CN100492400C - Matching identification method by extracting characters of vein from finger - Google Patents

Abstract

An identifying method based on pick-up and match of finger vein includes collecting finger vein image by image collection device, carrying out pretreatment on collected vein image, dividing vein image to be image block in specific size, applying wavelet decomposing algorithm to carry out wavelet decomposition and wavelet moment pick-up on sub-image, picking up characteristic by carrying out PCA and LDA transform and carrying out match and identification by utilizing fuzzy threshold means.

Description

Finger vein features is extracted and matching and recognition method

(1) technical field

What the present invention relates to is a kind of biological characteristics identity recognizing technology, particularly a kind of finger vein features identity identifying technology.

(2) background technology

Along with arrival of information age, the security of information and confidentiality have caused the great attention that people are general.The physiological characteristic that human body is abundant makes biometrics identification technology become an important means of field of identity authentication.

Biometrics identification technology (Biometric Identification Technology) is meant a kind of technology of utilizing human body biological characteristics to carry out authentication.With traditional different being of method, the foundation of living things feature recognition method is the thing that our human body itself is had, and is our individual character.In fact, any physiological feature can be used for discerning.Biological characteristic is divided into based on physical trait with based on behavioral characteristic two classes.Physical trait comprises: the vascular lake of fingerprint, palm type, eyes (retina and iris), human scent, the shape of face, skin pore, wrist/hand and DNA etc.; Behavioral characteristic comprises: the gait of signature, voice, walking, the dynamics of keystroke etc.

Fingerprint is the biological identification mode of using the most extensively at present, but but has certain secret worry.Its maximum secret worry is exactly to duplicate fraud easily, because fingerprint is the feature of body covering, the lawless person obtains the trace of fingerprint easily.In addition, fingerprint recognition system is to adopt optical glass or plastics prism to read fingerprint image, will overcome many technical matterss between finger and fingerprint capturer.As: the people loses available fingerprint suddenly, and fingerprint is by extraneous as wet excessively, and the material of overdrying covers.When fingerprint is wet excessively, fingerprint image can be damaged or fuzzy, and the possibility of success comparison will reduce like this.The acquisition instrument of this contact is antihygienic in addition.

And finger vena identification is as follows with respect to the advantage of fingerprint recognition and other recognition technologies:

(1) vivo identification

When carrying out authentication with finger vena, what obtain is the characteristics of image of finger vena, the feature that just exists when being the finger live body.The finger of non-living body can not get the vein image feature and can't discern, thereby also just can't fake.

(2) internal feature

When carrying out authentication with finger vena, what obtain is the vein image feature of finger interior, rather than the characteristics of image of finger surface.Therefore do not exist any because the cognitive disorders that drying or too wet etc. is brought is worn and torn in the damage of finger surface.Can overcome these shortcomings of fingerprint fully.

(3) contactless

Carry out authentication with finger vena, when obtaining finger venous image, finger need not contact with equipment.Finger one is stretched gently, can finish identification.Unhygienic, finger surface feature during not as the contact arrangement of fingerprint recognition may be replicated the safety problem of being brought, and it is uncomfortable to have avoided being taken as the psychology of examination object.

(4) safe class height

Because the vivo identification of front has been arranged, the feature of internal feature and contactless three aspects has guaranteed that user's finger vein features is difficult to be forged, so the safe class height is particularly suitable for the high place of safety requirements and uses.According to the test findings that Hitachi, Ltd provides, it is 0.01% that the product of their development is refused sincere, and False Rate is 0.0001%, has been used for fields such as bank, safety check.

Clearly, finger vena identification has overcome many shortcomings of fingerprint recognition etc., and wide application prospect is arranged.Therefore increased gradually in the research aspect the finger vena identification both at home and abroad in recent years.

In the field of vein identification, studying more is the vein of the back of the hand and palm.In this respect, the NEXTERN company of Korea S conducts a research the earliest, and the product of existing moulding is being sold.This cover vein recognition system when identification fast and accurately, solved bio-identification circle general pass through the comparison identification problem of carrying out for biometrics ID number.Adopt infrared C CD camera collection image, analyze hand back vein figure people's identity is discerned, operation interface is friendly and solid, and this system makes the biometric security rank forward further.

Fujitsu has announced the palm vein recognition device of their research in March, 2003.This technology writes down and analyzes the subcutaneous vein image of people's palm by infrared ray, and carries out authentication in view of the above.In Dec, 2005, the said firm announces novel palm vein identification product again.

In 2005, the Nanyang Technolohy University of Singapore was also relevant for the research of vein aspect, and delivered some papers.Different with the product of Japan and Korea S is that what they adopted is to clap the vein that the red heat image of getting the back of the hand extracts the back of the hand, it is said that FAR and FRR are 0.0%, but the temperature of environment and humidity has very big influence to its discrimination.

At home, the State Central Univ. in Taiwan also possesses some special knowledge aspect vein.People such as Kuo-Chin Fan described a kind of red heat image of the back of the hand that adopts in 2004 and have obtained vein image, FRR=1.5% that provides in the literary composition and FAR=3.5% in the paper of delivering.

People such as the Lin Xirong of Tsing-Hua University, Zhuan Bo utilize the near infrared blood-vessel image Acquisition Instrument of design voluntarily to extract the original image of blood vessel, after the image pattern that collects done normalization, enhancing, image segmentation, refinement etc. and handle, extracting features such as end points, point of crossing, to carry out aspect ratio right.Document has provided 65 coupling experimental results of sample among a small circle, is under 4.6% condition refusing sincere, and misclassification rate is 0%.

The research that people such as the Wang Dazhen of Harbin Engineering University, Zhuan Dayan, fourth aerospace also discern relevant for hand back vein in its document, and delivered some documents.

Below all be based on the research of palm hand back vein identification, aspect finger vena identification, only have the Hitachi, Ltd of Japan to possess some special knowledge abroad, and commercialization.

The initial development product in Hitachi is finger to be put in carry out room entry/exit management in the authenticating device, starts selling in September, 2003.Product structure is: the top is provided with led light source irradiation finger, takes vein image as shown in Figure 5 with photographic element below finger.

The Hitachi has released in October, 2004 and has been used for bank's window and authenticates my opening equipment, and only limit is in indoor use.With the adjusting sensitivity that is as the criterion of the light quantity around the vein image, the Hitachi improves as ultrared light source sunlight again to equipment.Sunlight is different with former led light source, and its intensity has very big difference in different time and occasion.The said firm is used as sunlight as the light source use by change photographic element sensitivity along with sunlight strength, and the equipment that makes can be in outdoor application.

On the Tokyo Motor Show of holding in October, 2005, Hitachi, Ltd has externally showed this cover security system.Be installed in a sensor behind the handlebar and utilize vein textured pattern on finger of near infrared ray identification.The unique design of handlebar can guide driver's hand just in time to enter the position that they open car door at every turn, and finger vena is all at same position during assurance system each reading of data.In Dec, 2005, Hitachi, Ltd was applied to this cover system above the notebook computer again.

In addition, Hitachi, Ltd also develops the portable finger vein reader of matchbox size.About this system of Hitachi, Ltd, people such as Naoto Miura in its paper to the development of harvester and the very detailed again elaboration of algorithm of extracting blood vessel, identification contrast.Adopt high precision CCD camera collection infrared image, because good imaging quality, thereby identification contrast algorithm has adopted the simplest masterplate matching algorithm.

In the world, the aspect is applied for a patent in identification about vein, mainly contains 11.Wherein Korea S has 4, mainly is based on the equipment or the system of hand back vein, palm vein.Some patents of Japanese publication that remaining mainly is about the equipment of finger vena aspect.It is to gather a package of fingerprint and finger vena simultaneously that a patent is wherein arranged, and has merged two kinds of biological informations of fingerprint and vein, and we of value go to use for reference.

(3) summary of the invention

The object of the present invention is to provide a kind of misclassification rate and reject rate low, the fast finger vein features of recognition speed is extracted and matching and recognition method.

The object of the present invention is achieved like this:

1, image collecting device carries out the collection of finger venous image;

2, the vein image of gathering is carried out pre-service, pre-service comprises: adopt the weighted mean value method carry out gray processing, adopt the method for iteration ask for optimal threshold come to image cut apart, adopt junction filter come filtering noise, the local dynamic thresholding method of employing carry out image segmentation, take after the area null method is cut apart denoising, the vein image after cutting apart is carried out highly standardized processing; Described employing junction filter comes filtering noise, junction filter is made up of salt-pepper noise detection, selective filter, elimination salt-pepper noise and 4 modules of elimination Gaussian noise altogether, original vein image to input carries out the salt-pepper noise detection earlier, can select corresponding wave filter to carry out the elimination of salt-pepper noise to the pixel that influenced by salt-pepper noise, then select corresponding wave filter to carry out the elimination of Gaussian noise to the pixel that not influenced by salt-pepper noise, at last two results are merged the finger venous image that obtains after the de-noising;

3, by vein image being divided into the image block of specific size, adopt the wavelet decomposition algorithm that subimage is carried out the extraction of wavelet decomposition and small echo square, carry out PCA and LDA shift step and extract feature;

4, adopt the threshold method of obfuscation to mate and identification.

Technical scheme provided by the invention is a kind of recognition methods of merging PCA and LDA conversion based on the small echo square.Its main design thought and characteristics are:

The collection of 1 finger venous image

Come images acquired by image collecting device.

The pre-service of 2 vein images

2.1 the conversion of colored bitmap and gray level image

Adopt the weighted mean value method to carry out gray processing, reduce computing and data quantity stored.

2.2 the location of finger areas

What extract is blood vessel in the finger, and the part for beyond finger contours and the profile needn't take in, so at first will mark the profile of finger.The present invention adopts the method for iteration to ask for optimal threshold to come image is cut apart.

2.3 the filtering of image

Because the image that collects contains noise in various degree,, need carry out filtering operation to this width of cloth gray level image in order to remove the salt-pepper noise of this point, bulk.The present invention adopts junction filter to come filtering noise.

2.4 cutting apart of image

Adopted a kind of local dynamic threshold algorithm.

2.5 the denoising after cutting apart

The method of taking is the area null method.Promptly calculating each piece black background or white background, is noise if background area less than given area, then is this piece, with its removal.

2.6 the cutting of finger areas and the standardization of finger width

Because the randomness of image acquisition and the difference of sample individuality, size, the ratio of the finger blood tube model that extracts at last are often inconsistent, and, need carry out height (i.e. Shou Zhi width) standardization to the vein image after cutting apart in order to make things convenient for further research.

3 feature extractions

3.1 the piecemeal of finger venous image

Because the length of different people's finger is different in size, is exactly same individual's finger, the length of gathering constantly in difference is also not necessarily identical.Therefore if directly image is asked for the small echo square, is carried out the PCA conversion and extract feature, the error of the individuality of feature representative is strengthened, thereby discrimination is reduced.The present invention solves the problems referred to above with the method that vein image is divided into the image block of specific size.

3.2 the extraction of the wavelet decomposition of subimage and small echo square

The small echo square is an invariant of well describing characteristics of image.The small echo square has the characteristics of rotation, translation and the convergent-divergent unchangeability of pair image, so a lot of successful application are arranged in pattern-recognition.So also adopt the wavelet decomposition algorithm in the present invention.

3.3 PCA and LDA conversion

As linear method, the PCA method is the optimum dimension compress technique on the least mean-square error meaning, that is under identical dimension, uses principal component method former data to be carried out will comprise in the data of conversion gained the information of maximum former data.The present invention proposes the feature extracting method that merges PCA and LDA based on the small echo square.

4 couplings and identification

The present invention has adopted a kind of threshold method of obfuscation, and the sample in to be identified and the database is mated.Adopt nearest neighbor classifier to classify at last, reach the purpose of identification.

Method of the present invention is verified by following experiment:

(consider user's convenience with finger, mainly be forefinger and the middle finger of gathering everyone) be unit, gather the vein image of 287 (class) finger altogether, each finger collection 5 times (every class sample number is 5), gather 287 * 5=1435 width of cloth vein image altogether, constituted the vein storehouse of finger vena identification thus.Each finger is adopted 2 width of cloth totally 287 * 2=574 width of cloth composition checking storehouse more in addition, for the validity of verification algorithm.

Each sample to the vein storehouse carries out one deck wavelet decomposition earlier, and high fdrequency component is extracted moment characteristics, and low frequency component carries out the PCA decomposition, and prerequisite is that we have utilized the sample in the vein storehouse to try to achieve the transition matrix that PCA decomposes certainly.

When carrying out dimensionality reduction with PCA, the dimension k after the compression choose proportion w with its component that can represent _kRelation as shown in table 1:

$w_k=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\λ}}_i/\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i$

The relation of the dimension after table 1 compression and the proportion of its component that can represent

Balance calculated amount and shared weight, the dimension of getting after the decomposition is 300.Under the Validation Mode of 1:1, verify that with 574 samples in the checking storehouse experimental result is as shown in table 2.

Reject rate is differentiated the result under the table 2 1:1 situation

Under the recognition mode for 1:n, discern experiment with 574 samples in the checking storehouse, experimental result is as shown in table 3.

Misclassification rate is differentiated the result under the table 3 1:n situation

From identification the result as can be seen, based on small echo, PCA is very low in conjunction with the misclassification rate and the reject rate of the method for LDA.Aspect recognition speed, under the pattern of 1:n, also can meet the demands, every identification time once of (n=287) approximately is 0.1s on our basis, vein storehouse.

