CN103839051A - Single-sample vein recognition method based on 2DPCA and partition LBP - Google Patents

Abstract

The invention discloses a single-sample vein recognition method based on 2DPCA and partition LBP. The problem that single-sample recognition performance is low is solved, only a palm vein image sample is utilized, and the trouble of collecting multiple biological characteristic samples is avoided. The experimental result shows that the influence of the single-sample situation in palm vein recognition is overcome to a great degree, the recognition rate under the single-sample situation is improved, and therefore the practicality of single-sample vein recognition is guaranteed.

Description

Single sample vein based on 2DPCA and subregion LBP is known method for distinguishing

Technical field

The invention belongs to palm vein recognition technology field, relate to a kind of single sample vein based on 2DPCA and subregion LBP and know method for distinguishing.

Background technology

Along with development and the progress of modern society, people are more and more higher to the requirement of social informatization and networking security.Vein identification, as current up-to-date a kind of biometrics identification technology, is also correspondingly subject to domestic and international research institution and company and more and more payes attention to.So-called vein identification, refers to the technology of carrying out identification using the vein blood vessel under skin as identity characteristic.Vein is as the one of blood vessel, and it, is easier to be detected and obtained by near infrared ray near skin than artery.The complexity that the curve of vein pattern and branch are suitable, everyone difference is fully aware of, and according to statistics, the likelihood that palm vein distributes is only less than 0.00008%, and than the identification of fingerprint, iris etc., it has higher degree of accuracy.Vein blood vessel is positioned at body surface inside, little with its changes in microstructure of age growth, and is difficult to forge or operation changes, once can avoid the undermined and defect that cannot carry out the identification of fingerprint, palmmprint of epidermis; Compare DNA, iris recognition, its gatherer process is very friendly.A typical vein recognition system is mainly divided into two parts: the one, and registration part, comprises the generation of image pre-service, feature extraction and property data base; The 2nd, verification portion, comprises image pre-service, feature extraction and carries out matching judgment with the feature in property data base.

But, in current vein Study of recognition, most is algorithm research based on multisample, and the discrimination of algorithm can improve along with the increase of training sample number, but in the time only having single sample, recognition performance can significantly decline.In order to solve single specimen discerning performance deficiency, what current research was mainly paid close attention to is multi-modal biological characteristic fusion method.But, in actual applications, often can run into the situation that lacks multiple biometric sample.

Summary of the invention

Object of the present invention, for fear of the trouble that gathers multiple biometric sample, provides a kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP, effectively improves single sample palm vein discrimination.

A kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP of the present invention, comprises the steps:

Step 1, generation 2DPCA feature database:

Step 11, every in training image storehouse original single sample palm vein image is generated to 4 virtual subnet images 4 times by image resampling method sampling, and then utilize the virtual image generating algorithm of singular value disturbance, the singular value of this original single sample palm vein image is carried out to disturbance and generate other 4 virtual subnet images 4 times;

Step 12, utilize 2DPCA algorithm to extract vein pattern from 8 virtual subnet images that generate, the projection properties image that obtains original single sample palm vein image forms 2DPCA feature database;

Step 2, generation subregion LBP feature database:

Step 21, the pattern of the ROI region of every in training image storehouse original single sample palm vein image being pressed to row 2 deciles and row 2 deciles, be divided into 4 equal-sized subimages;

Step 22, every number of sub images is utilized to circular LBP operator extraction textural characteristics, the steps include: the each pixel in subimage, in the window of the circular LBP operator that is 8 at radius, take the gray-scale value of window center point as threshold value, 16 sampled point pixels in window are made to binary conversion treatment, obtain 16 bits, computing formula is:

${\mathrm{LBP}}_{P,R}\left(I\right)=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}S(Z_p-Z_c)2^{p-1}---\left(4\right)$

Wherein $S\left(u\right)=\left\{\begin{array}{cc}1& u\&GreaterEqual;0\\ 0& u<0\end{array}\right.,$ Z _p, Z _crepresent the pixel value of sampled point and the pixel value of LBP operator central point in LBP operator;

Step 23, to the LBP textural characteristics obtaining, further extract its equivalent formulations, in the time that the corresponding circulation binary number from 0 to 1 or from 1 to 0 of LBP textural characteristics has twice saltus step at most, the binary number that this LBP textural characteristics is corresponding is just called equivalent formulations;

