correcting a histogram so as to smoothly change respective gaps in the case of preparing the histogram based on the distribution of the number of coincident windows, and precisely estimating the foot part of the distribution for the number of coincident windows by correcting a source function by inversely converting a corrected function. CONSTITUTION:This method is provided with eccentricity calculation processing 3 for calculating the standardized rate of a relevant function in the source distribution of the number of coincident windows standardized at respective points dividing the foot part of the selected approximate function (S3), eccentricity estimation processing 4 to correct this calculated rate so as to smoothly change it (S4), smoothing processing 5 to smooth the rate corrected by this eccentricity estimation processing 4 (S5), and correction processing 6 to correct the rate smoothed the distribution of the foot part by multiplying the rate smoothed by this smoothing processing 5 to the selected approximate function, and the estimation of the personal identification is performed with high accuracy by using the function corrected by this correction processing 6.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a personal identification estimation method for estimating a tail portion of the number of coincidence windows for personal identification.

[0002]

2. Description of the Related Art Estimating a population distribution from a fixed number of data obtained by experiments is an important issue not only in statistics but also in product development. In the product,
Its reliability is a problem and its ultimate performance must be estimated. In particular, in a pattern recognition device, for example, a fingerprint recognition device, an error in which the person cannot be recognized as the person (called a first type error) and an error in which another person is recognized as the person (second
Species error) must be minimized. In fingerprint recognition, the number of matching windows that are rectangular feature points (end points, branch points) is counted. Create a frequency distribution of the number of windows that match each other for the person and the person, and for the person and the person separately.
It can be seen that the first type of error in the former distribution has a left tail below the threshold, and the second type of error in the latter distribution has a right tail below the threshold.

From experiments, it has been attempted to accurately identify the person himself / herself and the person / others based on a finite number of frequency distributions.

[0004]

The verification is performed because the verification is performed based on the distribution of the number of matching windows of the windows of the conventional finite number of times of fingerprint verification, and the verification is performed to verify the person himself / herself and the person / others. If the number of results is small, there is a problem that it is difficult to accurately estimate the collation rate and the false collation rate for personal identification.

The present invention solves these problems.
When the histogram is created based on the distribution of the number of matching windows, it is modified so that the change in each interval changes smoothly, and the modified function is inversely transformed to correct the original function and the number of matching windows The purpose is to enable accurate estimation of the tail of the distribution and to accurately estimate the identity verification with a small number of data.

[0006]

[Means for Solving the Problems] Means for solving the problems will be described with reference to FIG. In FIG. 1, the function selection process 2 is to select a function to be approximated based on the input coincidence window number distribution.

The bias calculation processing 3 is to divide the tail portion of the selected approximate function and calculate the ratio of the function in the original matching window number distribution standardized at each divided point. The bias estimation process 4 is to correct the ratio so that it changes smoothly.

The smoothing process 5 smoothes the ratio corrected by the bias estimating process 4. Correction process 6
Is to correct the distribution at the tail by multiplying the smoothed ratio by the selected approximating function.

[0009]

According to the present invention, as shown in FIG.
Is selected based on the input matching window number distribution, the bias calculation processing 3 divides the tail of the selected approximating function, and the original matching window is standardized at each of the divided points. The ratio of the function of the number distribution is calculated, the bias estimation process 4 corrects the calculated ratio so that the calculated ratio changes smoothly, the smoothing process 5 smoothes the corrected ratio, and the correction process 6
Corrects the distribution of the tail part by multiplying the smoothed ratio by the selected approximating function, and uses the corrected function to accurately estimate the identity.

At this time, as the division of the tail portion of the selected approximate function, the number of matching windows of the tail portion of the function is equally divided. In addition, the entire approximate function selected by the function selection processing 2 is divided, and the ratio (histogram) of the function of the original matching window number distribution normalized at each divided point is calculated and selected (or in advance). (Specified) The function of the approximate tail part is divided, and the ratio (histogram) of the function of the original matching window number distribution normalized at each of the divided points is calculated separately, and the calculated former tail The product of the area of the ratio (histogram) of the area and the area of the ratio of the latter tail (histogram) is divided by the width of the foot (division width) to increase the height of the foot ratio (histogram). After synthesizing the two values, the ratio (histogram) after combination is corrected, smoothed, and corrected so that it changes smoothly, and the corrected function is used.
The person matching is estimated with high accuracy.

