CN105184266B - A kind of finger venous image recognition methods - Google Patents

Abstract

A finger vein image recognition method. It includes dividing the finger vein ROI image into training and testing samples; normalizing the image size to 46*102 pixels; using Gabor filter to enhance; comparing and selecting each pixel to obtain the ROI enhanced image; using PCA to The ROI enhanced image is reduced in dimension; the ROI enhanced image after dimensionality reduction is hypergranularized to obtain a new granule set; the distance between each finger vein ROI image in the test sample and each large super spheroid in the new granule set is calculated. Carry out steps such as identification. The present invention granulates and fuses all finger vein ROI images belonging to the same individual into one large hypersphere by combining PCA and hypergranulation, which better describes the commonality of samples from the same individual collected at different times. When identifying, it is only necessary to treat the sample to be tested as a hypersphere, and compare the distances with all large hyperspheres after fusion.

Description

A finger vein image recognition method

technical field

The invention belongs to the technical field of biological feature recognition, in particular to a finger vein image recognition method.

Background technique

At present, due to the low security of the traditional biometric identification technology, it cannot meet people's needs for high-precision identification. With the development of biometric identification technology, finger vein has many advantages as a new type of biometric identification technology. First of all, finger vein recognition is a living body recognition, and the structure of different human blood vessel networks is almost unchanged for life after adulthood; secondly, finger veins are internal features, and there is no recognition obstacle caused by any external factors (such as: wear, etc.); The advanced collection system is non-contact, and there is no problem of easy theft.

There are many methods for finger vein recognition, but most of the traditional methods do not take into account the commonality between samples from the same individual (category) in the database, and all need to compare the samples to be recognized with all samples in the database without distinction. , so the recognition efficiency is low.

Contents of the invention

In order to solve the above problems, the object of the present invention is to provide a finger vein image recognition method that can improve recognition efficiency.

In order to achieve the above object, the finger vein image recognition method provided by the invention comprises the following steps carried out in order:

1) Divide all finger vein ROI images to be detected into two parts, one part is used as a training sample, and the other part is used as a test sample, and then classify all the above images and give a category label. If some images come from the same finger, then the class labels are the same;

2) Normalize the size of all the above finger vein ROI images to 46*102 (4692) pixels to obtain ROI normalized vein images;

3) Use the Gabor filter to perform Gabor enhancement on all ROI normalized vein images above, and obtain ROIs in 8 directions, namely 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135° and 157.5° feature image;

4) Compare the pixel gray values at the same position in the ROI feature images of each of the above 8 directions one by one, and use the direction of the ROI feature image corresponding to the maximum gray value as the direction feature of the ROI enhanced image pixel, so After comparing and selecting each pixel in turn, the ROI enhanced image is obtained, and each ROI enhanced image is expressed as a ROI enhanced image matrix A _46*102 ;

5) Use PCA to reduce the dimension of all ROI enhanced images, and obtain the principal component features, and reduce the ROI enhanced images to different dimensions by adjusting the contribution rate of the principal component features to the entire feature space;

6) Each dimension-reduced ROI-enhanced image is represented by a dimension-reduced sample feature vector, that is, each dimension-reduced ROI-enhanced image is represented by a hypersphere in a high-dimensional space, and the hypersphere The center of the sphere is the feature vector of the sample after dimension reduction, and the radius is set to 0, so that each ROI enhanced image after dimension reduction is abstracted into an original particle with a center and radius in the high-dimensional space, and then Get x large super-spherulites by super-spherulizing, and thus get a new set of granules;

7) Use the Euclidean distance formula to calculate the distance between each finger vein ROI image G _t in the test sample and each large hypersphere in the above-mentioned new particle set, and select the large hypersphere G _m with the smallest distance to it, Then the category of the large hypersphere G _m is the category of the finger vein ROI image G _t in the test sample.

In step 3), the expression of described Gabor filter is as shown in formula (1):

Among them, σ represents the scale of the Gabor filter, σ=4,5,6; θ _k represents the angle value of the kth direction.

