KR0126464B1 - Finger print recognizing method - Google Patents

Abstract

A finger print recognition method discriminates a registered finger print and an input finger print, and counts an overlapped pixel number simultaneously moving vertically and horizontally a bitmap image expressing a finger pattern. The method detects a case that the overlapped area is maximized, detects a characteristic point indicating a similarity degree when the overlapped area is over a predetermined threshold value, and determines the same finger print on the basis of the similarity degree.

Description

Fingerprint Identification Method

Figure 1 is a schematic showing the principle of bitmap formatting.

(A) (b) of FIG. 2 is a schematic diagram showing the BMIN effect in bitmap formatting.

Figure 3 shows a 2 × 2 Mask schematic.

Figure 4 is a schematic diagram showing the principle of bitwise processing.

Figure 5 is a graph showing the irrationality of applying a statistical method.

(A) and (b) of FIG. 6 are graphs showing experimental probability distribution functions.

Sieve 7 is a graph showing the cumulative probability distribution function.

Figure 8 is a schematic of pseudo-feature points that appear in consultation.

Figure 9 is a flowchart of the restoration processing method.

Figure 10 is a schematic diagram of the restoration process.

Figure 11 is a schematic showing candidate elections for the removal of pseudo shortcomings.

Fig. 12 is a graph showing the restoration process for the pseudo branching point of the leg.

Figure 13 is a flow diagram of a method for enhancing registration by using two paths of preprocessing.

The present invention provides a fingerprint recognition method for comparing and discriminating a registered fingerprint and an input fingerprint when one device is used independently or when a plurality of devices are used at a short distance, and a situation where an equal sign is exchanged by a communication network such as a bank. The present invention relates to a fingerprint recognition method that minimizes the amount of registration data to be used in the present invention, and more specifically, a bitmap matching method capable of recognizing a fingerprint even in a noisy case by comparing binarized images. The patent application 89-17357 (US Patent Application No. 1618299, Japanese Application No. 325146) discloses a fingerprint recognition method using a bit-by-bit processing method that can process binary pixels at high speed by simultaneously processing a plurality of pixels. It is an improved invention.

In recognition of fingerprints, the most important thing is the processing speed and discrimination accuracy. The patent application No. 89-l7357 of the present applicant adopts a two-stage discrimination method to improve the processing speed and discrimination accuracy.

However, in the case of the printing by the feature points that occupy an important part of the printing, the reliability of the extracted feature points is greatly reduced when the fingerprint pattern is not clearly distinguished by a finger injury or noise such as wear or moisture. Characteristic points based on feature points are difficult to detect or have high error rates.

In addition, the binary image processing used in various stages of fingerprint recognition requires a lot of processing time, which is a bit-wise processing by a general purpose processor that refers to a storage device in bytes or more. Was not only very inconvenient but also quite inefficient in terms of speed.

The present invention compares the binarized images in pixel units to solve such phantom stylistic points, thereby processing binary images at a high speed by simultaneously processing a plurality of pixels and a bitmap matching method in which a fingerprint can be recognized even in case of noisy noise. It was developed a bit-by-bit processing method that can be processed, and in the case of using a multi-level matching method as in the present invention, each matching feature (Matching Feature) expressing the similarity between the registration fingerprint and the input fingerprint is obtained at each step The matching fitness should be appropriately judged.

Conventionally, a statistical method has been adopted for this purpose, but the pitching feature obtained in the present invention shows a completely different distribution from the general probability distribution function, and thus a new penalty method different from the statistical method is required.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

The bitmap matching method of the present invention is basically based on the fact that the same fingerprints will have a large number of matching pixels when two fingerprint images are overlapped. The position bozing value may have errors in pixel units, so the two fingerprints are moved by 4 pixels up, down, left, and right. When overlapping two fingerprints, check whether the area of the pixel to be compared is the background area or the boundary area (the area where no fingerprint exists in the image). Investigate how many pixels overlap each other using the bit-wise processing method described below for the telephone stations belonging to the area. The.

Through this penalty, the number of comparable areas that are not background or boundary areas in both fingerprints is RCNT, the number of overlapping areas per LC area is LCNT for the total comparable area, and the comparable total area. The sum of the number of overlapping pixels for an area is called SUM, and the average number of pixels overlapping per area is called SCORE (SUM / RCNT).