Therefore, when a histogram (ratio) is created based on the distribution of the number of coincident windows, it is corrected so that the change of each interval changes smoothly, and the corrected function is inversely transformed to the original function. By making the correction, it is possible to accurately estimate the tail portion of the distribution of the number of coincident windows, and it is possible to accurately estimate the identity confirmation with a small number of data.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the construction and operation of an embodiment of the present invention will be described in detail with reference to FIGS.

FIG. 1 shows a block diagram of an embodiment of the present invention.
In FIG. 1, an input process 1 is to input a coincidence window number distribution. Here, the matching window number is obtained by binarizing the image of the fingerprint read from the user's finger, performing pattern matching based on the fingerprint data retrieved from the dictionary, and matching the windows in the image in the matching window. Is the number of. The coincident window number distribution is the distribution of the coincident window numbers. For example, the distribution of the number of coincident windows is calculated and input each time when the matching is performed several hundred times.

The function selection process 2 is to select a function that approximates the coincidence window number distribution. The bias calculation processing 3 is to divide the tail portion of the selected approximate function and calculate the ratio of the function in the original matching window number distribution standardized at each of the divided points.

The bias estimation process 4 estimates the ratio of the ratio calculated by the bias calculation process 3 on a large number of populations. Here, a histogram in which the ratio changes smoothly is created. (See Figure 6).

The smoothing process 5 is for smoothing the shift ratio after being smoothed by the bias estimating process 4 (FIG. 7).
reference). The correction process 6 multiplies the selected function by the ratio. This multiplies the selected function by the ratio after being smoothed by the smoothing processing 5, and improves the accuracy of the tail portion of the function.

The output processing 7 calculates and outputs a discretized function. Next, the operation of the present invention will be described according to the procedure from S1 to S8. In S1, the distribution of the number of coincidence windows is input. This is because pattern matching is performed on the image binarized by reading from the user's finger based on the fingerprint data (coordinates of the matching window, image of the window) extracted from the dictionary, and the number of matching windows is calculated. Then, the distribution of the calculated coincidence window numbers is calculated and input.

In S2, a function approximating the coincidence window number distribution input in S1 is selected. For example, a function that has the smallest error is selected from the functions shown in FIG.

In step S3, the ratio of the original data distribution to the selected function is calculated. This calculates a standardized ratio of the original S1 coincidence window number distribution to the selected function, for example, a ratio obtained by dividing the ratio of FIG. 5A as shown in FIG. 5B.

S4 is to estimate what the ratio will be in the population,
That is, a smoothly changing histogram is created. This creates the ratio of FIG. 6, which is a histogram that changes smoothly with respect to the ratio of FIG. 5B calculated in S3.

In step S5, the ratio generated in step S4 is smoothed. This smooths, for example, the discrete ratio of FIG.
Ratio (normalized frequency / function value). S6 is
Multiply the selected function by the ratio. For this, the function selected in S2 is multiplied by the ratio smoothed in S5 to correct it, and a function (frequency) as shown in FIG. 8 is calculated.

In step S7, a discretized function is calculated. This calculates a discretized function as shown in FIG. 8B as a discretized function from the function (frequency) calculated in S6. S8 outputs. This outputs the discretized and corrected function calculated in S7. Then, using this corrected function, the collation is estimated with high accuracy.

As described above, the fingerprint read from the user's finger is subjected to pattern matching with the data extracted from the dictionary, the matching window number distribution is calculated based on the number of matching windows at that time, and this matching is performed. With the window number distribution as input, select a function to approximate. Then, with this selected function, the normalized ratio with the original distribution of the number of coincidence windows is calculated,
Divide the hem to make it smooth. After that, smoothing is performed, the selected function is multiplied, and the tail portion of the function is corrected to improve accuracy. By estimating the number of coincidences of the matching windows based on this corrected function, it becomes possible to estimate the coincidences of the windows with high accuracy and confirm the person.