In step 5), the specific process of the PCA dimensionality reduction is as follows:

(1) Take each column of the above-mentioned ROI enhanced image matrix A _46*102 as one dimension, subtract the mean value of the dimension from the data of each dimension, and centralize its features to obtain matrix B;

(2) Calculate the covariance matrix C of the matrix B;

(3) Calculate the eigenvalues and eigenvectors of the covariance matrix C;

(4) Arrange the eigenvalues from large to small, if the sum of the first n eigenvalues has exceeded 97% of the sum of all eigenvalues, then take the eigenvectors corresponding to the first n eigenvalues to get a new data set.

In step 6), the specific steps of performing hyperspheroidization on the ROI enhanced image after dimensionality reduction are as follows:

(1) The ROI enhanced image after dimensionality reduction in all training samples is used as the operation object;

(2) Use the Euclidean distance formula shown in the following formula (3) to calculate the distance d _ji between a certain original grain G _j and all original grains G _i (i∈[1,n]) in the training sample, and record it, Stored in the jth column of the matrix in ascending order until the last original grain G _n is calculated, thus obtaining an n×n matrix D; the Euclidean distance formula is as follows:

d(G _i ,Gj)＝||C _i -C _j || ₂ -r _i -r _j (3)

Among them, G _i ＝(C _i , r _i ), G _j ＝(C _j , r _j ), C _i , C _j are the centers of G _i , G _j respectively, r _i , r _j are G _i , the radius of _Gj ;

(3) Set the maximum value of row y+1 in the matrix D as the threshold ρ, y is the number of training samples of the same individual, assuming that a certain grain G _j and other original grains G _i (i∈[1,n], The distance of i≠j) is d _ji , and these distances are compared with the threshold in turn, that is, all the distance values in the j-th column are compared with the threshold. If d≤ρ is satisfied, it means that the characteristics of the two original grains are relatively similar, very If they may be the same type of grain, the original grain G _i is selected and stored in the j cell, so that n cells will be obtained in the end, and each cell contains several grains that may be of the same type as the original grain G _j The original grain; then compare the category labels of all the original grains stored in the j-th cell with the category labels of the j-th original grain. If they are consistent, it means that it is of the same type as the j-th original grain and will continue to be retained In the jth cell, if it is inconsistent, it means that it is not of the same type as the jth original particle, and the jth cell is removed; finally, a new cell set is obtained, which contains n cells ;

(4) Count the cells containing the same original particles, keep only one, and delete the rest, and finally select x cells from n cells, x is the number of categories that are finally divided, if there is no category Error, then x is the number of real categories, if there is an error, then x is close to the number of real categories, so that the training sample set is divided into x categories;

(5) Fusion the original grains in each of the above x categories using formula (4) to obtain x large hyperspheres, thus obtaining a new grain set; the fusion formula is:

Among them, P＝C _i -r _i (C _ij /||C _ij ||), Q＝C _j +r _j (C _ij /||C _ij ||), C _ij is pointed from C _i to C _j Vector, C _ij =C _j -C _i , C is the center of the large hypersphere after fusion, R is the radius of the large hypersphere after fusion.

The finger vein image recognition method provided by the present invention granulates and fuses all finger vein ROI images belonging to the same individual into one large hypersphere by combining PCA with hypergranulation, which better describes the The commonality of samples collected from the same individual. When identifying, it is only necessary to treat the sample to be tested as a hypersphere, and compare the distances with all the large hyperspheres after fusion. It is not necessary to match each sample from the same individual one by one, so as to improve The problem of low recognition efficiency has been solved.

Description of drawings

Fig. 1 is an ROI normalized vein image in the present invention.

Fig. 2 is the ROI characteristic image of 8 directions in the present invention.

Fig. 3 is a ROI enhanced image in the present invention.

Fig. 4 is a schematic diagram of the fused large supersphere in the present invention.

Detailed ways

The finger vein image recognition method provided by the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

The finger vein image recognition method provided by the invention comprises the following steps carried out in order:

1) Divide all finger vein ROI (region of interest) images to be detected into two parts, one part is used as a training sample, and the other part is used as a test sample, and then all the above images are classified, and a category label is given. If some images come from For the same finger, the category labels are the same.

In the present invention, any 7 of the 10 finger vein ROI images from the same finger (such as the right index finger) of the same person are selected as training samples, and the other 3 are used as test samples, and these 10 finger vein ROI images are because belong to the same finger, so the category is the same, then the same category label is given.

2) normalize the size of all above-mentioned finger vein ROI images to 46*102 (4692) pixels, obtain the ROI normalized vein image as shown in Figure 1;

3) Since the vein texture and features of the above-mentioned ROI normalized vein images are not very clear, the present invention utilizes the Gabor filter to carry out Gabor enhancement to all the above-mentioned ROI normalized vein images, and the expression of the Gabor filter is as formula (1 ) shown.

Among them, σ represents the scale of the Gabor filter, σ=4, 5, 6; θ _k represents the angle value of the kth direction. Through calculation, 8 directions (0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135° and 157.5°) ROI feature images, as shown in Figure 2.

4) Compare the pixel gray values at the same position in the ROI feature images of each of the above 8 directions one by one, and use the direction of the ROI feature image corresponding to the maximum gray value as the direction feature of the ROI enhanced image pixel, so After sequentially comparing and selecting each pixel, the ROI-enhanced image as shown in FIG. 3 is obtained, and each ROI-enhanced image is represented as an ROI-enhanced image matrix A _46*102 .

5) The representation of information granules in high-dimensional space has always been an important issue in granule computing. However, if each ROI-enhanced image is directly regarded as a hypersphere, the dimension of the feature space will be very high (4692 dimensions), which makes it difficult to represent hyperspheres in high-dimensional space, and the calculation is complex and the recognition efficiency is low. Very low, so this step uses PCA (Principal Component Analysis) to reduce the dimension of all ROI enhanced images, and obtain the principal component features, and reduce the ROI enhanced images to different levels by adjusting the contribution rate of the principal component features to the entire feature space. dimension.

The specific process of PCA dimensionality reduction is as follows:

(1) Take each column of the above-mentioned ROI enhanced image matrix A _46*102 as one dimension, subtract the mean value of the dimension from the data of each dimension, and centralize its features to obtain matrix B;

(2) Calculate the covariance matrix C of the matrix B;

(3) Calculate the eigenvalues and eigenvectors of the covariance matrix C;

(4) Arrange the eigenvalues from large to small, if the sum of the first n eigenvalues has exceeded 97% of the sum of all eigenvalues, then take the eigenvectors corresponding to the first n eigenvalues to get a new data set.

For example, if the sum of the first two eigenvalues exceeds 97% of the sum of all eigenvalues and meets the requirements, then the eigenvectors corresponding to the first two eigenvalues are taken to obtain a matrix M of 102*2.

A′ _46*2 ＝A _46*102 ×M _102*2 (2)

According to the formula (2), the 46*102 data set A is mapped to the 46*2 data set A', and the number of features is reduced from 102 to 2.

The results of PCA dimensionality reduction in the present invention are shown in Table 1. The category label of the data obtained after the dimensionality reduction of each ROI enhanced image is still consistent with the original finger vein ROI image.

Table 1 Dimensions of image features after PCA dimensionality reduction

6) Before dimensionality reduction by PCA, if each ROI-enhanced image R′ is directly expressed as a hypersphere, then there are 4692 features in the sample, that is, the row vector composed of features has 4692 dimensions, which is too high. Therefore, PCA dimensionality reduction is used. As shown in Table 1, when the contribution rate is set to 80%, the reduced dimension is 62 dimensions, and the row vector composed of features is reduced to 62 dimensions, which can be greatly reduced from 4692 dimensions to 62 dimensions. Computation. So far in this step, each dimension-reduced ROI-enhanced image is represented by a dimension-reduced sample feature vector, that is, each dimension-reduced ROI-enhanced image is represented by a hypersphere in a high-dimensional space. The center of the sphere is the feature vector of the sample after dimension reduction, and the radius is set to 0, so that each ROI enhanced image after dimension reduction is abstracted into an original particle with the center and radius in the high-dimensional space, and then the It performs hyperspheronization to obtain x large hyperspherulites, thereby obtaining a new set of particles. Specific steps are as follows:

(1) The ROI enhanced image after dimensionality reduction in all training samples is used as the operation object;

(2) Use the Euclidean distance formula shown in the following formula (3) to calculate the distance d _ji between a certain original grain G _j and all original grains G _i (i∈[1,n]) in the training sample, and record it, Stored in the jth column of the matrix in ascending order. For example, first calculate the distance between the first original grain G ₁ and n original grains, and then store the n distance values in the first column of the matrix in order from small to large, and then calculate the distance between the second original grain G ₂ and n original grains. The distance of the original particle is stored in the second column, and so on until the last original particle G _n is calculated, thus obtaining an n×n matrix D. The Euclidean distance formula is as follows:

d(G _i ,G _j )＝||C _i -C _j || ₂ -r _i -r _j (3)

Among them, G _i ＝(C _i , r _i ), G _j ＝(C _j , r _j ), C _i , C _j are the centers of G _i , G _j respectively, r _i , r _j are G _i , Radius of _Gj .

(3) Set the maximum value of row y+1 in the matrix D as the threshold ρ, and y is the number of training samples of the same individual, such as y=7 in the present invention. Assuming that the distance between a grain G _j and other original grains G _i (i∈[1,n],i≠j ⁾ is d _ji , compare these distances with the threshold in turn, that is, all the distance values in the jth column are compared with Threshold comparison, if d≤ρ is satisfied, it means that the characteristics of the two original grains are relatively similar, and they are likely to be the same type of grain, then the original grain G _i is selected and stored in the j cell. In this way, n cells will be obtained in the end, and each cell contains several protogranules that may be of the same type as the protogranule G _j . Then compare the category labels of all the original particles stored in the jth cell with the category labels of the jth original particle. If they are consistent, it means that it is of the same type as the jth original particle and will continue to be retained in the jth cell. If it is inconsistent in the first cell, it means that it is not of the same type as the jth original grain, and the jth cell will be removed. Finally, a new cell set is obtained, which contains n cells.

(4) Some of the n cells must contain the same original particles. For example, suppose the four original particles from G ₁ to G ₄ are samples belonging to the same category, and when the training accuracy is 100%. , the protoparticles stored in the 1st to 4th cells should be G ₁ , G ₂ , G ₃ , G ₄ , then only the first cell should be kept, because this cell has already represented G ₁ , G 4 ₂ , G ₃ , and G ₄ belong to the same class, and the other three cells are repeated and can be deleted. Count the cells containing the same protoparticles, keep only one, and delete the rest, and finally select x cells from the n cells, and the protoparticles contained in these x cells are not the same. x is the number of categories that are finally divided. If there is no error in the classification, x is the number of real categories. If there is an error, x is close to the number of real categories. This divides the training sample set into x classes.

(5) Fusion the original grains in each of the above x categories using formula (4) to obtain x large hyperspheres, and thus obtain a new grain set. The fusion formula is:

Among them, P＝C _i -r _i (C _ij /||C _ij ||), Q＝C _j +r _j (C _ij /||C _ij ||), C _ij is pointed from C _i to C _j Vector, C _ij =C _j -C _i , C is the center of the large hypersphere after fusion, R is the radius of the large hypersphere after fusion. For example, there are two protospheres in two-dimensional space, namely G ₁ = (2,6,2), G ₂ = (5,7,3), then calculated according to formula (4), the fused large supersphere G = (3.97, 6.66, 4.08), as shown in Figure 4.

7) Use the Euclidean distance formula shown in formula (3) to calculate the distance between each finger vein ROI image G _t in the test sample and each large hypersphere in the new particle set, and select the one with the smallest distance large hypersphere G _m , then the category of the large hypersphere G _m is the category of the finger vein ROI image G _t in the test sample.

Using this method of distance measurement to identify each finger vein ROI image in all test samples can be done without knowing the category of the test sample (the present invention has given the category label in advance because it wants to verify the correct rate of recognition, It is used to compare with the recognition results), and by comparing with the trained new particle set, its category can be judged. If the recognized category is consistent with the original given category label, the recognition is correct; otherwise, the recognition is wrong.

In order to verify the effect of the present invention, the inventor has carried out following experiments:

The finger vein image database of experimental samples in the present invention is collected by a self-made system. This database contains the fingers of 185 different individuals, each individual contains 10 finger vein ROI images, a total of 1850 finger vein ROI images. All finger vein images were normalized to 46*102 (4692), and the experimental environment was PC, Matlab R2010a.

Here we mainly discuss the recognition performance of the finger vein image recognition method provided by the present invention in terms of test accuracy Ts (%), test time Ts (s), training accuracy Tr (%), and training time Tr (s). The specific experimental results are shown in Table 2.

Table 2 Recognition performance

It can be seen from Table 2 that when the contribution rate is 75% (that is, when the dimension is reduced to 49 dimensions), the performance of the other three aspects is also optimal when the test accuracy Ts (%) is guaranteed to be the highest. In 99% of the cases, both the test accuracy and test time are relatively poor, the reason may be that the high number of features caused the over-matching problem. In the range of 95% to 70%, Ts(%) is almost the same, which shows that the dimension reduction method PCA selected by the present invention is very stable in a wide range, so it is more convenient to set the contribution rate percentage. Tr(%) has been kept constant because there are fixed several images in the database that cannot be correctly identified. As for Ts(s), Tr(s) is related to the reduced dimension, so when the test accuracy is guaranteed to be the highest, the lower the dimension, the better.

It can be seen from the results in the table that the finger vein image recognition method provided in the present invention is feasible.

At the same time, the present invention granulates and fuses all finger vein ROI images belonging to the same individual into one large hypersphere, which better describes the commonality of samples from the same individual collected at different times. When identifying, it is only necessary to treat the sample to be tested as a hypersphere, and compare the distance with all the large hyperspheres after fusion, without matching each sample from the same individual one by one, thus greatly Improved recognition efficiency.

Claims (3)

1. a kind of finger venous image recognition methods, it is characterised in that：The finger venous image recognition methods includes by suitable The following steps that sequence carries out：

1) all finger vena ROI images to be detected are divided into two parts, a part is used as training sample, and a part is as survey Then sample sheet classifies above-mentioned all images, and given class label, if certain images come from same root hand Refer to, then class label is identical；

2) size of above-mentioned all finger vena ROI images is normalized to 46*102 (4692) pixel, it is quiet obtains ROI normalization Arteries and veins image；

3) Gabor enhancings are carried out to above-mentioned all ROI normalization vein images using Gabor filter, obtains 8 directions respectively I.e. 0 °, 22.5 °, 45 °, 67.5 °, 90 °, 112.5 °, 135 ° and 157.5 ° of ROI feature image；

4) pixel gray value of same position in the ROI feature image in each above-mentioned 8 direction is compared one by one, it will The direction of the corresponding ROI feature image of maximum gradation value enhances the direction character of image slices vegetarian refreshments as ROI, so successively will be every After a pixel is compared selection, ROI enhancing images are obtained, and each ROI enhancing graphical representations are enhanced into image moment at ROI Battle array A_46*102；

5) it uses PCA to carry out dimensionality reduction to all ROI enhancing images, and obtains principal component feature, accounted for by adjusting principal component feature ROI enhancing images are dropped to different dimensions by the contribution rate of entire feature space；

6) the ROI enhancing images after each dimensionality reduction are characterized with the sampling feature vectors after dimensionality reduction, i.e., dropped each ROI enhancing images after dimension indicate that the centre of sphere of hypersphere grain is the sample after dimensionality reduction with a hypersphere grain in higher dimensional space Feature vector, radius are set as 0, ROI after dimensionality reduction each in this way enhancing image be conceptualized as having in higher dimensional space the centre of sphere and The former grain of one of radius, then carries out hypersphere granulation to it and obtains x big hypersphere grain, thus obtain a new grain collection；

