CN100414558C - Automatic Fingerprint Recognition System and Method Based on Template Learning - Google Patents

Abstract

The present invention relates to an automatic fingerprint recognition method based on template learning. The present invention comprises the steps of fingerprint image register, namely recording characteristic information of a fingerprint template and a fingerprint image to be matched; fingerprint image recognition, namely judging whether the characteristic information of the input fingerprint image is similar to some template information stored in a system database; detail information feedback, namely utilizing knowledge rules to feeding back the matched fingerprint image information to complete reliable detail point classes to which multiple templates correspond. In the present invention, the detail points of a plurality of templates of the same fingerprints are classified, the cores of the classes and the information of inter-class distance, class-in distance, etc. are stored, and a link of coarse matching is added; of the coarse matching fails, the ordinary matching operation is not carried out. The present invention can accurately extract and repair the characteristic information of multiple plates, enhances the recognition effect and performance of an automatic fingerprint recognition system and has an important application value in biologic recognition technology.

Description

Automatic Fingerprint Recognition System and Method Based on Template Learning

technical field

The invention belongs to the field of biometric feature identification, in particular to realizing the matching process of multi-template fingerprint images by cluster analysis and knowledge-based methods.

Background technique

In the 1990s, fingerprint identification technology gradually became a mature biometric identification method, which belongs to the field of "pattern recognition". First, input the extracted fingerprints into the computer, and then through a series of complex fingerprint recognition algorithms, anyone can be identified and authenticated based on the fingerprints in a very short period of time. At present, automatic fingerprint identification systems are widely used in fields that require identification. The application of automatic fingerprint identification is no longer limited to the field of law and public security. It can be used as a means for computers to confirm users, as an information security technology for accessing network resources, and can also be used to confirm the use of bank ATM cards and credit cards, and various smart ICs. Card double confirmation, employee certification and home electronic door locks and many other aspects. With the wide application of automatic fingerprint identification systems in civil fields such as access control, time attendance, and social security, people have put forward higher and higher requirements for the accuracy of fingerprint identification.

The composition of a typical automatic fingerprint identification system is shown in Figure 1.

A typical automatic fingerprint identification system mainly includes several modules such as fingerprint collection, feature extraction and matching (A.K.Jain and S.Pankanti, "Automated Fingerprint Identification and Imaging Systems", Advances in Fingerprint Technology, 2nd Ed. (H.C.Lee and R.E.Gaensslen), CRC Press, 2001.). From the perspective of modern scientific research, the identification methods of this type of system mainly involve issues such as fingerprint image acquisition, fingerprint image enhancement, feature extraction, data preservation, and feature matching. Among them, feature extraction and fingerprint matching are two core problems of fingerprint recognition system, and they are also two basic problems and important research topics of pattern recognition.

According to the standards of the US FBI, fingerprints are usually divided into ridges and valleys, and the end points and bifurcation points of the ridges are considered to be invariant and unique throughout life, as shown in Figure 2.

Therefore, the commonly used fingerprint recognition algorithm is to extract the terminal points and bifurcation points of the fingerprint image as feature information, and then perform the rotation and translation calibration between the fingerprint feature point sets, and then calculate the similarity through the feature points. A.K.Jain, an authoritative scholar in the field of international biometric identification, led the image processing and pattern recognition research group of the University of Michigan in the United States to conduct in-depth research on this, and applied for a number of US invention patents, such as: U.S.Patent 6,185,318, Feb.6, 2001.