FIG. 2 shows an explanatory diagram of the identity verification of the present invention. Here, the horizontal axis represents the number of matching windows in which the windows match, and the vertical axis represents the frequency of each matching window number. FIG. 2A shows how the person-to-person matching is performed. Here, the shaded portion is the frequency (degree) when the number of matching windows is equal to or less than the threshold (when the person is no longer the person). That is, it is the frequency at which the fingerprint read from the user's finger is subjected to pattern matching with the fingerprint data extracted from the dictionary, and when the number of matching windows is equal to or less than the threshold, it is determined that the user is not the person.

FIG. 2B shows how the person is collated with another person. Here, the shaded portion is the frequency (degree) when the number of matching windows is equal to or greater than the threshold value (when another person becomes the principal).
That is, it is the frequency at which the fingerprint read from the user's finger is subjected to pattern matching with the fingerprint data extracted from the dictionary, and when the number of matching windows is equal to or greater than the threshold value, another person is determined to be the person.

FIG. 3 shows a function selection explanatory diagram of the present invention.
FIG. 3A shows an example of the matching window number distribution. This is a calculation of the matching window number distribution based on a large number of matching windows subjected to pattern matching.

FIG. 3B shows an example of the function. this is,
This is a function registered in advance for selection as a function that approximates the coincidence window number distribution of FIG. Here, functions such as exponential distribution, lognormal distribution, Weibull distribution, and gamma distribution are registered in advance as candidates, and the function that most approximates them is selected.

FIG. 4 shows an example of selected functions of the present invention.
This shows the function g (k) selected from FIG. 3B as the function that best matches the distribution of the number of matching windows. A function obtained by accumulating the area of this function g (k) is represented as Q (k).

FIG. 5 shows an explanatory diagram of calculating the ratio according to the present invention. Here, the horizontal axis represents the area of the selected function, and the vertical axis represents (normalized frequency / value of function). FIG. 5 (a)
Is the normalized function of the selected function and the coincidence window number distribution, and represents the ratio thereof. When the selected function and the matching window number distribution are equal, the ratio is 1. When all of them are the same, the ratio is a straight line. Here, the ratio is as shown.

FIG. 5B shows an example of division. this is,
The area of the selected function is equally divided (0.05 equal division) with respect to the ratio shown in FIG. FIG. 6 shows an example of bias estimation according to the present invention. This is a correction of the values so that the divided ratios in FIG. 5B change smoothly. As a method of correcting this, for example, Bayes' theorem may be used.

FIG. 7 is an explanatory view of the smoothing processing of the present invention. This is obtained by smoothing the discrete ratio (histogram) of FIG. 6 by smoothing. FIG. 8 shows an example of the corrected function of the present invention.

FIG. 8A shows an example of the function. this is,
It is a function after the ratio (histogram) smoothed in FIG. 7 is multiplied by the function selected in FIG. 4 and corrected. In this corrected function, the accuracy of the tail portion is particularly improved.

FIG. 8B shows a function obtained by discretizing the function shown in FIG. This is the number of coincidence windows (a) in FIG.
Is a discrete function of. The black portion of the obtained discretized function represents the probability that the stranger described in FIG. 2B impersonates himself. Similarly, in FIG.
The probability that the person will disappear is also the accuracy of the hem.

Next, the division example of FIG. 5B will be sequentially described in detail with reference to FIGS. 9 to 13. Figure 9
The division example (1) of this invention is shown.

FIG. 9A shows an example of division of positive data. FIG. 9A-1 shows an example in which the right skirt portion of the above-described function is divided at equal intervals. Here, the matching frequency on the horizontal axis is divided at equal intervals from A0 to Ai as illustrated.

FIG. 9A-2 shows the area of the selected function. Here, the horizontal axis represents the area of the selected function, and here, the area becomes smaller toward the right. FIG. 9B shows an example of division of positive and negative data.