7) Euclidean distance formula is utilized to calculate each finger vena ROI image G in test sample_tIt is concentrated with above-mentioned new grain every The distance of a big hypersphere grain selects that big hypersphere grain G minimum with its distance_m, then big hypersphere grain G_mClassification be test Finger vena ROI image G in sample_tClassification；

In step 6), the ROI enhancing images to after dimensionality reduction carry out that hypersphere is granulated to be as follows：

(1) image is enhanced as operation object using the ROI after dimensionality reduction in all training samples；

(2) the former grain G of some in training sample is calculated using Euclidean distance formula shown in following formula (3)_jWith all former grain G_iAway from From d_ji, wherein i ∈ [1, n], and record, it is preserved in jth row in a matrix by sequence from small to large, until having been calculated The last one former grain G_n, thus obtain the matrix D of a n × n；Euclidean distance formula is as follows：

d(G_i,G_j)=| | C_i-C_j||₂-r_i-r_j (3)

Wherein, G_i=(C_i,r_i), G_j=(C_j,r_j), C_i, C_jRespectively G_i, G_jThe centre of sphere, r_i, r_jRespectively G_i, G_jRadius；

(3) maximum value of y+1 rows in matrix D is set as threshold value ρ, y is the number of training of same individual, it is assumed that some G_j With other former grain G_iThe distance of (i ∈ [1, n], i ≠ j) is d_ji, these distances are compared with threshold value successively, i.e., are arranged jth In all distance values and threshold value comparison illustrate that the feature of the two former grains is more similar, it is likely to same if meeting d≤ρ Class grain, then by former grain G_iIt picks out and is stored in j-th of cellular, can finally obtain n cellular in this way, include in each cellular Several may be with former grain G_jFor of a sort former grain；Then by the class label of all former grains preserved in j-th of cellular with The class label of j-th of former grain compares, if unanimously, illustrating that it with j-th of former grain is same class, will remain in the In j cellular, if it is inconsistent, explanation is not a kind of with j-th of former grain, then it will be rejected in its j-th of cellular；Finally just New cellular collection is arrived, n cellular is contained in the inside；

(4) cellular of contained identical former grain is come out, only retains one, x is finally selected in remaining deletion from n cellular A cellular, the classification number that x is as finally marked off, if the no any mistake of classification, x is true classification number, if gone out Show mistake, then training sample set is thus divided into x classes by x close to true classification number；

(5) it will be merged respectively using formula (4) per the former grain in one kind in above-mentioned x classes, obtain x big hypersphere grains, thus The grain collection new to one；Fusion formula is：

Wherein, P=C_i-r_i(C_ij/||C_ij| |), Q=C_j+r_j(C_ij/||C_ij| |), C_ijIt is by C_iIt is directed toward C_jVector, C_ij=C_j- C_i, C is the center of circle of big hypersphere grain after fusion, and R is the radius of big hypersphere grain after fusion.

2. finger venous image recognition methods according to claim 1, it is characterised in that：It is described in step 3) Shown in the expression formula of Gabor filter such as formula (1)：

Wherein, σ represents the scale of Gabor filter, σ=4, and 5,6；θ_kIndicate the angle value in k-th of direction.

3. finger venous image recognition methods according to claim 1, it is characterised in that：In step 5), the PCA Dimensionality reduction detailed process is as follows：

(1) by above-mentioned ROI enhancing image arrays A_46*102Each row as one-dimensional, every one-dimensional data are all subtracted into the dimension Mean value makes its eigencenter and obtains matrix B；

(2) the covariance matrix C of calculating matrix B；

(3) characteristic value and feature vector of covariance matrix C are calculated；

(4) it arranges characteristic value is descending, if the sum of preceding n characteristic value has been over the sum of all characteristic values 97%, then the corresponding feature vector of n characteristic value, obtains a new data set before taking.