This redundancy is performed by moving up, down, left, and right by 4 pixels around the direction difference of the two fingerprints, and in each case SCORE (i), RCNT (i), and LCNT (i). (i = 1… CNT, CNT is the total number of trials).

Through these values, we can get the comparison characteristic that shows the similarity between two fingerprints as follows.

BMIN = min {SCORE (i) i = 1 ... CNT}

BMAX = max {SCORE (i). i = 1 ... CNT}

BVAR = {∑ SCORE (i) * SCORE (i)} / CNT- {MEAN (i) * MEAN (i)}

Where MEAN (i) = {∑ SCORE (i)} / CNT, (i = 1 CNT)

BRATlO = MAXLCNT / MAXRCNT

Where MAXLCNT = maX {LCNT (i), i = 1... CNT}

MAERCNT = RCNT (max), max = {j: LCNT (j) = MAXLCNT}

That is, BMAX is the largest value among SCORE (i), and in the case of the same fingerprint, the value of BMAX is naturally increased.

On the other hand, BMllN is the smallest value of SCORE (i), and in the case of the same fingerprint, the value tends to be considerably smaller, as shown in the figure 2.

In addition, in the case of the same fingerprint, the value of SCORE (i) shows a big change, as can be seen as the difference in value between BMAX and BMlN increases. Therefore, BVAR, which represents SCORE (i) variance, is also an important factor in determining whether the same fingerprint is present.

BRATlO, on the other hand, is the ratio of the area where the two images overlap best (when SCORE (i) becomes the maximum), which determines whether a part of the image coincides overall, not intensively overlapping. It becomes a reference value and also has a high value in the case of the same fingerprint.

The bitmap matching method judges whether the same fingerprint is considered in consideration of the above comparison characteristics. Since this method uses all pixels as the basis of the comparison penalty, the method of extracting the feature points and performing the comparison penalty on the basis of the comparison penalty is compared. It is more resistant to noise.

Next, the bit unit processing method for speeding up the binary image processing will be described. The basic principle of the bit unit processing method is that all arithmetic logic operations of a computer can be represented by combinatorial logic. By using the combinational logic of the binary image processing used and applying it to the bit image, it is possible to speed up the multiple bits by processing them simultaneously. For example, as shown in Figure 3, four pixels (2 × 2) If the value of) is equal to or greater than 3, 1 is output. Otherwise, 0 is output. In the conventional method, lF ((A + B + C + D) = 3) output = 1, EISE output = For the processing of 1 pixel, because it is performed in the same way as 0,

4 READ

ADD 3 times

Comparation 1 time

1 write (WRlTE)

Operation is required.

On the other hand, if the above penalty is composed of combinatorial logic, a combinatorial logic such as body 4 is obtained.

F (A, B, C, D) = ABC + ABD + ACD + BCD

In addition, the present invention uses the following method to perform the above process, where A _1, B _1, C _1, D ₁ (i = 0, ... n) is each pixel value, 0i Is the processing result for the pixel, and A, B, C, and D are word data representing successive 16 pixels, respectively, and indicate the following data.

A = word data of the original bit image

B = Weed data of left 1 bit shifted image

C = A Word data under one line

D = B Word data one line below

If you enter it as combinatorial expression at the same time

4 READ

AND product 8 times

OR 3 times

Write once

1 bit shift

By processing the result of processing 16 pixels (when one word is 16 bits), the processing speed is remarkably improved compared to the conventional method.

At this time, pixels corresponding to the data bus size processed by the general-purpose operator are processed in parallel, and in the case of the MC68000, the 16-bit general-purpose operator adopted in the present invention, processing for 16 pixels is performed in parallel.

As stated above, the fingerprint recognition method of the present invention uses a multi-step matching method, and thus, a direction pattern. The use of various features such as bitmap images and feature points (Minutiae) are used to determine whether the same fingerprints are used, and the use of such a variety of matching features must be very helpful in recognizing fingerprints with various characteristics. In order to use this, a study on how to make final judgment efficiently by integrating multiple matching features is essential.

As such, a lot of researches have been conducted in statistics on how to determine which class (C1ass: same or other) of a fitness vector that contains several elements. It is known that the most probabilistic judgment can be made by using the method.