FIG. 9 (b-1) shows an example in which the right hem portion and the left hem portion of the above-mentioned function are divided at equal intervals. Here, the matching frequency on the horizontal axis is changed from A0 to Ai at equal intervals.
It is divided as shown in FIG.

FIG. 9B-2 shows the area of the selected function. Here, the horizontal axis represents the area of the selected function, and here, the area decreases from the center to the left and from the center to the right.

FIG. 10 shows a division example (No. 2) of the present invention. Here, the horizontal axis represents the coincidence window number X, and the vertical axis represents the frequency Y. This is an example in which the number of coincidence windows X is equally divided into the skirt portion and the remaining portion is equally divided into frequencies Y.

FIG. 10A shows the selected function g (x) and the area (histogram) Q of the selected function g (x).
Represents (x). Here, for the portion described as the (N−K) category in the bottom portion of the figure, the number of coincidence windows X is equally divided, and K, K
+1, K + 2 ... N are divided here.

On the other hand, (C-N + K) of the part other than the skirt part
The part described as a category is equally divided by the area Y. The interval at this time is equally divided as Q (K) / (C- (NK)) = Q (K) / 7 (1).

Here, C: (the number of division points) +1 N: the upper limit of the number of coincidence windows FIG. 10B shows an example of division of the category (N−K) and the category (C−N + K). Show. Y on the horizontal axis is the area of the selected function.

Here, in the parts other than the skirt part of the category (C-N + K), Y is equally divided at the intervals of the above (1). On the other hand, in the skirt portion of the category (N−K), X is equally divided, so that Y here is not equally divided.

FIG. 11 shows a division example (3) of the present invention. Here, the horizontal axis represents the coincidence window number X, and the vertical axis represents the frequency Y. This is an example in which the number of coincidence windows X is equally divided into the skirt portion and the remaining portion is equally divided into frequencies Y.

FIG. 11A shows the selected function g (x) and the area (histogram) Q of this selected function g (x).
Represents (x). Here, the left hem portion and the right hem portion equally divide the matching window number X.

On the other hand, the parts other than the skirt part are equally divided by the area Y. FIG. 11B shows an example of division of the left hem portion, the right hem portion and other portions. Y on the horizontal axis is the area.

Here, Y is equally divided into portions other than the skirt portion. On the other hand, since the left hem part and the right hem part are equally divided into X, Y here is not equally divided. FIG. 12 shows a division example (No. 4) of the present invention. This is Y
This is an example in which the division interval of is changed exponentially.

FIG. 12A shows a state in which the Y division interval is changed exponentially. Here, the horizontal axis represents the serial number of each divided category, and the vertical axis represents the interval y of Y division. This interval y is represented by the following equation y = A-λexp (-λ (mx)) in the figure. Here, A, λ, and m are parameters. In the (N−K) category at the bottom, X is equally divided. Therefore, this portion is not equally divided in Y.

FIG. 12B shows an example of division when the interval between division points is determined exponentially. Y on the horizontal axis is the area. FIG. 13 shows a division example (5) of the present invention. This is the case where both the left hem portion and the right hem portion are provided, as compared with the case of FIG. 12, and is similar to FIG.

FIG. 14 shows an example of a method of estimating a histogram in two steps according to the present invention. In (a) of FIG. 14, a model (ratio) on the histogram is configured for the entire distribution of the number of coincidence windows. As described above, this is because the bias calculation processing 3 in FIG. 1 divides the selected approximating function and standardizes at each of the divided points the ratio of the corresponding function of the original matching window number distribution (histogram). Is calculated and used as a model.
Here, the height of each part is as follows.

(Height of each portion) = (area of each one of the histograms designated by) / (width of the histogram) forms a model (ratio) on the histogram for the skirt portion. This is the same as in, dividing the function that approximates the selected tail part (or a pre-specified function (for example, uniform function)) and standardizing the original matching window number distribution at each of the divided points. The ratio (histogram) of the relevant function of is calculated and used as a model.