The method proposed by them has higher requirements on the template information of the fingerprint. It requires that the number of fingerprint minutiae points contained in the template should not be too small, and the information such as position and direction must be accurate. When the effective area of the fingerprint template collected is small, the number of detail points of the fingerprint template is less (for example: the collector of Authentic company can only collect the image of 128 * 128 pixels, and the fingerprint image collected can only be a part of the actual fingerprint, like this It is difficult to match the extracted minutiae of the image; when the quality of the fingerprint image is poor, there may be false minutiae in the extracted minutiae such as position and direction. So in the above two cases, the methods they adopt are prone to misrecognition or rejection. Therefore, in the automatic fingerprint identification system, according to the matched fingerprint information, removing false minutiae points and adding more accurate fingerprint features to repair the fingerprint template is an effective way to improve the identification effect, and it is also a problem to be solved.

Contents of the invention

The purpose of the present invention is to propose and design a practical automatic fingerprint identification method and system. It can process multiple template images of the same fingerprint, fuse the common information of multiple templates into the overall multi-template detail information, and then apply this type of information to the matching algorithm of fingerprints; at the same time, it can feedback matching during use The information on the fingerprint image is used to learn and train the parameters of the multi-template detail class information. Make template information rich and accurate, and reduce rejection and misrecognition caused by individual differences in templates.

In order to achieve the above-mentioned purpose of the invention, according to one aspect of the present invention, the automatic fingerprint identification method based on template learning includes steps:

(1) Fingerprint image registration based on multi-template detail information clustering, recording the feature information of the fingerprint template and the fingerprint image to be matched, including:

Multiple template image calibration;

Multi-template detail information clustering, calculate the core of reliable detail points corresponding to multiple templates and the distance between classes within a class;

(2) Fingerprint image recognition based on detailed information feedback, using knowledge rules, feedback matching fingerprint image information, and judging whether the feature information of the input fingerprint image is similar to a certain fingerprint template information in the system database, including:

Rough fingerprint image matching, judging the similarity between the minutiae feature point set of the fingerprint image and the multi-template minutiae class;

The fingerprint image is finely matched, and the matching result is output;

Detailed information feedback.

According to another aspect of the present invention, an automatic fingerprint recognition system based on template learning includes:

(1) The fingerprint image registration module based on multi-template detailed information clustering records the feature information of the fingerprint template and the fingerprint image to be matched, including:

Multiple template image calibration;

Multi-template detail information clustering, calculate the kernel and intra-class inter-class distance of reliable detail points of multiple templates to English;

(2) The fingerprint image recognition module based on the feedback of detailed information uses the knowledge rules to feed back the matched fingerprint image information to determine whether the feature information of the input fingerprint image is similar to a certain fingerprint template information in the system database, including:

Coarse fingerprint image matching, judging the similarity between the minutia feature point set of the fingerprint image and the multi-template minutiae class;

The fingerprint image is finely matched, and the matching result is output;

Detailed information feedback.

The present invention classifies the minutiae points of multiple templates of the same fingerprint, records information such as the cores of these classes and the intra-class distances between classes, and adds a rough matching link. If the rough matching is unsuccessful, the general matching operation is no longer performed. The invention can accurately extract and repair the feature information of multiple templates, improve the recognition effect and performance of an automatic fingerprint recognition system, and has important application value in biometric technology.

Description of drawings

Figure 1 is the composition of a typical automatic fingerprint identification system;

Figure 2 is the fingerprint feature;

Figure 3 is a schematic diagram of an automatic fingerprint identification system with feedback

Fig. 4 is the fingerprint image that same finger is collected multiple times;

Fig. 5 is the composition of the automatic fingerprint identification system based on template learning;

Fig. 6 is the enhanced processing flow of fingerprint image;

Fig. 7 is an 8-connected neighborhood diagram of M points;

Figure 8 is a detail point model;

Figure 9 is a detail point class diagram and kernel features;

Figure 10 shows several possible results in the clustering process, in which, picture a is a set of similar points in the same class; picture b is a set of points with dissimilar angles; picture c is an isolated point; picture d is a pair of similar points but in image edge or out of bounds

Figure 11 is a fingerprint image recognition system designed and realized by itself;

Figure 12 shows the generation of multiple templates for the same finger, where the test data is a fingerprint image with a resolution of 300×300×256. Picture a, picture b and picture c are the template images collected three times by the same finger respectively; picture d is the calibration image of picture a, picture b and picture c; picture e is the core set of multi-template clustering of picture d;

Figure 13 is an example of the identification process;

Fig. 14 is a statistical ROC curve diagram of the matching result;

Figure 15 is the FVC algorithm evaluation results.

Detailed ways

Our method proposes a new form of template representation, and the successfully matched fingerprint information can be used to repair and improve the fingerprint template, that is, to add a feedback link in the traditional automatic fingerprint identification system, as shown in Figure 3. At the same time, our fingerprint image recognition method can remove the pseudo minutiae of the original fingerprint image due to the influence of noise, record the feature information of the fingerprint as accurately as possible, and can identify and process fingerprint images with poor quality or few minutiae. Therefore, our fingerprint image recognition method adapts to the requirements of the miniaturization of new fingerprint collectors and ensures the robustness of the automatic fingerprint recognition system.

An efficient automatic fingerprint identification system should be able to perform real-time matching and identification operations on large-capacity fingerprint databases. The usual method is to classify the input fingerprints according to bucket type, left-handed type, right-handed type, arch type, pointed arch type, etc., and then perform one-to-one matching in the fingerprint database of the class. For this problem, our method solution strategy is: according to the special template representation method, two steps of coarse matching and fine matching are designed. In the template library, only the templates with successful rough matching and the minutiae point set of the input fingerprint are finely matched one-to-one, and only the templates with successful fine matching can participate in parameter learning and repair. Such operations make our method efficient and accurate.

Due to various reasons such as the fingerprint collector and the person to be collected, sometimes the fingerprint area collected by the fingerprint sensor is relatively small, and such a fingerprint cannot provide sufficient information, such as minutiae, for an automatic fingerprint system with a high recognition rate. Moreover, sometimes only a small part of the fingerprint images collected by the same finger may overlap with each other, which will also affect the matching performance of the fingerprint recognition system. Therefore, during the registration process, the fingerprints of multiple angles continuously collected by the same finger are stored as templates, which is a measure to improve the matching rate of fingerprints. As shown in Figure 4, it is a fingerprint image collected multiple times by the same finger.

However, using each registered fingerprint independently as a template does not take into account the correlation and common features between these templates, which will result in massive growth of redundant information in the template database and waste of resources. How to extract the public information of multiple fingerprint templates and remove the false information of individual templates is a problem that needs to be solved in order to improve the system identification performance. Our method adopts the improved Clique graph and other clustering analysis methods to fuse the information of multiple fingerprint templates for fingerprint identification, and the implementation results show that the method is practical and reliable.

The core idea of the present invention is to adopt an effective clustering analysis method and a knowledge-based training and learning method and use a computer to simulate the method of manual fingerprint image recognition. Since the fingerprint image has its own characteristics, there are two main prior knowledge that can be used for identification and matching, one is the distribution of the fingerprint minutiae position, and the other is the texture near the fingerprint minutiae. The distribution of fingerprint minutiae positions among multiple templates of the same fingerprint and the nearby texture directions corresponding to fingerprint minutiae points are similar, but for images collected by different fingers, they are completely different. We can summarize the commonality of these visual information among multiple templates of the same fingerprint, and express the structural information such as the distribution of fingerprint details in the computer, and then in the subsequent identification process, we can accurately distinguish the similarities and differences of fingerprints. It is necessary and possible to introduce people's understanding of fingerprint structure into the process of fingerprint image matching, and it is necessary and possible to use a computer to simulate manual image matching. This image matching algorithm is based on people's understanding of the fingerprint structure in the form of rules (that is, the two main prior knowledge of the fingerprint image) and uses the structure information of the fingerprint image to guide the image matching process.

The design of fingerprint image matching algorithm and recognition system based on template learning is described in detail below. As a specific identification system, the main modules are: fingerprint image registration module, fingerprint image recognition module and feedback module. For the specific recognition algorithm, the main steps are: minutiae point extraction, multi-template detail information clustering, fingerprint image rough matching, fingerprint image fine matching, and multi-template detail information restoration. The following introduces them one by one.

As shown in Figure 5, the modules of the system are mainly divided into a fingerprint image registration module, a fingerprint image recognition module and a feedback module.

Fingerprint image registration means that in the process of offline fingerprint collection, each fingerprint needs to collect multiple fingerprint images with good quality as templates (generally, 3≤N≤10), extract and record the detailed feature information of fingerprints, and store them in the in the template database.

The main processing processes of this module are: image acquisition, image processing and detail point extraction

Step 1: Acquisition of Fingerprint Image

There are two collection methods: ink pressing and instrument collection. A certain optical sensor, CMOS fingerprint sensor, thermal sensor, ultrasonic sensor and other new sensors can be selected as the device for fingerprint collection. It is required to place the area where the center of the singular point of the fingerprint is located horizontally in the center of the acquisition chip as far as possible, and press with a certain force.

Each fingerprint needs to collect N fingerprint images with better quality as templates, generally, 3≤N≤10.

Step 2: Enhancement processing of fingerprint image

Fingerprint image enhancement refers to the process of using some image processing means to process the fingerprint image. In our fingerprint algorithm, this step is critical. The processing flow is shown in Figure 6.

The specific processing operations include: 1. Equalization of gray scale, which can eliminate the contrast difference between different images. 2. Use a simple low-pass filter algorithm to remove speckle noise and Gaussian noise. 3. Calculate the boundary of the image and crop the image. This reduces the computational workload in the next step and increases the speed of the system. 4. Estimation of the direction field, calculating the direction of each pixel of the fingerprint image. 5. Binarization, according to the direction of each pixel, the fingerprint image is processed into a black and white image. 6. Thinning, according to the binarized image, the ridge line width of the fingerprint is thinned to only one pixel, and a fingerprint thinning map is generated. 7. Post-thinning processing, remove some obvious broken lines in the thinned image, obvious bridges between ridges, burrs on ridges, too short ridges and single spots and other bad ridge structures.

Step 3: Details Extraction

We use the following algorithm to detect the details: as shown in Figure 7,

Set point M to represent the grayscale value on the thinned image, M=0 to represent this point as a black point, and M=255 to represent this point as a white point.

If M=0, and $\underset{i=0}{\overset{7}{\mathrm{\Σ}}}|N_{(i+1)/8}-N_i|=2\×225,$ Then M is the endpoint;

If M=0, and $\underset{i=0}{\overset{7}{\mathrm{\Σ}}}|N_{(i+1)/8}-N_i|=6\×225,$ Then M is the bifurcation point.

Because the record information of the minutiae is determined according to a specific matching algorithm. Figure 8 shows the minutiae point model of our method. We record the following information according to our own matching algorithm:

1) The x and y coordinates of the detail points

2) The direction θ of the minutiae point, which is defined as the direction of the local ridge where the minutiae point is located.

3) The type t of the minutiae point, that is, the tip of the ridge or the branch of the ridge.

In this way, a fingerprint image is transformed into a plane point set M={M _k , 1≤k≤L} composed of minutiae points. where L is the number of detail points in the point set. For any one of the minutiae points, its feature vector is

The fingerprint image recognition module is to match the input fingerprint image to be recognized with the template information in the system database, and then judge whether the input fingerprint is from the same finger as a certain fingerprint in the template library. This module can be divided into two parts: offline and online. Among them: in the offline part, after calibrating the N templates of the same fingerprint, the feature data is clustered and analyzed to calculate the core of the feature points and other data; in the online part, the input fingerprint After the minutiae point set of the image is calibrated, it is matched with the data in the template library. The operation methods of these two parts are described below.

1. Offline part

Step 1: Calibration with N template images from the same finger

Due to the influence of factors such as the time and environment of fingerprint collection, even multiple fingerprint images collected by the same fingerprint will not completely overlap, but will rotate and translate. Different fingerprint images must be calibrated before template clustering and fingerprint matching. Respectively rotate the 2nd to N template point sets M _{i, j} ^T (j=2,...,n) of each finger in the fingerprint library based on the corresponding 1st template point set M _{i, 1} ^T and translation transformations. Specifically, the method of calibrating the minutiae point set of one fingerprint image to the minutiae point set of another fingerprint image is as follows:

Let the minutiae point sets of the two fingerprint images be

为细节点的方向，t_P ⁱ为细节点的类型；对于点集Q的第j个点q^j(1≤j≤N)，(q_x ^j，q_y ^j)是细节点的x和y轴坐标，为细节点的方向，t_j ^Q为细节点的类型。Among them, the point set P has M points in total, and the point set Q has N points in total. For the i-th point p ⁱ (1≤i≤M) of the point set P, (p _x ⁱ , p _y ⁱ ) is the x and y axis coordinates of the detail points,

is the direction of minutiae, t _P ⁱ is the type of minutiae; for the jth point q ^j (1≤j≤N) of point set Q, (q _x ^j , q _y ^j ) is the x and y of minutiae axis coordinates, is the direction of minutiae, t _j ^Q is the type of minutiae.

The purpose is to find the best transformation F _{s, θ, Δx, Δy} : R ² → R ² ,

为参考细节点。Let F _{s, θ, Δx, Δy} (p)=q. Here Δθ is the rotation parameter, (Δx, Δy) is the translation parameter, which belong to the attitude calibration parameters;

For reference details.

Our method is to search for the most similar point pairs of k (recommended value k=5) in the two detail point sets, and then use each point pair as a reference point, and estimate the rotation parameters of the local detail point set according to A.K.Jain's method and translation parameters. That is, each parameter is discretized into a finite set:

Δθ∈{Δθ ₁ , Δθ ₂ , ... Δθ _L }, Δx ∈ {Δx ₁ , Δx ₂ , ... Δx _L }, Δy ∈ {Δy ₁ , Δy ₂ , ... Δy _L }

Among them, the calculation method of the k most similar point pairs of two detail point sets is similar to the method of Xudong Jiang (Xudong Jiang, Wei-Yun Yau. Fingerprint Minutiae Matching Based on the Local and Global Structures. ICPR 2000: 6038-6041), namely Calculate the top k pairs of points where the point pair similarity function sl(*,*) takes the maximum value.

Respectively rotate the 2nd to N template point sets M _{i, j} ^T (j=2,...,n) of each finger in the fingerprint library based on the corresponding 1st template point set M _{i, 1} ^T After transformation and translation, the obtained calibration template point set is denoted as M _{i, j} ^A , T(j=1,...,N), where $m_{i,1}^{A,T}=m_{i,1}^T$ (i=1, . . . , L ₁ ).

Step 2: Multi-template detail information clustering

It is generally believed that the minutiae features of fingerprints are unique and invariant throughout life, so there must be some similarities between the minutiae points corresponding to each other between the minutiae point sets obtained after the templates collected multiple times are processed. The true minutiae points of a template point set should find similar point correspondences in other template point sets; while the pseudo minutiae points of this template point set have no corresponding points in other template point sets. We adopt the idea of cluster analysis, so that the similar minutiae points among the template point sets are included in the class, and the dissimilar minutiae points are all outside the class. That is, the similarity of the same class is the largest, and the similarity of different classes is the smallest. Then we define the class kernel feature vector to describe the commonality of similar minutiae points within the class. The specific method is as follows:

Input: the calibration template point set is denoted as M _{i, j} ^{A, T} (j=1,...,N)

Output: minutiae information class C _i ={C ₁ ,...,C _l } and the kernel feature vector of the class $K_i^T=\{K_{i,1}^T,...,K_{i,1}^T\}$

Then after the previous calibration, the point corresponding to a certain detail point q in the first template point set M i, ₁ ^T in the second to N template point sets M i _{, j} ^T (j=2,..., n) p _j (j=2,...,n) will fall in the vicinity of point q, as shown in Figure 9 is the detail point class diagram and kernel features.

Then we define the similarity function as follows:

If M _i,1 (1≤i≤L ₁ ) is the minutiae point of the template point set M ₁ , and M _j,2 (1≤j≤L ₂ ) is the minutiae point of the template point set M ₂ , we define M _{i, 1} and M _{j, 2} are similar if the condition

(1)|M _{i, 1} -M _{j, 2} |<Thre and

(2) There are M′ _{i, 1} and M″ _{i respectively, and 1} is M _{i, the two neighbors of 1} in its template point set M ₁ , there are M′ _{j, 2} and M″ _{j respectively, and 2} is in its template The two nearest neighbors in the point set M _2. Satisfy

|M′ _{i, 1} -M′ _{j, 2} |<T and |M″ _{i, 1} -M″ _{j, 2} |<T

Here, Thre and T are threshold parameters.

As shown in Figure 10, the classification rules are as follows:

1. If the number of similar points r ∈ [4, N], record these similar points as credible classes

2. If the number of similar points r ∈ [2, 3] and located on the edge of the image, record these similar points as candidate classes

3. Other isolated discrete points are neither marked nor recorded as classes.

In this way, we classify the points satisfying the similarity condition into one class. Next, for each class C, we use a feature vector to describe local structural features such as the degree of aggregation and the average direction and position of similar points within the class. We define this eigenvector as the kernel eigenvector K of the class.

The kernel eigenvector calculation method of a class is as follows:

Let the class of template similarity points be C _i ={M _{i, j} , j=1,...,L}, where

j＝1，…，L

j=1,...,L

记为第i个类C_i的核向量，这里，vector

Denoted as the kernel vector of the i-th class C _i , here,

The average coordinates of detail points in the class: $x_{c,i}=\left(\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{x}_{i,j}\right)/L,{\mathrm{the\; y}}_{c,i}=\left(\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{i,j}\right)/L$

The average direction of minutiae points in the class:

Class bounding box radius: $r_i=\mathrm{\&lambda;}\&Center\; Dot;\underset{j,k\&Element;\mathrm{Landj}\&NotEqual;k}{\max }\left(\mathrm{dis}(m_{i,j},m_{i,k})\right)$

(λ is a parameter greater than 1)

2. Online section

Step 1: Coarse fingerprint image matching. If successful, record the template number that may match, and go to step 2 below; if unsuccessful, output the matching result of "no matching fingerprint".

The specific rough matching method is as follows:

Step (1) Extract the minutiae point vector set M ^I after the input fingerprint image is subjected to the aforementioned enhanced processing.

Step (2) Compare the kernel feature vector set K ^T of each template in the template library with it, and determine the number of points R that the input image detail point set falls into the r-radius area of the template clustering kernel,

When the number R in step (3) is greater than a given threshold, the rough matching is considered successful. Indicates that the template is likely to match the input image. Add this potentially matching template label to the candidate list, and go to the fine matching operation (ie step 2). Otherwise, go to step (2) and search for the next template until all templates in the library have been traversed once, then exit and return a matching failure.

Step 2: Fingerprint image fine matching. Match the possible fingerprint templates one by one. If there is a matching template, output the matching result of "successful matching"; if not, output the matching result of "no matching fingerprint".

The detailed matching method is as follows:

Step (1) Calibrate the minutiae point vector set M ^I of the input fingerprint and the minutiae point set M ₁ ^T of the candidate template.

Step (2) Use the method of Xudong Jiang (Xudong Jiang, Wei-Yun Yau. Fingerprint Minutiae Matching Based on the Local and Global Structures. ICPR 2000: 6038-6041) to perform polar coordinate transformation and calculate the matching score

Step (3) If the matching score is greater than the given threshold, it is considered that the fine matching is successful, and ^MI is returned for system learning. Otherwise, find the next candidate template, return to step (1), until all candidate templates are traversed once, exit, and return matching failure.

The purpose of the feedback module is to use the matching input fingerprint image information to repair the template data of the fingerprint, and make it more consistent through repeated training and learning of the template data. Its flow chart is shown in the figure. Let the minutiae point set of the input fingerprint image be M ^I , after identification it matches a certain fingerprint in the database, the template set of the fingerprint M ^T

There are mainly 4 steps.

Step 1: Merge M ^I and M ^T into a new minutiae set M.

Step 2: Use the cluster analyzer to calculate the clustering result C _i ^* of the new set of minutiae points M

Step 3: Calculate the average similarity H _ave ^* between the same kind of detail points in the clustering result C _i ^* and the weighted average similarity S _ave ^* between the classes. Similarly, for the clustering result C _i of M ^T , calculate H _ave and _Save . Methods as below:

${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; ave</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Among them, F(M _i ) and F(K) are feature vectors of minutiae point _Mi and central kernel K, respectively. N is the number of detail points in the class. S(*, *) is the similarity judging function of the feature vector.

For the class C1,...,Ct formed after the minutiae point set clustering operation,

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; ave</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Step 4: If the conditions are met: $h_{\mathrm{ave}}^{*}>h_{\mathrm{ave}}$ and $S_{\mathrm{ave}}^{*}<S_{\mathrm{ave}},$ Modify the template information corresponding to the fingerprint, replace C _i with C _i ^* , recalculate the class of the minutiae points, and the kernel and class radius of each class.

Tests have shown that this fingerprint image recognition algorithm can classify and compress the information of multiple templates of the same finger, and the extracted common detail feature point set is accurate. After adding two links of rough matching and feedback to the system, the effect is very good. It can be well applied in the process of fingerprint identification.

Example

As shown in Figure 1, we design and realize the fingerprint image recognition system by ourselves.

The fingerprint image processing system is based on Window98/95, adopts object-oriented design method and software engineering norms, realizes with C++ language, and is an image processing and analysis system oriented to the field of fingerprint identification. This system has rich graphics and image processing and analysis functions, not only has perfect two-dimensional image processing and analysis functions, but also can dynamically load various fingerprint recognition algorithms. The system provides a series of functions such as image input, image storage, image processing, algorithm loading, file conversion, and FVC testing tools.

The following is the specific implementation process of the automatic fingerprint recognition method based on template learning. The test data is the database of FVC2000, and the resolution is 300×300×256.

1) Read in multiple fingerprint template images by opening the file or the Open button.

2) Click the load module menu to load the fingerprint enhancement algorithm.

3) Click to extract minutiae points to get a smooth fingerprint refinement map, including terminal points and bifurcation points, as shown in Figure 12.

4) Calibrate feature point sets of multiple fingerprint template images, as shown in Figure 12(d).

5) Generate template data, as shown in Figure 12(e).

6) Read in a single input fingerprint image by opening the file or collecting fingerprints button.

7) Process fingerprints to be identified by fingerprint enhancement and detail extraction operations, as shown in Figure 13(b).

8) Select the fingerprint to be identified and the file in the template library.

Methods The identification algorithm standard evaluation method of the international fingerprint identification competition and the standard fingerprint database FVC2000 are used for testing. The experimental results are shown in Figure 14 and Figure 15.

The above results are consistent with the theoretical analysis conclusions of the inventors on fingerprint image matching algorithm research and system design. It has high reliability, applicability and adoptability.

Claims (7)

1. An automatic fingerprint identification method based on template learning comprises the following steps:

(1) registering fingerprint images based on multi-template detail information clustering, recording the characteristic information of the fingerprint templates and the fingerprint images to be matched, wherein the fingerprint image registration method comprises the following steps:

calibrating a plurality of template images;

clustering multi-template detail information, and calculating the distance between a core and an intra-class of reliable detail nodes corresponding to a plurality of templates;

(2) fingerprint image identification based on detail information feedback, using knowledge rule to feed back matched fingerprint image information, judging whether the characteristic information of input fingerprint image is similar to some fingerprint template information in system database, including:

rough matching of the fingerprint images, and judging the similarity degree between a minutiae feature point set of the fingerprint images and a multi-template minutiae class;

fine matching of the fingerprint images and outputting a matching result;

and feeding back the detail information.

2. The method of claim 1, wherein said calibrating of said plurality of template images comprises the steps of:

calculating a plurality of most similar point pairs among the template point sets;

estimating a rotation parameter and a translation parameter of each pair of most similar point pairs;

the set of minutiae points is calibrated in the local region of the most similar point pairs.

3. The method of claim 1, wherein said clustering of multi-template detailed information comprises the steps of:

judging a similarity point set according to a defined similarity function;

classifying according to defined classification rules.

4. The method of claim 1, wherein said clustering of multi-template detailed information comprises the recording step of:

mean coordinate (x) of minutiae within class_c，y_c)；

Mean direction of minutiae within class

Class bounding box radius <math><mrow> <msub> <mi>r</mi> <mi>i</mi> </msub> <mo>=</mo> <mi>&lambda;</mi> <mo>&CenterDot;</mo> <munder> <mi>max</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> <mo>&Element;</mo> <mi>L and j</mi> <mo>&NotEqual;</mo> <mi>k</mi> </mrow> </munder> <mrow> <mo>(</mo> <mi>dis</mi> <mrow> <mo>(</mo> <msub> <mi>M</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mo>,</mo> <msub> <mi>M</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mo>.</mo> </mrow></math>

5. The method of claim 1, wherein said coarse fingerprint image matching comprises the steps of:

after the input fingerprint image is undergone the process of enhancement treatment, the minutiae vector set M is extracted¹；

Using kernel feature vector set K of each template in template library^TComparing with the above-mentioned table, judging the number of points R of input image detail point set falling into R radius region of template clustering kernel,

if the number R is larger than a given threshold value, the rough matching is considered to be successful, the template label which is possibly matched is added into the candidate list, the fine matching operation is switched, otherwise, the step (2) is switched, the next template is searched until all the templates in the library traverse once, and the matching failure is returned.

6. The method of claim 1, wherein said detail information feedback comprises the steps of:

merging the point sets, and merging the input fingerprint image characteristic point set and the corresponding template characteristic point set;

clustering, namely calculating a clustering result of the new minutiae set by using a clustering analyzer;

calculating the class attribute, the average similarity in the class and the average similarity between the classes;

and modifying the template, and recalculating the kernel feature vector of the class.

7. An automatic fingerprint identification system based on template learning, comprising:

(1) fingerprint image registration module based on multi-template detail information clustering records fingerprint template and the characteristic information of fingerprint image to be matched, including:

calibrating a plurality of template images;

clustering multi-template detail information, and calculating the distance between the kernel and the class of reliable detail points of a plurality of templates for English;

(2) the fingerprint image recognition module based on detail information feedback utilizes knowledge rules to feed back fingerprint image information on matching, and judges whether the characteristic information of an input fingerprint image is similar to certain fingerprint template information in a system database, wherein the fingerprint image recognition module comprises:

rough matching of the fingerprint images, and judging the similarity degree between a minutiae feature point set of the fingerprint images and a multi-template minutiae class;

fine matching of the fingerprint images and outputting a matching result;

and feeding back the detail information.

Images