(4) description of drawings

Fig. 1 is a finger vein recognition system block diagram of the present invention;

Fig. 2 is the design frame chart of junction filter;

Fig. 3 is the effect after the image segmentation;

Fig. 4 is the effect after the denoising;

Fig. 5 is the image after the finger width standardization;

Fig. 6 is the synoptic diagram that finger extracts the subimage block method;

Fig. 7-1, Fig. 7 the-the 2nd, the image subblock that extracts and the result of wavelet decomposition;

Fig. 8 is that matrix transform vectorial synoptic diagram as;

Fig. 9 is a sample classification synoptic diagram in the PCA conversion.

(5) embodiment

For example the present invention is done description in more detail below in conjunction with accompanying drawing:

Finger vein features is extracted the method with coupling identification, come images acquired by the vein harvester, then to the image that collects carry out that gray scale transforms, after the pre-treatment step such as standardization of the cutting of the cutting apart of the location of finger areas, filtering, image, denoising, finger areas and finger width, merge the method extraction feature of PCA and LDA conversion again by the small echo square, finish identification at last.In conjunction with Fig. 1, method of the present invention contains following steps:

(1) collection of finger venous image

The ultimate principle of the finger vena harvester of selecting for use is to utilize near infrared ray irradiation finger, and points the light that transmission is come by image sensor senses.The haemoglobin that key wherein is to flow in the vein red blood cell can lose deoxidation part because of irradiation, and the haemoglobin of this reduction has absorption near the near infrared ray the wavelength 760nm, what adopt in the present embodiment is the infrared light supply of 850nm, cause the transmission of vein segment less, on imaging device, will produce vein pattern.So the finger vena collector of selecting for use utilizes the near infrared power of transmission to show vein blood vessel especially.

(2) pre-service of vein image

The vein image that collects if directly image is cut apart, is difficult to extract exactly the model of blood vessel owing to be subjected to The noise.For this reason, before image segmentation, earlier image is carried out filtering and remove noise.Through experiment showed, adopt after the combined filter method relatively good to the image segmentation effect again.Described combined filter method comprises:

(2.1) conversion of colored bitmap and gray level image

Through gray processing, the data volume of image only is original 1/3, thereby has reduced the calculated amount of subsequent treatment, after also being convenient to the transplanting of ARM embedded device.

The color that we say usually can be represented by RGB, HIS, YIQ, YUV color model such as (YCbCr).In the RGB color model, (B) three kinds of colors are represented chromatic information for R, G by color.(R, G B) are converted to the gray level image of 256 looks by calculating.

Adopt the weighted mean value method to carry out gray processing: give R according to importance or other indexs, G, B give different weights, and make R, G, and the value weighting of B, that is:

Gray＝(W _RR+W _GG+W _BB)/(W _R+W _G+W _B)

In the formula, W _R, W _G, W _BBe respectively R, G, the weights of the correspondence of B, W _R, W _G, W _BGet different values, the weighted mean value method will form different gray level images.Because human eye is the highest to the susceptibility of blueness, and red component is taken second place, and is minimum to the susceptibility of blueness, therefore make W _GW _RW _BTo obtain more rational gray level image.Experiment and theoretical derivation prove, work as W _R=0.30, W _G=0.59, W _B, can obtain the most rational gray level image, that is: at=0.11 o'clock

Gray＝0.30×R+0.59×G+0.11B

Adopt following formula to carry out gray processing in the present invention.

(2.2) location of finger areas

Among the present invention, what extract is blood vessel in the finger, and for the part beyond finger contours and the profile, we will not take in, so at first will mark the profile of finger.The present invention adopts the method for iteration to ask for optimal threshold to come image is cut apart.

This method is threshold value of hypothesis in starting condition at first, and constantly upgrades this given threshold in to the interative computation of image, to obtain optimal threshold.Initial threshold is generally got average gray, like this, behind the average gray value split image, the mean value in 2 class zones after the computed segmentation, the mean value that is lower than the initial threshold zone is designated as T _b, another mean value of areas is designated as T ₀, calculate (T then _b+ T ₀)/2, and be worth as new threshold value with this are repeated above-mentioned steps then, no longer change up to the threshold value of calculating for 2 times, have at this moment promptly obtained optimal threshold, iteration stopping.With T ₀As the estimation of initial threshold, then be estimated as for the k time of threshold value in the iteration:

$T_k=\frac{\underset{i=0}{\overset{T_{k-1}}{\mathrm{\Σ}}}i\×h\left[i\right]}{2\underset{i=0}{\overset{T_{k-1}}{\mathrm{\Σ}}}h\left[i\right]}+\frac{\underset{j=T_{k-1}+1}{\overset{N}{\mathrm{\Σ}}}i\×h\left[i\right]}{2\underset{j=T_{k-1}+1}{\overset{N}{\mathrm{\Σ}}}h\left[i\right]}$

(2.3) filtering of image

Because the image that collects contains noise in various degree,, need carry out filtering operation to this width of cloth gray level image in order to remove the salt-pepper noise of this point, bulk.The present invention adopts a kind of junction filter to come filtering noise.

In conjunction with Fig. 2, junction filter is made up of 4 modules altogether: salt-pepper noise detection, selective filter, elimination salt-pepper noise and elimination Gaussian noise.For our given image, earlier detect the pixel that influenced by salt-pepper noise with the salt-pepper noise monitor, can carry out noise removing with the wave filter of median filter one class to these pixels, wave filter to then available smoothing filter one class of remaining pixel carries out the filtering of noise, at last two parts result combinations is got up to obtain to Gaussian noise and salt-pepper noise result of filtering all.

Concrete step is earlier the original vein image of importing to be carried out salt-pepper noise to detect, can select corresponding wave filter to carry out the elimination of salt-pepper noise to the pixel that influenced by salt-pepper noise, then select corresponding wave filter to carry out the elimination of Gaussian noise to the pixel that not influenced by salt-pepper noise, at last two results are merged the finger venous image that obtains after the de-noising, next step the image segmentation of being convenient to us.

(2.4) image cuts apart

Examine given vein image, can find that gray distribution of image is extremely inhomogeneous, the angiosomes gray-scale value that has is identical with the gray-scale value of other local background field, so be can not extract blood vessel exactly with the method for a simple threshold value at all.So the present invention adopts the method for pointwise threshold value to extract blood vessel to filtered image.This is a kind of fairly simple effective local dynamic threshold algorithm, and this basic idea is to the every bit in the image, in its r * r neighborhood, calculates the average and the standard deviation of pixel in the neighborhood, carries out binaryzation with the following formula calculated threshold then:

T(x，y)＝m(x，y)+k×s(x，y)

Wherein, (x, y), (x y) is the threshold value of this point to T, and (x y) is the average of the r * r neighborhood interior pixel point of this point to m, and (x is the standard deviation of the r * r neighborhood interior pixel point of this some y) to s, and k is a correction factor for each pixel.

Like this, just can obtain the model of blood vessel later, but also mix the block noise of many points simultaneously through image segmentation.As shown in Figure 3, so also will remove these noises.

(2.5) denoising after cutting apart

The method that the present invention takes is the area null method.Promptly calculating each piece black background or white background, is noise if background area less than given area, then is this piece, with its removal, as shown in Figure 4.

(2.6) standardization of the cutting of finger areas and finger width

At last, just can reduce out the zone of finger, thereby obtain the model of blood vessel in conjunction with the finger areas array of the extraction of last joint.

Because the randomness of image acquisition and the difference of sample individuality, size, the ratio of the finger blood tube model that extracts at last are often inconsistent, and, need carry out height (i.e. Shou Zhi width) standardization to the vein image after cutting apart in order to make things convenient for further research.Image after the standardization as shown in Figure 5.In experiment, we are standardized as 80 pixels with the height of image.

(3) feature extraction

(3.1) piecemeal of finger venous image

Because the length of different people's finger is different in size, is exactly same individual's finger, the length of gathering constantly in difference is also not necessarily identical.Therefore if directly image is asked for the small echo square, is carried out the PCA conversion and extract feature, certainly will make the feature representative the error of individuality strengthen, thereby discrimination is reduced.The present invention's employing solves the problems referred to above with the method that vein image is divided into the image block of specific size.

In practical operation, the size of the sub-piece of the image that is divided into is 80 * 80.If directly image is split, the length of given image is generally about 200 pixels, can only be divided into 2 complete subimages.The quantity of feature is very few can to influence discrimination to a certain extent.At original image, every certain length (what get in the experiment is 20 pixels), get an image subblock (size is 80 * 80) in invention, former like this figure just can be split as the 6-7 number of sub images, thereby enough characteristic quantities is arranged for identification.

If (x y) uses matrix form A to the image f after the standardization _{M * n}Represent:

A _mn＝[A ₀，A ₁，A ₂，...，A _n-1]

In the formula: A _iBe column vector, i ∈ [0, n-1].

Width of sub-piece of this definition image be the image after w, height are got standardization for h height (w=80 in the experiment, h=80).When extracting subimage, that get is r (r=20 in the experiment) at interval.

So just can obtain the matrix of subimage:

B ₁＝[A ₀，...，A _w-1]

B ₂＝[A _r，...，A _w-1+r，]

B _k＝[A _kr，...，A _w1+kr]

In the formula $k=[\frac{n-w}{r}+1];$ [x] is for getting the maximum integer less than x.

So just obtained B ₁, B ₂..., B _kBe total to k subgraph, and the size of each subimage is w * h.The back all is at each subimage B when extracting feature _iFinish.The piecemeal of whole finger as shown in Figure 5.

(3.2) extraction of the wavelet decomposition of subimage and small echo square

The small echo square is an invariant of well describing characteristics of image.The small echo square has the characteristics of rotation, translation and the convergent-divergent unchangeability of pair image, so a lot of successful application are arranged in pattern-recognition.

Among the present invention, for each subimage B _i(x, y), its size is w * h.Adopt two-dimentional Mallat decomposition algorithm, can be to B _i(x y) carries out wavelet decomposition, as shown in Figure 7.

If f (x, y)=B _i(x, y) ∈ L ²(R ²) be vein subimage block to be analyzed, its one deck wavelet decomposition is:

$f(x,y)=A_1+D_1^1+D_1^2+D_1^3$

Wherein, A ₁Be the low frequency component under this yardstick (also promptly approaching component), and

Then be the level under this yardstick, vertical and oblique details component.As follows:

$A_1=\underset{(m,n)}{\mathrm{\Σ}}{c}_1(m,n){\mathrm{\φ}}_1(m,n),$

c ₁(m，n)＝<f(x，y)，φ ₁(m，n)>

$D_1^k=\underset{(m,n)}{\mathrm{\&Sigma;}}{d}_1^k(m,n){\mathrm{\&psi;}}_1^k(m,n),$

$d_1^k(m,n)=<f(x,y),{\mathrm{\&psi;}}_1^k(m,n)>$

Wherein, k=1,2,3, c ₁(m n) is low frequency component A ₁Coefficient,

Be the coefficient of 3 high frequency subgraphs, m, n are the horizontal ordinate sequence number of each matrix of coefficients, φ ₁(m n) is the scaling function of wavelet decomposition,

Wavelet function.

According to experiment, finally select the Daub4 wavelet basis to carry out wavelet decomposition, recognition effect is best than other wavelet basiss.That can utilize wavelet decomposition approaches coefficient c ₁(m, n) and high frequency coefficient

Ask for Wavelet moment for images.If

Expression f (x, (p+q) rank small echo square that the low frequency coefficient behind wavelet transformation y) is expressed, Be (p+q) rank small echo square that each high frequency coefficient is expressed, specific as follows:

$\left\{\begin{array}{c}{w}_{p,q}^0=2^{-(p+q+1)j}\underset{m,n\&Element;Z}{\mathrm{\&Sigma;}}{m}^p{n}^q{c}_1(m,n)\\ w_{p,q}^k=2^{-(p+q+1)j}\underset{m,n\&Element;Z}{\mathrm{\&Sigma;}}{m}^p{n}^q{d}_1^k(m,n)\end{array}\right.$

The small echo square vector of taking here is $w_{22}=[w_{22}^0,w_{22}^1,w_{22}^2,w_{22}^3].$

(3.3) PCA conversion

Adopting another advantage of wavelet transformation in invention is to reduce calculated amount.For each subimage B _i, if directly carry out the PCA conversion, the tagsort ability of Ti Quing not only, and also calculated amount is also not little.Low frequency subgraph image set behind the wavelet transformation has suffered most of energy of original image, and the image size only is original 1/4, decomposes, then can reduce calculated amount largely so adopt the low frequency subgraph picture to carry out PCA again.

(3.3.1) transition matrix asks for

With B _iThe low frequency subgraph A of one deck wavelet decomposition ₁As the object of analyzing.In order to carry out the PCA conversion, we earlier must be with A ₁Transform, make it become single-row wh/4 dimension image vector ξ=Vec (A ₁), as shown in Figure 8.

Here, in order to say something, total c the people's who supposes finger vena sample, everyone same finger sample has 5 width of cloth images, and as shown in Figure 9, m people's n sample image can be divided into k _{M, n}Number of sub images piece (Fig. 9 only is for the simple declaration problem, and r=80 comes the sub-piece of abstract image with the interval, is to use r=20 in the reality) for each sample image, is only got preceding k _MinIndividual subgraph participates in calculating:

k _min＝min(k _1，1，k _1，2，...，，k _1，5，k _2，1，...，k _c，5)

With reference to Fig. 9, classification is to be that unit classifies with " low frequency subgraph behind the wavelet transformation of image subblock " in the present embodiment.In order to narrate conveniently, be that unit classifies with " image subblock " here.