Combines the models of and. That is,
The height of the hem portion = (total area of the hem estimated by) × (area of the divided hem portion estimated by) ÷ (width of the histogram) for each of the divided hem portions repeat.

However, here, (the area of the hem specified by)
Is 1 for the entire hem. As described above, the histogram (ratio) was calculated for the selected entire function, and then the histogram (ratio) was calculated for the selected tail function according to the input of the distribution of the number of coincidence windows. After that, the area of the hem part of the former is multiplied by the area of the hem part of the latter and divided by the division width of the hem part to obtain the height of the hem part, to synthesize the hem part. . Regarding the histogram (ratio) after this combination, S in FIG.
Processing after 4 (correction so that it changes smoothly, smoothing,
And correction), and the corrected function is used to estimate the personal identification with high accuracy.

[0054]

As described above, according to the present invention,
When a histogram is created based on the distribution of the number of coincident windows, it is modified so that the changes in each interval change smoothly, and this modified function is inversely transformed to correct the original function. Therefore, the tail portion of the distribution of the number of coincident windows can be accurately estimated, and the estimation probability of the tail portion can be increased particularly when the number of data is small, and accurate identification confirmation can be performed.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention.

FIG. 2 is an explanatory diagram of personal identification according to the present invention.

FIG. 3 is an explanatory diagram of function selection according to the present invention.

FIG. 4 is an example selected function of the present invention.

FIG. 5 is an explanatory diagram of calculating a ratio according to the present invention.

FIG. 6 is an example of bias estimation according to the present invention.

FIG. 7 is an explanatory diagram of smoothing processing according to the present invention.

FIG. 8 is an example of a corrected function of the present invention.

FIG. 9 is a division example (1) of the present invention.

FIG. 10 is a division example (2) of the present invention.

FIG. 11 is a division example (3) of the present invention.

FIG. 12 is a division example (4) of the present invention.

FIG. 13 is a division example (5) of the present invention.

FIG. 14 is an example of a method of estimating a histogram in two steps according to the present invention.

[Explanation of symbols]

 1: Input processing 2: Function selection processing 3: Bias calculation processing 4: Bias estimation processing 5: Smoothing processing 6: Correction processing 7: Output processing

Claims (3)

[Claims] 1. A function selection process (2) for selecting an approximate function based on an input distribution of the number of matching windows in a principal matching estimation method for estimating a tail portion of the number of matching windows when performing the individual matching. A bias calculation process (3) of dividing the tail portion of the approximating function selected by the function selection process (2) and calculating the ratio of the function of the original matching window number distribution normalized at each of the divided points ), A bias estimation process (4) for correcting the ratio calculated by the bias calculation process (3) so as to change smoothly, and a smoothing process for smoothing the ratio corrected by the bias estimation process (4). The processing (5) and the ratio after smoothing by this smoothing processing (5) are
A correction process (6) for correcting the distribution of the tail portion by multiplying the selected approximating function, and using the function corrected by the correction process (6),
A personal identification estimation method, characterized in that the personal identification estimation is configured with high accuracy. 2. The person-to-person matching estimation according to claim 1, wherein, as the division of the tail portion of the selected approximate function, the number of matching windows of the tail portion of the function is divided evenly. Method. 3. The entire approximate function selected by the function selection processing (2) is divided, and the ratio (histogram) of the original function of the matching window number distribution standardized at each divided point is calculated. , And the bias of separately calculating the ratio (histogram) of the function of the original matching window number distribution standardized at each of the divided points by dividing the function of the selected (or specified in advance) approximate tail part The calculation processing (3) is provided, and the ratio of the calculated skirt portion of the former (histogram)
The product of the area of the area and the area of the latter skirt (histogram) is divided by the width of the hem (division width) to obtain the height of the skirt (histogram), and the two are combined. After that, the ratio (histogram) after composition is corrected, smoothed, and corrected so that it changes smoothly as described above, and the function after correction is used to accurately estimate the identity verification. The personal identification estimation method according to claim 1, wherein: