CN106326886A - Finger-vein image quality evaluation method and system based on convolutional neural network - Google Patents

Abstract

The invention provides a finger-vein image quality evaluation method and system based on a convolutional neural network. The method includes: labeling the quality of finger-vein gray images, building a training sample set, and using the training sample set to train the convolutional neural network; inputting one optional gray image and a binary image into trained models, selecting the output of second full connection layers in two convolutional neural network models as the depth feature vectors of the input gray image and the binary image; connecting the two depth feature vectors to form a united expression vector, inputting the united expression vector into a support vector machine for training, and using a probability support vector machine to calculate the quality of a predicted finger-vein image. By the evaluation method and system, finger-vein image quality evaluation precision can be increased to a large extent, and the identification performance of a certification system can be improved.

Description

Finger vein image quality appraisal procedure based on convolutional neural networks and assessment system

Technical field

The invention belongs to biometrics identification technology field, particularly to a kind of finger vena based on convolutional neural networks Image quality measure method and assessment system.

Background technology

Along with Internet technology is fast-developing and the growth of information security threats, the most effectively discriminate one's identification to protect individual Urgent problem is become with property safety.Compared with traditional authentication mode such as key and password, based on physiology and behavior Biological characteristic be difficult to be stolen, replicate and lose.Therefore, biometrics has been studied widely and has been successfully applied to In personal identification.Following two can be divided into: 1 outside mode such as face, fingerprint, palmmprint and rainbow based on physiological biological mode Film；2 Internal biological mode: finger vena, palm vein and hand back vein.System based on outside biological mode is subject to attack Hit.Such as, it is easy to steal and forge a width fingerprint template to attack fingerprint recognition system.It is different from external biological mode, In biological mode be positioned at finger epidermis under make it difficult to be stolen and forge, therefore they have higher safety Energy.

Owing to the acquisition process of finger venous image is affected by many factors, such as ambient light, ambient temperature, light dissipate Penetrate, the change of physiological feature and the behavior of user, therefore finger vena identification remains a challenging task.As Fruit can not overcome these factors well, then comprises a large amount of low-quality image in the image of collection.In general, these low-qualitys Spirogram picture eventually reduces the performance of Verification System.Finger vein image quality assessment as a kind of effective solution by Study widely.In existing finger vein image quality assessment algorithm, researcher assumes these factors such as picture contrast Relevant with the quality of image with vein quantity.Then, some description artificially designed are utilized such as: Radon conversion, Gaussian Energy These attributes of model, Gabor filter, curvature measuring.The method existed utilizes mankind's intuition or biometric image quality Priori determine and affect the attribute of quality and utilize the manual son that describes these attributes to be extracted, such as CN101866486 discloses a kind of finger vein image quality judging method, and the method is by obtaining finger venous image Contrast mass fraction, position offset mass mark, effective coverage mass fraction, direction fuzziness mass fraction and then acquisition matter Amount mark adds up by weights and carries out overall merit, sets up the overall merit mass function of finger venous image, and then next right Finger vein image quality is estimated, but these methods yet suffer from following shortcoming:

The 1 very difficult attribute proving to choose by hand is certain is relevant to the picture quality of finger vena.Such as, some based on The high quality graphic the most certified system refusal of human vision or understanding.

2 researcheres can not have an impact the attribute of picture quality by inquiry agency.

Even if 3 these attributes are positively correlated with the quality of image, it is also difficult to set up effective mathematical model and go to retouch them.

Summary of the invention

In order to overcome above-mentioned the deficiencies in the prior art, the invention provides a kind of finger vena based on convolutional neural networks Image quality measure method and assessment system.First, overcome conventional finger vein image quality appraisal procedure rely on intuition or The shortcoming that person's priori carrys out evaluation image quality, it is possible to the most objectively picture quality is evaluated.Second, this invention energy Enough quality tabs automatically producing image, thus decrease the produced hard work of artificial mark and error.3rd, this Bright can automatically from original finger venous image learning to the feature relevant to picture quality, it is to avoid select artificially and carry The problem taking discrimination feature.

The concrete technical scheme of the present invention is as follows:

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, the method comprises the steps:

S1: the quality of finger vena gray level image in data base is labeled, it is thus achieved that with the gray-scale map of quality tab Picture, i.e. marks out low quality gray level image and high-quality gray level image, and obtains the vein of the gray level image with quality tab Feature, obtains bianry image after encoding；

S2: with the bianry image training sample set of quality tab in establishment step S1；

S3: with the gray level image training sample set of quality tab in establishment step S1；

S4: extract the convolutional neural networks model of gray level image depth characteristic；Described convolutional neural networks model includes: defeated Enter layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the 3rd convolutional layer, the first full articulamentum, second Full articulamentum and output layer；

S5: extract the convolutional neural networks model of bianry image depth characteristic；Described convolutional neural networks model includes: defeated Enter layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the first full articulamentum, the second full articulamentum and Output layer；

S6: the training of convolutional neural networks model

The random number utilizing Gaussian distributed initializes each layer wave filter, and the initial value of side-play amount is arbitrary constant；Adopt By stochastic gradient descent method, convolutional neural networks is trained；By step S2 set up bianry image training sample set and The gray level image training sample set that step S3 is set up is divided into different subclass, is separately input to step S5 and step in batches In the convolutional neural networks model that rapid S4 is applied, when the image of all batches carries out a forward direction at convolutional neural networks model After propagation, calculate gradient and carry out back propagation with update wave filter power and side-play amount, by iterate searching wave filter and The optimal solution of skew；

S7: after completing training, is input to the convolutional neural networks mould of step S4 and step S5 by prediction finger venous image In type, in selecting step S4 and S5 step, in convolutional neural networks model, the second full articulamentum is output as inputting a width gray-scale map The depth characteristic vector of picture and bianry image；Connect two depth characteristic vectors and form a width input prediction finger venous image Combined expression vector；

S8: be input in support vector machine be trained by the Combined expression vector that step S7 is formed, uses probability support Vector machine calculates the quality of prediction finger venous image.

Further improve, the quality of finger vena gray level image in data base be labeled method particularly includes:

S11: the selection of enrollment image

Select a finger appoints piece image, utilizes ripe recognizer method to extract and coupling two width finger venas Image, and calculate this image and remaining width image averaging value distance；Select the image of minimum average B configuration distance correspondence as this finger Enrollment image, other image as test image；

The mark of S12: picture quality

Calculate in the distance between every width test image and its enrollment image of same finger obtains class and mate mark； Calculate the distance between each enrollment image and obtain mating between class mark；Mark is mated between mark and class according to mating in class, The receptance FAR of mistake in computation and the reject rate FRR of mistake；Preset a threshold value, when the receptance FAR of mistake is equal to this threshold value Time such as FAR=0.1%, then according to whether the image refused by system mistake or the correct image accepted are to distinguish low-quality spirogram Picture or high quality graphic.

Further improve, in first volume lamination, volume Two lamination or the 3rd convolutional layer, the characteristic image of l layerPress Calculate according to equation below:

$z_n^l=\underset{0<m\&le;M^{l-I}}{max}(w_{n,m}^l*x_m^l)+b_n^l$

Wherein,It is the input spectrum of l layer,Being the convolution kernel between m-th input and n output characteristic spectrum, * is Convolution operation, M^l-1It is the quantity of input feature vector spectrum,It it is the skew of the n-th output spectra.

Further improve, first volume lamination, volume Two lamination or the 3rd convolutional layer use and revises linear unit conduct Excitation function, it is defined as follows:

$y_n^l=max(z_n^l,0)$

Wherein,Represent the output spectra of l layer.

Further improve, in the first pond layer, the second pond layer, the output of first volume lamination and volume Two lamination is special Levy spectrum and divide the region of non-overlapping copies, choose the representative value as this region of the average of p maximum before in each region to the The output of one convolutional layer and volume Two lamination is sampled；If I^kRepresent output spectra after kth convolution kernel carries out convolution,Represent I^kAll elements in middle s × s regional areaCarry out from greatly to The set obtained after little sequence, T=s × s represents the number of element；I^kThe output characteristic obtained after after samplingAccording to as follows Formula calculates:

$R_{i,j}^k=\frac{\underset{t=0}{\overset{p-1}{\&Sigma;}}{c}_t^k}{p},(p\&le;T).$

Further improve, kth step wave filter power w_kMore new regulation be:

${\mathrm{\&Delta;}}_{k+1}=0.9\&CenterDot;{\mathrm{\&Delta;}}_k-0.004\&CenterDot;\mathrm{\&lambda;}\&CenterDot;w_k-\mathrm{\&lambda;}\&CenterDot;\frac{\&part;L}{\&part;w_k}$

w_k+1=Δ_k+1+w_k

Wherein Δ represents momentum, and λ is learning rate,For w_k.Gradient.

Further improving, the probabilistic SVMs of described use is by combined depth characteristic vector V and its matter Amount label q ∈ 0,1}, and probabilistic SVMs is trained, its output probability value is p

$p(q=1|\mathrm{\&xi;}\left(v\right))=\frac{1}{1+\exp (\mathrm{\&omega;}\&CenterDot;\mathrm{\&xi;}(v)+\mathrm{\&gamma;})}$

ξ (v) represents the output of traditional support vector machine, ω and γ represents that probabilistic SVMs trains two ginsengs obtained Number.

Another aspect of the present invention provides a kind of finger vein image quality based on convolutional neural networks assessment system, and this is commented Estimating the data base that system includes assessing unit and communicating with assessment unit, described databases contains finger vena gray-scale map Picture, described assessment unit includes:

Quality annotation module, for being labeled the quality of finger vena gray level image, it is thus achieved that with quality tab Gray level image, and obtain the vein pattern of the gray level image with quality tab, obtain bianry image after encoding；

Training sample set builds formwork erection block jointly, for by quality annotation module obtain the bianry image with quality tab and Gray level image sets up bianry image training sample set and gray level image training sample set respectively；

Model building module, extracts bianry image and the convolutional neural networks of gray level image depth characteristic for setting up respectively Model；

Convolution training module, for building bianry image training sample set and the ash that formwork erection block is set up jointly by training sample set Degree image training sample set is divided into different subclass, is separately input in batches extract bianry image and the gray level image degree of depth In the convolutional neural networks model that feature is corresponding, it is trained；

Connect processing module, for obtaining gray level image and bianry image in the convolutional neural networks model after training Depth characteristic vector；And form Combined expression vector for connecting two depth characteristic vectors；

Computing module, for being input in support vector machine be trained by Combined expression vector, calculates prediction finger The quality of vein image.

Further improving, quality annotation module includes:

Image selects and feature extraction submodule, for optionally selecting a width gray-scale map from some images of same finger As and carry out feature extraction and obtain bianry image；

Enrollment image selects submodule, is used for calculating image and selects with the image of feature extraction submodule selection with same A piece finger is left the meansigma methods distance of image, selects minimum average B configuration apart from corresponding image as enrollment image, other Image is as test image；

Calculating sub module, the distance between every width test image and its enrollment image calculating same finger obtains In class, mate mark, calculate the distance between each enrollment image and obtain mating between class mark；And according to coupling point in class Mark, the receptance FAR of mistake in computation and the reject rate FRR of mistake is mated between number and class；

Judge submodule, for wrongheaded receptance FAR whether equal to the threshold value preset, if the receptance of mistake FAR is equal to this threshold value, sends sort instructions to classification submodule；

Classification submodule, for classifying the image and the correct image accepted that are labeled with False Rejects, and to mark Note submodule sends the instruction of mark；

Mark submodule, for the image or the correct image accepted that are labeled with False Rejects are carried out quality annotation, and Set corresponding quality tab.

Compared with prior art, the invention has the beneficial effects as follows: the finger based on convolutional neural networks that the present invention provides Vein image quality appraisal procedure and assessment system, it is possible to largely promote the precision of finger vein image quality assessment, Improve the recognition performance of Verification System, compared with other finger vein image quality appraisal procedures, its have the beneficial effect that with Under several aspects:

1. the present invention provide finger vein image quality appraisal procedure based on convolutional neural networks and assessment system energy Enough automatically finger vena gray level image is labeled, thus decreases and artificially mark the hard work brought and error.

2. the present invention provides finger vein image quality appraisal procedure based on convolutional neural networks and assessment system are first The depth characteristic of secondary fusion finger vena bianry image and gray level image realizes the quality of finger vein image and estimates.

3., compared with traditional convolutional neural networks model, the convolutional neural networks model difference that the present invention uses is: First, for all convolutional layers, between employing calculating this layer of characteristic spectrum of input, the maximum of correspondence position element is as this layer Characteristic spectrum is also entered in activation primitive；Second, in the layer of all ponds, calculate front p in the regional area of characteristic image Input feature vector image is sampled by the average of individual maximum.

4. the present invention utilizes probability support vector merge depth characteristic and predict the quality of finger venous image, from And it is effectively improved the precision of image quality measure.

5. the present invention provide finger vein image quality appraisal procedure based on convolutional neural networks and assessment system, no It is only applicable to the quality evaluation of finger venous image, and may apply in other biological characteristic image quality evaluation.

Accompanying drawing explanation

Fig. 1 is the flow chart of embodiment 1 finger vein image quality based on convolutional neural networks appraisal procedure；

Fig. 2 is the convolutional neural networks model structure schematic diagram that embodiment 4 extracts the depth characteristic of gray level image；

Fig. 3 is the convolutional neural networks model structure schematic diagram that embodiment 4 extracts the depth characteristic of bianry image；

Fig. 4 is the structured flowchart of embodiment 6 finger vein image quality based on convolutional neural networks assessment system；

Fig. 5 is the structured flowchart of embodiment 7 quality annotation module.

Detailed description of the invention

Embodiment 1

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, as it is shown in figure 1, the method include as Lower step:

S1: the quality of finger vena gray level image in data base is labeled, it is thus achieved that with the gray-scale map of quality tab Picture, it is thus achieved that low quality gray level image and high-quality gray level image, and obtain the vein pattern of the gray level image with quality tab, Bianry image is obtained after encoding；

S2: the bianry image training sample set with quality tab obtained in establishment step S1；

S3: with the gray level image training sample set of quality tab in establishment step S1；

S4: extract the convolutional neural networks model of gray level image depth characteristic；Described convolutional neural networks model includes: defeated Enter layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the 3rd convolutional layer, the first full articulamentum, second Full articulamentum and output layer；

S5: extract the convolutional neural networks model of bianry image depth characteristic；Described convolutional neural networks model includes: defeated Enter layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the first full articulamentum, the second full articulamentum and Output layer；

S6: the training of convolutional neural networks model

The random number utilizing Gaussian distributed initializes each layer wave filter, and the initial value of side-play amount is arbitrary constant；Adopt By stochastic gradient descent method, convolutional neural networks is trained；By step S2 set up bianry image training sample set and The gray level image training sample set that step S3 is set up is divided into different subclass, is separately input to step S5 and step in batches In the convolutional neural networks model that rapid S4 is applied, when the image of all batches carries out a forward direction at convolutional neural networks model After propagation, calculate gradient and carry out back propagation with update wave filter power and side-play amount, by iterate searching wave filter and The optimal solution of skew；

S7: after completing training, is input to the convolutional neural networks mould of step S4 and step S5 by prediction finger venous image In type, in selecting step S4 and S5 step, in convolutional neural networks model, the second full articulamentum is output as inputting a width gray-scale map The depth characteristic vector of picture and bianry image；Connect two depth characteristic vectors and form a width input prediction finger venous image Combined expression vector；

S8: be input in support vector machine be trained by the Combined expression vector that step S7 is formed, uses probability support Vector machine calculates the quality of prediction finger venous image.

The finger vein image quality appraisal procedure based on convolutional neural networks that the present invention provides can be largely Promoting the precision of finger vein image quality assessment, improve the recognition performance of Verification System, the method that the present invention provides is melted first The depth characteristic closing finger vena bianry image and gray level image realizes the quality estimation of finger vein image.

Embodiment 2

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, described method is different from embodiment 1 , described the quality of finger vena gray level image in data base is labeled method particularly includes:

S11: the selection of enrollment image

Select a finger appoints piece image, utilizes ripe recognizer method to extract and coupling two width finger venas Image, and calculate this image and remaining width image averaging value distance；Select the image of minimum average B configuration distance correspondence as this finger Enrollment image, other image as test image；

The mark of S12: picture quality

Calculate in the distance between every width test image and its enrollment image of same finger obtains class and mate mark； Calculate the distance between each enrollment image and obtain mating between class mark；Mark is mated between mark and class according to mating in class, The receptance FAR of mistake in computation and the reject rate FRR of mistake；Preset the threshold value safe class as system, when mistake Receptance FAR equal to during this threshold value such as FAR=0.1%, then according to whether the image refused by system mistake or correct acceptance Image distinguishes low-quality image or high quality graphic.

Finger venous image in the sample set that the present invention provides derives from the data base http of The Hong Kong Polytechnic University: // Www4.comp.polyu.edu.hk/～csajaykr/fvdatabase.htm.This data base comprises the 3132 width handss of 156 people Refer to vein image；In two stages, in each acquisition phase, each finger provides 6 width image patterns, often in the collection of data Individual provides two fingers, so everyone provides 24 width images two stages；Wherein, front 105 people gather rank at two Section provides 2520 images；Remaining 51 people are only involved in the image acquisition of second stage, a total of 612 width images.Due to Two phase acquisition images more tally with the actual situation, so the present invention has simply used 2520 images gathered from 105 people It is introduced as a example by width image × 2 stage of finger × 6,105 people × 2 piece.

Illustrating the quality of finger venous image of the present invention, to be labeled concrete grammar as follows:

The extraction of first step vein pattern with mate

The extraction of 1.1 vein patterns: image is strengthened mainly by following Gabor wavelet

$G{\left(p\right)}_{\mathrm{\&theta;}}=\frac{1}{2p|C{|}^{\frac{1}{2}}}{\mathrm{cos\&omega;}}_m^T{(p_n-p_0)}^{{\&lsqb;-\frac{1}{2}(p_n-p_0)C^{-1}(p_n-p_0)\&rsqb;}^T}$

In formula, p_n=[x, y]^TRepresent at coordinate axes both horizontally and vertically, p₀=[x₀,y₀]^TIt is the distance to initial point, ω_mBeing flat rate, C is 2 × 2 positive definite covariance matrixes, | | represent dot-product operation.Pass through coordinate transformWithThe Gabor filter of not Tongfang can be arrived, whereinθ_nBe the anglec of rotation and by from Dispersion K direction is such asWherein q=1,2 ..., K (K=8).

Finger vein features is strengthened by equation below；

$F(x,y)=max\left(\stackrel{\&OverBar;}{G_{\mathrm{\&theta;}}}\right(x,y)*f(x,y\left)\right)$

WhereinRepresent G_θ(x, average y)) .* represents volume Long-pending, (x y) is finger venous image to f.

Morphological operation is utilized to further enhance vein pattern；

$Z(x,y)=F(x,y)-\left(F\right(x,y)\&CirclePlus;b)\&CircleTimes;b$

WithRepresent, by structural element b, image is carried out gray scale expansion and corrosion.

Then, utilize equation below that characteristic image Z is carried out coding and obtain bianry image；

$R(x,y)=\left\{\begin{array}{cc}255& ifZ(x,y)>0\\ 0& ifZ(x,y)\&le;0\end{array}\right.$

The coupling of vein pattern mainly realizes by the following method:

The coupling of 1.2 vein patterns:

Assume two-value registered images and test image that R and T represents m × n respectively；By being extended obtaining template to R ImageSuch as by its length and width are expanded to 2w+m and 2h+n, to obtain template representation as follows:

$\stackrel{\&OverBar;}{R}(x,y)=\left\{\begin{array}{cc}R(x-w,y-h)& x\&Element;\&lsqb;w+1,w+m\&rsqb;,y\&Element;\&lsqb;h+1,h+n\&rsqb;\\ -1& otherwise\end{array}\right.$

Coupling mark between R and T is calculated as follows:

$N(T,R)=\underset{0\&le;i\&le;2w,0\&le;j\&le;2h}{min}\left(\frac{\underset{x=1}{\overset{m}{\&Sigma;}}\underset{y=1}{\overset{n}{\&Sigma;}}\mathrm{\&Phi;}\left(T\right(x,y),R(x,y\left)\right)}{m\&times;n}\right)$

In formula, w and h represents the distance moved in the horizontal and vertical directions；

Φ is defined as follows:

$\mathrm{\&Phi;}(X,Y,M,N)=\left\{\begin{array}{cc}1& ifX-Y=255\\ 0& else\end{array}\right..$

The selection of second step enrollment image

Data base there are 210 fingers, every finger have 12 width finger venous images.All images are through method above Extract vein pattern bianry image.The most optionally take a width bianry image and utilize aforesaid matching algorithm calculate this image with Remaining all image averaging value distances.Repeat this operation, calculate the average distance of other images.Finally select minimum average B configuration distance Corresponding image is as the template image of this finger, and other image is as test image.Therefore, data have 210 width registration moulds Plate and 2310 width test image.

The mark of the 3rd step picture quality

The distance between same finger every width test image and its template image, common property raw 2320 is calculated according to matching algorithm Coupling mark in class.Correspondingly, can obtain mating between 21945 classes mark by the distance between calculation template.According between class With receptance FAR and the reject rate (FRR) of mistake mating mark mistake in computation in class；At one, there is relatively high safety grade Under the conditions of FAR equal to 0.1%, 0.1% be the value preset, and the image that mark is refused by system mistake is low-quality image, and it is marked Label are set as 0；The image labeling correctly accepted by system is high quality graphic, and its label is set as 1.

The finger vein image quality appraisal procedure based on convolutional neural networks that the present invention provides and other finger venas Image quality measure method is compared, it is possible to automatically finger vein image is labeled, thus decreases artificial mark and bring Hard work and error.

Embodiment 3

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, described method is different from embodiment 1 , the bianry image training sample set with quality tab of acquisition in step S2 establishment step S1 method particularly includes: After mask method according to step S1 marks all test images, choose 1155 width images of 105 fingers as training image, Remaining image is as test image；In training set, a total of 101 width low-quality images and 1054 panel height quality images. In test set, high-quality and low-quality image are respectively 110 width and 1045 width.Owing in training set, low-mass ratio is high The sample of quality is few, causes all kinds of imbalance；In order to overcome this problem, method below is utilized to produce low-quality image； Such as, in order to produce the synthesis sample of low-quality image x, first optional two width low-quality image x from training set₁And x₂.So After, utilize equation y_l=x₁+rand(0,1)(x₂-x₁) (l=1,2 ..., L) produce an interim image pattern.Finally, logical Cross equation p_l=x₁+rand(0,1)(y_l-x) (l=1,2 ..., K) calculate the new samples synthesized.According to the method, can produce The image of 953 width synthesis, so that total low-quality image is 1054 width in training set.

Gray level image training sample set with quality tab in establishment step S1 in step S3 method particularly includes: by In the corresponding width gray level image of every width bianry image, therefore according to bianry image training sample set above can obtain based on The gray level image training sample set of gray level image.