For example, 5 width of cloth images of 1# finger, the 1st sub-piece of every width of cloth image constitute the 1st class (total 5 corresponding image subblocks) altogether, and the 2nd image subblock of every width of cloth image constitutes the 2nd class ..., the k of every width of cloth image _MinThe height piece constitutes k _MinClass is so the 1# finger just can form k _MinIndividual class.

Then, the 2# finger is then classified with subimage block, can branch away k _Min+ 1 class, k _Min+ 2 classes ..., k _Min+ k _MinClass.

By that analogy, like this, the available pattern class one total C=c * k of c people's finger _MinClass, i.e. ω ₁, ω ₂..., ω _i..., ω _CThere are 5 corresponding sample to be designated as in the i class: ξ _{I, 1}, ξ _{I, 2}, ξ _{I, 3}, ξ _{I, 4}, ξ _{I, 5}, and all be the column vector of wh/4 dimension.Training sample sum N=5C.

The average of i class training sample is:

${\stackrel{\&OverBar;}{\mathrm{\&xi;}}}_i=\frac{1}{5}\underset{j=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_{i,j}$

The average of all training samples is:

$\stackrel{\&OverBar;}{\mathrm{\&xi;}}=\frac{1}{N}\underset{=i1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_{i,j}$

Then all kinds of overall scatter matrix S _tFor:

$S_t=\frac{1}{N}\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n_i}{\mathrm{\&Sigma;}}}({\mathrm{\&xi;}}_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{\&xi;}}){({\mathrm{\&xi;}}_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{\&xi;}})}^T$

Prove overall scatter matrix S easily _tBe non-negative definite matrix.

We can be in the hope of S so _tEigenvalue ₁, λ ₂..., (these eigenwerts by from big to small order sequence preface λ ₁〉=λ ₂〉=...) and the characteristic of correspondence vector

D the pairing orthonormal proper vector of eigenvalue of maximum constitutes transformation matrix before getting

(3.3.2) Feature Extraction

For each subimage block B _i, through the low frequency subgraph A after the wavelet decomposition ₁, with A ₁Method according to the front is converted into column vector ξ, and the transition matrix P that tries to achieve in the joint on utilizing extracts feature, following formula:

e＝P ^Tξ

E=[e like this ₁, e ₂..., e _d] just as subimage block B _iThe proper vector that extracts of PCA.

Just can obtain all well and good effect when experimental results show that d=120, d=300, when promptly boil down to 300 is tieed up, recognition effect the best.

(3.3.3) LDA mapping

In general, the feature that the PCA method obtains is best description feature, rather than the optimal classification feature.In order to obtain better classifying quality, the present invention adopts the LDA method that the PCA feature is further classified.

For utilizing PCA projection matrix P that each sample is converted into proper vector in the d dimension space behind the dimensionality reduction in the preamble $e_i=[e_1^i,e_2^i,...,e_d^i],$ I=1 wherein, 2 ..., N is the sample sequence number.So here, the design of our sorter is exactly again by the PCA proper vector e of each sample behind the dimensionality reduction ₁, e ₂..., e _NConstitute scatter matrix S between interior scatter matrix of class and class _wAnd S _bCalculate corresponding matrix then Preceding l eigenvalue of maximum characteristic of correspondence vector α ₁, α ₂..., α _lConstitute LDA transition matrix W by this l eigenvalue of maximum characteristic of correspondence vector _LDA=[α ₁, α ₂.., α _l].Next just can utilize LDA transition matrix W _LDAProper vector behind the PCA dimensionality reduction has been carried out the LDA projective transformation again.That is: $z_i=[z_1^i,z_2^i,...,z_{\text{l}}^i]=W_{\mathrm{LDA}}^T{e}_i,$ I=1,2 .., N are the sample sequence number.

Like this, we can replace above-mentioned e proper vector to carry out discriminator with optimal classification feature z proper vector.

(4) coupling and identification

By top wavelet decomposition and PCA conversion, at each subimage B _i, can obtain small echo square w ₂₂Proper vector z with PCA and LDA extraction.So B _iBe characterized as v _i=[w ₂₂Z].

For identification and matching, now to point 1 proper vector group V=[v ₁, v ₂..., v _k] and point 2 proper vector group $V\&prime;=[{v\&prime;}_1,{v\&prime;}_2,...,{v\&prime;}_{k\&prime;}]$ For example is analyzed.