However, in order for such a traditional method to be applicable, each element is basically required to be distributed according to a gonggyu distribution or a known probability distribution function. However, an experimentally obtained matching score is usually different from a feature probability distribution function. This statistical approach did not provide satisfactory judgment results. Moreover, in this case, the average of the class was found to be ideal because the general probability theory was applied as it is. For example, as shown in Figure 5, when the score is 100%, the probability of the same fingerprint is the largest, but the probability of the percentage of M%, which is the average of the same fingerprint, is more likely to be the same. It implies irrationality such as being evaluated.

Therefore, in the present invention, a characteristic probability distribution function can be used by experimentally obtaining a stochastic distance to an ideal state of each class (same fingerprint or cognitive fingerprint) in order to obtain a proper final result of the irregular distribution as described above. In the absence of this, they have developed a way to make a comprehensive judgment.

To describe the method developed in the present invention in more detail, a FIT vector representing the similarity between two fingerprints is called F, F is _{n elements} from f ₁ to f _n (15 in this fingerprint recognition method). If that is configured obtains the empirical probability distribution to obtain order of M _d with respect to the case of the same M _s of pitchwo vector, other fingerprints for the case of a fingerprint, by performing much of the maetching experiment for each case of the class pitchwo vector .

The following distribution is obtained for each element of the fitness vector using the fitness vectors obtained by the experiment.

First, divide the range of values of each element into k uniform ranges, and obtain a distribution such as sieve 6 (a) (b) by the frequency of the elements of the fitness vectors of each class in each range. By calculating the relative probability of occurrence, the probability that the elements E _{s, i} of the same vector's feature vector are in the k-th interval can be expressed as follows.

P _{s, j} (k) = Q _sj (K) / Ms

(However, Q _{s, j} (k) is the number of _{vectors in} which the i th element of the feature vector of the same fingerprint obtained experimentally is included in the k th interval)

Similarly _{, the} probability that the element E _{d, i} of the fitness vector of the cognitive statement exists in the k-th interval can be expressed as follows.

P _{d, j} (k) = Q _{d, i} (K) / Md

(Where Q _{d, i} (k) is the number of vectors in which the i th element belongs to the k th interval among the Feature Vectors of other advisors obtained experimentally)

By using the experimentally obtained probability distribution function as described above, a judgment with high accuracy can be performed without using the feature probability distribution function. However, since the probability distribution function is obtained by a finite number of experiments, An irregular distribution, such as ringing in a feature area or a very small value in a feature area, may occur.

In addition, as described above, when the matching score is better than the average, the existence probability tends to be lowered, and thus, irrational aspects are not solved. Thus, in the present invention, instead of using the experimental probability distribution function, the ideal state of each class is used. The stylistic point is solved by using the cumulative probability distribution, which is the stochastic distance of, and the cumulative probability distribution function is conceptually equivalent to the area of the hatched portion of body 7 degrees.

In the case of the same fingerprint, in the case of the figure where the value of the element increases (the ideal value of the same fingerprint is K)

In the case of the same fingerprint, in the case of the figure where the value of the element becomes small (if the ideal value of the same fingerprint is 0)

Based on the cumulative probability distribution obtained above, the final penalty is calculated by calculating the distance Ds to the same fingerprint class and the distance Dd to the cognitive fingerprint class for the given figure vector F, respectively.

(W1 is a weighting factor that indicates the importance of each element, and k indicates that the element belongs to the k-th period when each element of the input feature vector is divided into k sections.)

The final penalty is as follows:

In the case of D _s + T _h D _d , the same fingerprint is used. Otherwise, the same fingerprint is used.

Here, Tb is a constant that acts as a threshold value. In general, fingerprint recognition recognizes the same fingerprint as the same fingerprint, so it is much more dangerous.

The following relates to fingerprint recognition methods suitable for use over a telecommunication network. The methods stated include Direction, Pattern, Bitmap Matching. Although the accuracy of the sentence is increased by using various characteristics of fingerprints such as feature points, it is not desirable to have a large amount of data in order to register a fingerprint when data must be exchanged through a communication network. A fingerprint identification method is required to provide sufficient punctual accuracy by registration data. In the present invention, by using only the feature point (Mlinutiae) among the various data of the advisory, the amount of registration data is reduced to about 200 bytes, and the following stylistic points must be solved in order to realize fingerprint recognition using only the feature point. .

1) The method of complementing the difference in the position and direction existing between the enrolled fingerprint and the input fingerprint

2) Consideration to extract feature points accurately

3) Method of screening fingerprints that cannot be processed by this recognition method at registration step and reinforcing the appearance of feature points.

In the conventional method, the position matching can be done simply and efficiently by attempting template matching between the direction pattern tables. However, in the present invention, since only the feature point is included in the registration data, the above method is not applicable, so We developed a method based on the shift-invariant feature that enables position complementing according to the correspondence relationship.This method applies only the matching by the feature point, so the accuracy of the feature point is very important. Therefore, it is necessary to extract feature points accurately at the expense of some processing time.

In the present invention, for this purpose, by applying a filtering on the original image known in the prior art and developing a restoration method for the line image after thinning, the accuracy of feature point extraction is improved. To increase.

The restoration processing method used in the present invention will be described later in the restoration processing method for the removal of pseudo feature points, and as described above, a pattern in which a fingerprint has a large wound, a severely dried, or a wet fingerprint is very bad. Since it was poor, it was almost impossible to judge by the feature point. Therefore, such fingerprint should be warned that the use of the fingerprint recognition method of the present invention is impossible at the registration step, and only the matching by the feature point is used. There is a need to enrich the collection as much as possible, and a detailed description of the advantages is a method for reinforcing fingerprint registration penalty, which will be described later.

Even if the image is obtained from the same fingerprint, the fingerprint obtained every time has a difference in position. Therefore, in order to recognize the fingerprint and determine whether it is the same fingerprint, the difference must be compensated. In recognition, for this purpose, the method of finding the center point of the fingerprint and preserving the location based on the point is mainly used, but it is not easy to extract the center point from the fingerprint image automatically and accurately, and the center point does not exist like the arch. In some cases, it was difficult to apply in general.In addition, by constructing a tree with feature points like MST (Minimal Spanning Tree), an algorithm that can perform comparative panning regardless of rotation and parallel movement is proposed. If present in the overall Demonstrate a weak propensity to noise or the like is stronger than a physical structure variations to the noise, generally applicable position bojing method developed is required.

Therefore, in the present invention, in order to meet such demands, a pair corresponding to the feature point list of the registration and input fingerprints using a property that is not affected by the rotation (D0) or the ping-sin movement (DX, DY) among the characteristics of each feature point By finding and monitoring the watch by its correspondence, each feature point is provided independently without relying on the feature points such as the center point, thereby developing a position bossing method that is not significantly affected by the presence of pseudo feature points. It obtains the feature points through preprocessing, smoothing, binarizing, thinning, and restoring the circular image received from the camera. The feature points have various characteristics depending on their position, orientation, and relationship with the surrounding feature points.

In the position retaining method of the present invention, such characteristics are divided into three categories as follows.

1) Che 1 class

Properties that do not depend on differences due to translation or rotation

2) Body 2 class

Properties that depend on rotational differences but not on translational differences

3) Body 3 class

Properties affected by both rotation and translational differences

To describe in more detail the flow of the position boss method of the present invention as described above,

1) Each feature point of the registered fingerprint feature point list is compared with each feature point of the input fingerprint list, and the pairs matching within the range of characteristics of Che 1 class are elected as primary candidates.

2) If the primary candidate extracted in l) is a correct pair, the characteristics of the sieve 2 class of each pair, that is, the characteristics that depend only on the rotation, will appear where the difference is concentrated. Using this property, we can see the distribution of the difference between the two class characteristics of the single-character candidate pairs, and select the place where it is concentrated as the rotation value of the registration and input fingerprints. Delete from inging and pair list. At this time, the list of pairs remaining without being deleted is called a secondary candidate pair list.

3) The difference between the registered fingerprint and the input fingerprint by the rotation of the input fingerprint by the rotation value obtained in 2).

4) As in the case of 2), if the secondary candidate pairs are pairs that correspond correctly, since the rotation is complemented in 3) for the characteristics of the Che 3 class, that is, the properties that depend on parallel movement and rotation, the parallel movement is performed. Will be collected at the place where the car of the Using this property, we select the concentration of the difference between the registration and input fingerprints as the parallel movement between the registration and input fingerprints. Erase. At this time, the list of pairs that will remain without being deleted is called a final pair list.

5) In the above, the effect of parallel movement and rotation was complemented, and the purpose of position boring was basically achieved. However, in the fingerprint image, distortion may occur due to the pressing of a finger. There is a possibility of lack. Considering such a point, we apply the Least Square Error Method to the final pair list to find the translation and rotation values that reduce the conversion error between pairlists.

To explain the necessity of the restoration processing in the restoration processing method for removing the pseudo feature point, the input fingerprint image inevitably includes some noise in the image due to the limitation of the performance of the input device and the state of the fingerprint. The noise reduction and the preprocessing penalty to improve the quality of the fingerprint image have to go through, but there is still a lot of noise left in the final thinning image needed to find the feature points, and the pseudo feature points are distributed with the feature points.

Therefore, the processing penalty for protecting the feature points and removing the pseudo feature points is called the restoration process. In this class, the rule-based adaptation method is used to streamline the application order among the restoration processes. In addition, the real image itself is restored to improve the efficiency of subsequent processing. By using the direction values of the local area, the accuracy of the restoration is increased, and the processing time is reduced by the efficient search and discretization method. In terms of body 8 degrees, the types and shapes of noise appearing on the fingerprint images are as follows.

1) SHRT LlNE (Refer to Body 8 (a))

2) LOOP (see sieve 8 (b))

3) SHORT BRANCH (refer to Figure 8 (c))

4) BRIDGE (see FIG. 8 (d))

5) CUT LINE (See Fig. 8 (e))

Figure 9 shows the overall flow of the restoration process, and its contents are the same as Figure 10.

1) Type of feature points

The form is determined by obtaining the number of changes (CROSSlNG COUNT) of the peripheral pixels from the 3x3 mask in which the center pixel is the ridge region (i.e., X (0) = RlDGE).

TYPE = ∑│X (i) -Xi (i + 1) │ / 2, where i = 0,1,... 7

If TYPE is 1, it is disadvantage, if 3 is branch point, 4 sides intersection

2) Justice and treatment of the curvature of the shortcomings

The slope of the line connecting the starting point and the ending point is called `` CURVATURE '' by tracking the range of distances (currently 20 pixels) from each of the disadvantages. After finding the curvature of the point as the end point, return with a value such as `` ENDRTN '' or `` BIFRTN ''.

In this case, in case of short-term regression, the type of the characteristic point that is the point of time is represented by the short-term short-term (E-CUTTED) '', and in the case of branch point regression, the short-term regression and beard (to be described later). It can be used in the restoration process for.

3) Candidate election for the removal of pseudo shortcomings

For each disadvantage E (i), for the disadvantages where the distance d from the other disadvantage E (i) is within the working distance (current 20) and the pier m between the curvatures of the two disadvantages is more than the working angle (current 90) Calculate the element (see also body 11)

α = slope of the connecting segment of the two disadvantages

β (i) = piers of curvature of α and E (i)

Difference between δ = α and Local Direction of the region to which point M (i, j) belongs

Where M (ij) is the midpoint of E (i) and E (j)

D = 10β (i) + 10β (j) + 10δ + 100d

Where the coefficients of each element are weighted with respect to D

E (j) having the smallest value among Ds obtained for each E (j) is indicated by the reconstruction code of E (i).

4) Restoration process for cutting line

As a result of obtaining candidates for restoration of each shortcoming from the punishment for the elimination of pseudo shortcomings, it is judged that the pair of shortcomings extracted from each other as candidates can be restored, and these are called pseudo-feature points (PSEUDO MlNUTlAE). Connect two points.

However, restoration is stopped in the following cases.

i) δ defined in the penalty (3) is greater than or equal to the daily value (currently 50), or

ii) Searching for pixels in 8 directions while connecting two points and encountering ridged pixels instead of target pixels

5) Restoration treatment for disconnection and beard

With the image and the feature point list restored first by the cutting line restoration process. Among the shortcomings, re-tracing is carried out only on the points judged as shortcomings of the disconnection and the shortcomings of the beard in exaggeration (2). If two points are connected in the line restoring process, it is not already a disadvantage to reconfirm the mold at the position of the two ends, so it is determined as a pseudo feature point. '', The half of this value (10) is used, which may vary depending on the whiskers and disconnections. Do not delete the focus point. In case of returning to the shortcomings of the tracking result, it is judged as a short disconnection and the image is removed from the viewpoint to the midpoint.

6) Discipline and Treatment of Curvature at the Junction

Unlike the shortcomings, there are three traceable directions unlike the shortcomings, so the question is which one of the lines should be defined as the curvature of the bifurcation. The choice is as follows. After considering them as α (i), calculate the following factors.

δ (i) = │ α (i)-α (i + 1) │

α (i) = ∑│δ (i) -δ (j) │, where j ≠ i

0≤j3j

Of all [alpha] (i), i (i + 2) mod3 of i having the minimum value becomes the main line (STEM) of the bifurcation point. The direction of the main line is defined as the curvature of the branch point.

7) Care and reconstruction of small loops

As a result of tracing in each direction of the branching point, when two directions of the branching point regress at the same branching point, it is determined as a branching point forming a small stream, and one of the two tracking paths is removed from the image.

8) The treatment of the bridge (BRIDGE) and its restoration

A bridge is a form of noise that runs between ridges and neighboring ridges, almost perpendicular to their direction of travel. This type of noise appears relatively more than other noises and is often concentrated in a complex manner, and thus takes up a large portion of the reconstruction algorithm.

That is, in the current algorithm, when starting from the branch point and returning from the branch point as shown in Fig. 12, the direction of the branch is obtained and the difference from the sailing direction dp of the average direction (davg) of the local region to which the viewpoint belongs ( θ) to determine that the branch with the lowest value is the most likely 'leg' starting at the current fork, from which the value is below the established threshold (currently 50), and this is the branch of the bridge (B-BRANCH). Mark as θ and record θ at this time.

For all branching points, the following processing penalties are executed in sequence for all the points identified as the branching points of the bridge where the above penalty is processed.

i) Ascending order based on θ.

As a result of the alignment, the smaller the θ of each point, that is, the branching point having the branch most similar to the shape of the bridge connecting the ridge, is located at the front of the list.

ii) Re-track starting at each branch in the order of order.

The starting direction of the trace refers to the previously recorded direction value, in which case it is necessary to reconfirm the current fork before the trace, because there is a possibility that the image may change through a series of bridge strengths that follow. If the reconfirmation result is not the branch point, the result is that the bridge started from this branch point is erased earlier.

If it is also found as a bifurcation point, this point is tracked in the direction of progress of the previously recorded leg (20) to determine the advantage as a pseudo feature point when the tracking result is a bifurcation regression. Tips up. If the divergence point is not met even after the exhaustive distance has been met, this point is classified as the bifurcation point because the noise reduction of the point where the noise is more likely to be noise in the previous restoration penalty is less likely to be the pseudo feature. do.

9) Final reconfirmation penalty at the fork

For the points classified as bifurcation, the sentence is reconfirmed based on the updated image. The intersection point (when the number of branches is 4) is determined as a pseudo feature point.

Finally, to explain the method for strengthening the registration penalty of the advisory, the reliability and anxiety of the registration notice may be stylized due to the noise included in the fingerprint image or the coarseness of the pattern itself. In order to improve the quality of the image and the culling, various types of image processing work are performed in the first half of the process, which is called preprocessing. Different preprocessing penalties bring different results to the same image example. Finally, the list of feature points finally obtained varies depending on whether or not they have been subjected to the present invention. In the present invention, two paths are preprocessed on the same fingerprint image as shown in Fig. 13, and the difference between the two paths is a binarization method.

1) Path (1) uses local binarization, which shows the shape of fingerprints relatively cleanly and clearly, but is weak in areas with poor contrast.

2) Path (II) uses circular mask binarization, which shows the shape of the fingerprint well in the region with poor contrast ratio and maintains the ridge connectivity, but the shape of the fingerprint image is relatively There are stylistic points such as changing the connectivity by connecting rough and different ridges.

Therefore, it extracts the feature points (Minutiae) from each of the two regions through the above two paths and uses them as a sign for reinforcing registration, and the list of two feature points obtained through different preprocessing paths is slightly different in the distribution and shape of the feature points. However, this difference may vary depending on the type of noise or fingerprint pattern included in the input image, which can be used as an evaluation data for the image stability. In other words, the higher the Anjing degree image, the smaller the difference seen in the two lists. Therefore, in the present invention, the registration penalty is settled by using this, and the registration of the very low fingerprint fingerprint is not allowed.

In addition, the distribution of feature points and forms of retribution can be regarded as having a probabilistic character, and weights are added to feature points that are identified as having the same or more likely to exist in the list obtained through different processing penalties. The weighting method used in the present invention and used in the present invention is as follows.

Priority applies

1) Point each other as a matched pair,

If the types match

2) Point to each other in combination pairs, but the types of feature points do not match

3) A combination pair is found unsuccessfully

4) Combination pair does not exist and exists only on one side

In addition, the simple combination method using the feature points has calculated the similarity between the corresponding feature points as the combination rate, but this method is insufficient to identify the overall topological structure to which each feature point is connected.

However, the phase structure of the feature points can be used as another feature of the fingerprint. In the present invention, the connectivity of the feature points is defined as a new feature and used in combination, and the connectivity tracks the ridges starting from each feature point. It refers to the connection status with other feature points that are encountered. The elements used to express it are as follows.

1) The type of other connected feature points

2) the distance to that point

3) the direction of the point

Combination method using the above elements can be divided into two as follows.

1) Calculate the combination rate for the connecting elements of the corresponding feature points.

Combination rate (M) = W1 * type match + W2 * distance match + W3 * direction match

Here, Wi is the weighting factor

2) Obtain the maximum matched component and use the weighted proportional to the size of the linker in the combination result.

When all three factors such as type, distance, and direction coincide in calculating the combination ratio as described above, the connection relationship at the point is reconsidered and repeated until all three elements of the penalty are inconsistent. The final obtained form is called the maximum combination connector, and the size of the maximum combination connector can be expressed by the number of edges connected, and a value proportional thereto is used as a weight.

As described above, the present invention uses a bitmap matching method to compare binarized images at the minimum unit, and the ridges and curved portions of the fingerprint are distinctly distinguished by noise such as moisture according to a severe wound on the finger or wear or an individual. Even if it is not possible, it is possible to recognize the fingerprint securely, and by processing the binary image at a high speed by using a bit unit processing method which processes a plurality of pixels simultaneously, the efficiency can be expected in terms of speed. The effect is very large.

Claims (6)

A method of comparing and comparing a registered fingerprint and an input fingerprint by counting the number of overlapping pixels by moving a bitmap image representing a pattern of a fingerprint slightly up, down, left, and right, wherein the overlapped area is A fingerprint recognition method comprising extracting feature points representing similarities, such as the maximum case and the proportion of the areas where the variance and overlapping areas therebetween are equal to or greater than the ilzing threshold value, and whether or not the same fingerprint is determined based on the value. In the method for performing various purposes processing of noise reduction, linear image extraction, direction pattern extraction for the binary image represented by 0 and 1. A method of fingerprint recognition, comprising: expressing a plurality of pixels in a word that is a basic unit referred to by a general-purpose operator, and expressing the contents of the process by a combinational logic to simultaneously perform the processing on the plurality of pixels, thereby speeding up processing. Given a number of feature points expressing the similarity of the fingerprints in the body 1 term, the probability distribution function experimentally obtained as a means for making an accurate judgment on whether the same fingerprint is valid, and the probability from the ideal identical fingerprint to the ideal cognitive fingerprint on the function A fingerprint recognition method characterized by increasing the accuracy of the plate by using the cumulative probability distribution function obtained by calculating the respective distance. A method of extracting the difference between parallel movement and rotation in a method for compensating the effects of parallel movement and rotation existing between a registered fingerprint and an input fingerprint based only on the characteristic points of the fingerprints in the first and third paragraphs and their characteristics. Among the characteristics, the effect of the rotation is extracted by extracting the initial pair by the shift-invariant feature, and then calculating the figure affected by the rotation between the pairs. In order to compensate for the effect of the parallel movement by calculating the difference of the fitness affected by the shift, and to apply the least square error method to the pair of feature points finally obtained through such a penalty, the registration fingerprint and input Fingerprint recognition method characterized in that the effect of the parallel movement and rotation existing between the fingerprints. In order to remove the pseudo feature points from the initial feature point list extracted from the line images of the body of paragraph 4, and to apply the rule for sequencing with respect to various types of pseudo feature points to leave only the actual feature points. And a local directional pattern table is used to increase the accuracy of the pseudo-pointing point determination. The similarity in the method of comparing the feature list obtained by using two preprocessing paths to enhance the anjing property of the registration data to confirm whether they are the same by using the pseudo-feature points of Article 5, and to enrich the registration data. Based on the registration data, it is automatically determined whether it is a valid fingerprint image, and if it is found to be valid, weighting is applied to the overlapping feature points, thereby increasing the punishment of the feature points, and including the punishment regarding the connectivity between the feature points in the registration data. Fingerprint recognition method characterized in that to obtain registration data including a more unsettled by.