Embodiment 4

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, described method is different from embodiment 1 , step S4 is extracted the convolutional neural networks model of gray level image depth characteristic, as in figure 2 it is shown, described first volume lamination, In volume Two lamination or the 3rd convolutional layer, the characteristic image of l layerIn each element right equal in all characteristic patterns of last layer Answer the maximum of position, its characteristic imageCalculate according to equation below:

$z_n^l=\underset{0<m\&le;M^{l-I}}{max}(w_{n,m}^l*x_m^l)+b_n^l$

Wherein,It is the input spectrum of l layer,Being the convolution kernel between m-th input and n output characteristic spectrum, * is Convolution operation, M^l-1It is the quantity of input feature vector spectrum,It it is the skew of the n-th output spectra.

In described first volume lamination, volume Two lamination or the 3rd convolutional layer, use correction linear unit is as excitation function, It is defined as follows:

$y_n^l=max(z_n^l,0)$

In formula,Represent the output spectra of l layer.

The output characteristic of first volume lamination and volume Two lamination is composed by described first pond layer, the second pond layer and divides the most not Overlapping region, chooses the representative value as this region of the average of p maximum before in each region to first volume lamination or the The output of two convolutional layers is sampled；If I^kRepresent output spectra after kth convolution kernel carries out convolution, Represent I^kAll elements in middle s × s regional areaObtain after sorting from big to small Set, T=s × s represents the number of element；To I^kThe output characteristic obtained after samplingCalculate according to equation below:

$R_{i,j}^k=\frac{\underset{t=0}{\overset{p-1}{\&Sigma;}}{c}_t^k}{p},(p\&le;T).$

Described first full articulamentum and the second full articulamentum use and abandons method and discharge the neuron of half randomly.

In described output layer, utilize softmax function to predict the probability of N=2 class；

$y_m=\frac{\exp \left(z_m\right)}{{\mathrm{\&Sigma;}}_{n=1}^N\exp \left(z_n\right)}$

Wherein,The output x of last hidden layer_mLinear combination.

Step S5 extracts the convolutional neural networks model of bianry image depth characteristic as shown in Figure 3.

Compared with traditional convolutional neural networks model, the convolutional neural networks model difference that the present invention uses is: the One, for all convolutional layers, between employing calculating this layer of characteristic spectrum of input, the maximum of correspondence position element is as the spy of this layer Levy spectrum and be entered in activation primitive；Second, in the layer of all ponds, before calculating in the regional area of characteristic image, p is individual Input feature vector image is sampled by the average of maximum.

Embodiment 5

A kind of finger vein image quality appraisal procedure based on convolutional neural networks, described method is different from embodiment 4 , the training of convolutional neural networks model method particularly includes:

1. the random number utilizing Gaussian distributed initializes each layer wave filter, and the initial value of side-play amount is arbitrary constant； Stochastic gradient descent method is used to carry out the convolutional neural networks model shown in training step S4 and step S5.

2. for piece image F, its quality tab is that { 0,1}, wherein 0 represents low-quality image to q ∈, and 1 represents high-quality Spirogram picture；Training set is expressed as { (F₁,q₁),(F₂,q₂),…,(F_N,q_N)}；Training dataset is divided into different subclass, It is input to step S4 in batches and in convolutional neural networks model that step S5 is applied；When the image of all batches enters at network After propagated forward of row, calculate gradient and carry out back propagation to update wave filter power w_kWith side-play amount b_k；Such as: kth step is repeatedly Weight w in Dai_kMore new regulation be:

${\mathrm{\&Delta;}}_{k+1}=0.9\&CenterDot;{\mathrm{\&Delta;}}_k-0.004\&CenterDot;\mathrm{\&lambda;}\&CenterDot;w_k-\mathrm{\&lambda;}\&CenterDot;\frac{\&part;L}{\&part;w_k}$

w_k+1=Δ_k+1+w_k

Wherein Δ represents momentum, and λ is learning rate,For w_k.Gradient.