What at first will do is, the length of V and V ' may be different, and promptly k is not necessarily identical with k '.Here we define:

K＝min(k，k′)

Be that K vector done contrast before getting among V and the V '.

Analyze corresponding subimage block v first _iAnd v ' _iGet match condition.

v _i＝[w ₂₂；z]，v′ _i＝[w′ ₂₂；z′]

According to experiment, set two threshold vector w _t, z _tDefinition δ _iBe V and the corresponding B of V ' _iTwo proper vector w of subimage and between w ' Euclidean distance.

The B of definition V and V ' correspondence _iThe small echo moment characteristics w of subimage ₂₂The coupling mark is:

$w\_{\mathrm{mark}}_i=\left\{\begin{array}{cc}\frac{w_t-{\mathrm{\&delta;}}_i}{w_t}& {\mathrm{if}}_{}{\mathrm{\&delta;}}_i<w_t\\ 0& \mathrm{else}\end{array}\right.$

Can be in the hope of the small echo moment characteristics coupling mark of whole finger at last:

$w\_\mathrm{mark}=\underset{i=0}{\overset{K}{\mathrm{\&Sigma;}}}w\_{\mathrm{mark}}_i$

In like manner we can ask for the z proper vector coupling mark z_mark of V and V '.

At last, comprehensive mark:

total_mark＝

s ₁×w_mark+s ₂×z_mark

s ₁, s ₂Be two kinds of proportions that the characteristic matching mark is shared, and s ₁〉=0, s ₂〉=0, s ₁+ s ₂=1.

Like this, whether mate the size of the value that just is converted into total_mark for finger vena 1 and 2, if greater than given coupling score threshold, two fingers mate so, otherwise do not match.Certainly, also can classify by minimum distance classifier, promptly each finger vein features contrasts one by one in finger vein features and the database with waiting to know, seek coupling mark total_mark the highest and greater than given threshold value, want identifying object exactly, thereby finished identification mission.

Claims (3)

1, a kind of finger vein features is extracted and matching and recognition method, it is characterized in that:

(1) image collecting device carries out the collection of finger venous image;

(2) vein image of gathering is carried out pre-service, pre-service comprises: adopt the weighted mean value method carry out gray processing, adopt the method for iteration ask for optimal threshold come to image cut apart, adopt junction filter come filtering noise, the local dynamic thresholding method of employing carry out image segmentation, take after the area null method is cut apart denoising, the vein image after cutting apart is carried out highly standardized processing; Described employing junction filter comes filtering noise, junction filter is made up of salt-pepper noise detection, selective filter, elimination salt-pepper noise and 4 modules of elimination Gaussian noise altogether, original vein image to input carries out the salt-pepper noise detection earlier, can select corresponding wave filter to carry out the elimination of salt-pepper noise to the pixel that influenced by salt-pepper noise, then select corresponding wave filter to carry out the elimination of Gaussian noise to the pixel that not influenced by salt-pepper noise, at last two results are merged the finger venous image that obtains after the de-noising;

(3) by vein image being divided into the image block of specific size, adopt the wavelet decomposition algorithm that subimage is carried out the extraction of wavelet decomposition and small echo square, carry out PCA and LDA shift step and extract feature;

(4) adopt the threshold method of obfuscation to mate and identification.

2, finger vein features according to claim 1 is extracted and matching and recognition method, it is characterized in that: the method for described employing iteration is asked for optimal threshold and is come image cut apart and be: threshold value of hypothesis in starting condition at first, and in to the interative computation of image, constantly upgrade this given threshold, initial threshold is generally got average gray, behind the average gray value split image, the mean value in 2 class zones after the computed segmentation, the mean value that is lower than the initial threshold zone is designated as T _b, another mean value of areas is designated as T ₀, calculate (T then _b+ T ₀)/2, and be worth as new threshold value with this are repeated above-mentioned steps then, no longer change up to the threshold value of calculating for 2 times, have at this moment promptly obtained optimal threshold, iteration stopping.

3, finger vein features according to claim 1 is extracted and matching and recognition method, it is characterized in that: it is to the every bit in the image that the local dynamic thresholding method of described employing carries out image segmentation, in its r * r neighborhood, calculate the average and the standard deviation of pixel in the neighborhood, carry out binaryzation with the following formula calculated threshold then:

T(x，y)＝m(x，y)+k×s(x，y)

Wherein, (x, y), (x y) is the threshold value of this point to T, and (x y) is the average of the r * r neighborhood interior pixel point of this point to m, and (x is the standard deviation of the r * r neighborhood interior pixel point of this some y) to s, and k is a correction factor for each pixel.

Images