Step 24, the Uniform LBP histogram of every number of sub images corresponding original single sample palm vein image is coupled together and forms the vein pattern expression figure of entire image, form subregion LBP feature database;

Step 3, double-deck screening:

The 2DPCA feature of step 31, extraction test pattern is mated by nearest neighbor method with 2DPCA feature database, select the sample of similarity rank front 10%, then utilize this sample class of 10%, choose the search sample space of corresponding subregion LBP feature database as step 32;

The subregion LBP feature database dwindling in step 32, the subregion LBP feature of extracting test pattern and step 31 is selected optimum matching Pattern Class by nearest neighbor method, from the subregion LBP feature database dwindling, select and the sample class of subregion LBP characteristic distance minimum, matching process finishes.

Described image resampling method is by original single sample palm vein image is sampled to obtain new virtual subnet image, detailed process is the sub-block that every in training image storehouse original single sample palm vein image is divided into P × Q size, then the pixel of sampling from the sub-block of each P × Q size forms a virtual subnet image of original single sample palm vein image, described P and Q are respectively the length of sampling interval and wide, make P=Q=2, if the size of original single sample palm vein image is M × N, M, N is respectively the length of image and wide, a virtual subnet image I ij (m who generates through image resampling, n) be:

I _ij(m,n)=I((m-1)×P+i,(n-1)×Q+j) （1）

In formula (1), 1≤m≤M/P, 1≤n≤N/Q, 1≤i≤P, 1≤j≤Q.

The described virtual image generation method based on singular value disturbance is to generate new virtual subnet image by the singular value of original single sample palm vein image being carried out to disturbance, suppose I (x, y) grey value profile of the vein image of a M × N of expression, wherein x ∈ [1, M], y ∈ [1, N], through type (2) obtains a new virtual subnet image P:

P=UΣ ⁿV ^T （2）

In formula (2), n represents disturbance factor, U, V are expressed as an orthogonal matrix, the transposition of T representing matrix, and Σ represents a diagonal matrix, value on its diagonal line is image array I (x, y) singular value, U, Σ, V is determined by the svd formula (3) of image array I (x, y): I=U Σ V ^t(3).

Described 2DPCA algorithm is directly to utilize virtual subnet image to obtain covariance matrix, try to achieve the corresponding proper vector composition of front d the larger eigenwert projection matrix of covariance matrix, then above-mentioned 8 virtual subnet images are all projected in above-mentioned proper vector, obtain the projection properties image of original single sample palm vein image.

The present invention is a kind of vein identification method based on 2DPCA feature and subregion LBP feature, overcome the lower problem of single specimen discerning performance, only utilize palm vein image sample, avoid gathering the trouble of multiple biometric sample, experimental result shows, the present invention has overcome the impact of single sample situation in palm vein identification to a great extent, has improved the discrimination in single sample situation, thereby has guaranteed the practicality of single sample palm vein identification.

Accompanying drawing explanation

Fig. 1 is palm vein identification process figure of the present invention;

Fig. 2 is the virtual image (P=Q=2) that image resampling of the present invention generates;

Fig. 3 is the basic LBP operator in the present invention;

Fig. 4 is Simulation identification result diagram in the present invention.

Below in conjunction with the drawings and specific embodiments, the invention will be further described.

Embodiment

As shown in Figure 1, a kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP of the present invention, specifically comprises the steps:

Step 1, generation 2DPCA feature database:

Step 11, every in training image storehouse original single sample palm vein image is generated to 4 virtual subnet images 4 times by image resampling method sampling, and then utilize the virtual image generating algorithm of singular value disturbance, the singular value of this original single sample palm vein image is carried out to disturbance 4 times, to generate other 4 virtual subnet images;

Wherein image resampling method is by original single sample palm vein image is sampled to obtain new virtual subnet image, detailed process is as follows, every in training image storehouse original single sample palm vein image is divided into the sub-block of P × Q size, then the pixel of sampling from the sub-block of each P × Q size forms a virtual subnet image of original single sample palm vein image, described P and Q are respectively the length of sampling interval and wide, unification of the present invention makes P=Q=2, image resampling result as shown in Figure 2, if the size of original single sample palm vein image is M × N, M, N is respectively the length of image and wide, the virtual subnet image I generating through image resampling _ij(m, n) is:

I _ij(m,n)=I((m-1)×P+i,(n-1)×Q+j) （1）

In formula (1), 1≤m≤M/P, 1≤n≤N/Q, 1≤i≤P, 1≤j≤Q;

Wherein the virtual image generation method based on singular value disturbance is to generate new virtual subnet image by the singular value of original single sample palm vein image being carried out to disturbance, detailed process is as follows: suppose I (x, y) grey value profile of the vein image of a M × N of expression, wherein x ∈ [1, M], y ∈ [1, N], through type (2) obtains a new virtual subnet image P:

P=UΣ ⁿV ^T （2）

In formula (2), n represents disturbance factor, U, V are expressed as an orthogonal matrix, the transposition of T representing matrix, and Σ represents a diagonal matrix, value on its diagonal line is image array I (x, y) singular value, U, Σ, V is determined by the svd formula (3) of image array I (x, y): I=U Σ V ^t(3);

Step 12, utilize 2DPCA algorithm from generate 8 virtual subnet images extract vein pattern, the projection properties image that obtains original single sample palm vein image forms 2DPCA feature database;

Described 2DPCA algorithm is directly to utilize virtual subnet image to obtain covariance matrix, try to achieve the corresponding proper vector composition of front d the larger eigenwert projection matrix of covariance matrix, then above-mentioned 8 virtual subnet images are all projected in above-mentioned proper vector, obtain the projection properties image of original single sample palm vein image;

Step 2, generation subregion LBP feature database:

Step 21, the pattern of the ROI region of every in training image storehouse original single sample palm vein image being pressed to row 2 deciles and row 2 deciles, be divided into 4 equal-sized subimages;

Step 22, every number of sub images is utilized to circular LBP operator extraction textural characteristics, the steps include: the each pixel in subimage, in the window of the circular LBP operator that is 8 at radius, gray-scale value take window center point is made binary conversion treatment as threshold value to 16 sampled point pixels in window, obtains 16 bits; As shown in Figure 3, computing formula is as follows for the Method And Principle of a simple LBP operator:

${\mathrm{LBP}}_{P,R}\left(I\right)=\underset{p=1}{\overset{P}{\mathrm{\&Sigma;}}}S(Z_p-Z_c)2^{p-1}---\left(4\right)$

Wherein $S\left(u\right)=\left\{\begin{array}{cc}1& u\&GreaterEqual;0\\ 0& u<0\end{array}\right.,$ Z _p, Z _crepresent the pixel value of sampled point and the pixel value of LBP operator central point in LBP operator;

Step 23, to the LBP textural characteristics obtaining, further extract its equivalent formulations (Uniform Pattern), in the time that the corresponding circulation binary number from 0 to 1 or from 1 to 0 of LBP textural characteristics has twice saltus step at most, the binary number that this LBP textural characteristics is corresponding is just called equivalent formulations;

Step 24, the Uniform LBP histogram of every number of sub images corresponding original single sample palm vein image is coupled together and forms the vein pattern expression figure of entire image, form subregion LBP feature database;

Step 3, double-deck screening:

The 2DPCA feature of step 31, extraction test pattern is mated by nearest neighbor method with 2DPCA feature database, select the sample of similarity rank front 10%, then utilize this sample class of 10%, choose the search sample space of corresponding subregion LBP feature database as step 32;

The subregion LBP feature database dwindling in step 32, the subregion LBP feature of extracting test pattern and step 31 is selected optimum matching Pattern Class by nearest neighbor method, from the subregion LBP feature database dwindling, select and the sample class of subregion LBP characteristic distance minimum, matching process finishes.

Fig. 4 is Simulation identification effect diagram of the present invention, as we know from the figure, the present invention adopts double deck screen choosing method than the vein identification method that only adopts 2DPCA or LBP, the raising highly significant of discrimination, average recognition rate can reach 93.116%, has good practical value.

The above, it is only preferred embodiment of the present invention, not technical scope of the present invention is imposed any restrictions, therefore any trickle modification, equivalent variations and modification that every foundation technical spirit of the present invention is done above embodiment all still belong in the scope of technical solution of the present invention.

Claims (4)

1. the single sample palm vein recognition methods based on 2DPCA and subregion LBP, is characterized in that comprising the steps:

Step 1, generation 2DPCA feature database:

Step 11, every in training image storehouse original single sample palm vein image is generated to 4 virtual subnet images 4 times by image resampling method sampling, and then utilize the virtual image generating algorithm of singular value disturbance, the singular value of this original single sample palm vein image is carried out to disturbance and generate other 4 virtual subnet images 4 times;

Step 12, utilize 2DPCA algorithm to extract vein pattern from 8 virtual subnet images that generate, the projection properties image that obtains original single sample palm vein image forms 2DPCA feature database;

Step 2, generation subregion LBP feature database:

Step 21, the pattern of the ROI region of every in training image storehouse original single sample palm vein image being pressed to row 2 deciles and row 2 deciles, be divided into 4 equal-sized subimages;

Step 22, every number of sub images is utilized to circular LBP operator extraction textural characteristics, the steps include: the each pixel in subimage, in the window of the circular LBP operator that is 8 at radius, take the gray-scale value of window center point as threshold value, 16 sampled point pixels in window are made to binary conversion treatment, obtain 16 bits, computing formula is:

${\mathrm{LBP}}_{P,R}\left(I\right)=\underset{p=1}{\overset{P}{\mathrm{\&Sigma;}}}S(Z_p-Z_c)2^{p-1}---\left(4\right)$

Wherein $S\left(u\right)=\left\{\begin{array}{cc}1& u\&GreaterEqual;0\\ 0& u<0\end{array}\right.,$ Z _p, Z _crepresent the pixel value of sampled point and the pixel value of LBP operator central point in LBP operator;

Step 23, to the LBP textural characteristics obtaining, further extract its equivalent formulations, in the time that the corresponding circulation binary number from 0 to 1 or from 1 to 0 of LBP textural characteristics has twice saltus step at most, the binary number that this LBP textural characteristics is corresponding is just called equivalent formulations;

Step 24, the Uniform LBP histogram of every number of sub images corresponding original single sample palm vein image is coupled together and forms the vein pattern expression figure of entire image, form subregion LBP feature database;

Step 3, double-deck screening:

The 2DPCA feature of step 31, extraction test pattern is mated by nearest neighbor method with 2DPCA feature database, select the sample of similarity rank front 10%, then utilize this sample class of 10%, choose the search sample space of corresponding subregion LBP feature database as step 32;

The subregion LBP feature database dwindling in step 32, the subregion LBP feature of extracting test pattern and step 31 is selected optimum matching Pattern Class by nearest neighbor method, from the subregion LBP feature database dwindling, select and the sample class of subregion LBP characteristic distance minimum, matching process finishes.

2. a kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP according to claim 1, it is characterized in that: described image resampling method is by original single sample palm vein image is sampled to obtain new virtual subnet image, detailed process is the sub-block that every in training image storehouse original single sample palm vein image is divided into P × Q size, then the pixel of sampling from the sub-block of each P × Q size forms a virtual subnet image of original single sample palm vein image, described P and Q are respectively the length of sampling interval and wide, make P=Q=2, if the size of original single sample palm vein image is M × N, M, N is respectively the length of image and wide, a virtual subnet image I ij (m who generates through image resampling, n) be:

I _ij(m,n)=I((m-1)×P+i,(n-1)×Q+j) （1）

In formula (1), 1≤m≤M/P, 1≤n≤N/Q, 1≤i≤P, 1≤j≤Q.

3. a kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP according to claim 1, it is characterized in that: the described virtual image generation method based on singular value disturbance is to generate new virtual subnet image by the singular value of original single sample palm vein image being carried out to disturbance, suppose I (x, y) grey value profile of the vein image of a M × N of expression, wherein x ∈ [1, M], y ∈ [1, N], through type (2) obtains a new virtual subnet image P:

P=UΣ ⁿV ^T （2）

In formula (2), n represents disturbance factor, U, V are expressed as an orthogonal matrix, the transposition of T representing matrix, and Σ represents a diagonal matrix, value on its diagonal line is image array I (x, y) singular value, U, Σ, V is determined by the svd formula (3) of image array I (x, y): I=U Σ V ^t(3).

4. a kind of single sample palm vein recognition methods based on 2DPCA and subregion LBP according to claim 1, it is characterized in that: described 2DPCA algorithm is directly to utilize virtual subnet image to obtain covariance matrix, try to achieve the corresponding proper vector composition of front d the larger eigenwert projection matrix of covariance matrix, then above-mentioned 8 virtual subnet images are all projected in above-mentioned proper vector, obtain the projection properties image of original single sample palm vein image.

Images