3. by searching wave filter and the optimal solution of skew of iterating.When precision meet require time, stop iteration, thus Complete the training of this deep neural network model.

Step S7 method particularly includes: after completing training, remove the output layer of convolutional neural networks, when inputting a width gray scale Image is in the convolutional neural networks model of step S4, and the second full articulamentum will export a depth characteristic vector；This vector is i.e. The degree of depth for input gray level image is expressed；When inputting a width bianry image in the convolutional neural networks model of step S5, second The depth characteristic vector of full articulamentum one two straight image of output；Assume v₁And v₂It is respectively a width gray level image and corresponding two The depth characteristic vector of value image；The vector of the Combined expression to a width input picture is formed by connecting two depth characteristic vectors V=[v₁ v₂], then, it is input in support vector machine be trained by this vector；

Step S8 method particularly includes: Evaluation Model on Quality based on support vector machine:

During image quality measure based on support vector machine closes, probabilistic SVMs is used to carry out the quality of prognostic chart picture. Its probabilistic SVMs used is defined as follows: by combined depth characteristic vector v and its quality tab q ∈ 0,1}, Being trained probabilistic SVMs, its output probability value is p

$p(q=1|\mathrm{\&xi;}\left(v\right))=\frac{1}{1+\exp (\mathrm{\&omega;}\&CenterDot;\mathrm{\&xi;}(v)+\mathrm{\&gamma;})}$

ξ (v) represents the output of traditional support vector machine, ω and γ represents that probabilistic SVMs trains two ginsengs obtained Number.After training, probabilistic SVMs can calculate the quality of any input feature value v corresponding image.

The present invention utilizes probability support vector that depth characteristic merges and predicts the quality of finger venous image, thus It is effectively improved the precision of image quality measure.

Embodiment 6

A kind of finger vein image quality based on convolutional neural networks assessment system, as shown in Figure 4, this assessment system bag Including assessment unit 1 and the data base 2 communicated with assessment unit 1, described data base 2 internal memory contains finger vena gray level image, Described assessment unit 1 includes:

Quality annotation module 11, for being labeled the quality of finger vena gray level image, it is thus achieved that with quality tab Gray level image, and obtain the vein pattern of the gray level image with quality tab, after encoding, obtain bianry image；

Training sample set builds formwork erection block 12 jointly, for the binary map with quality tab quality annotation module 11 obtained Picture and gray level image set up bianry image training sample set and gray level image training sample set respectively；

Model building module 13, extracts bianry image and the convolutional Neural net of gray level image depth characteristic for setting up respectively Network model；

Convolution training module 14, for building the bianry image training sample set that formwork erection block 12 is set up jointly by training sample set It is divided into different subclass with gray level image training sample set, is separately input in batches extract bianry image and gray level image In the convolutional neural networks model that depth characteristic is corresponding, it is trained；

Connect processing module 15, for obtaining gray level image and binary map in the convolutional neural networks model after training The depth characteristic vector of picture；And form Combined expression vector for connecting two depth characteristic vectors；

Computing module 16, for being input in support vector machine be trained by Combined expression vector, calculates prediction hands Refer to the quality of vein image.

Finger vein image quality based on the convolutional neural networks assessment system energy that the present invention that the present invention provides provides Enough precision largely promoting finger vein image quality assessment, improve the recognition performance of Verification System, with other fingers Vein image quality appraisal procedure is compared the depth characteristic of fusion finger vena bianry image and gray level image first and is realized opponent Refer to that the quality of vein image is estimated.

Embodiment 7

A kind of finger vein image quality based on convolutional neural networks assessment system, this assessment system is with embodiment 6 not With, as it is shown in figure 5, quality annotation module 11 includes:

Image selects and feature extraction submodule 110, for optionally selecting a width ash from some images of same finger Degree image also carries out feature extraction and obtains bianry image；

Enrollment image selects submodule 111, for calculating the figure that image selects and feature extraction submodule 110 selects As being left the meansigma methods distance of image with same finger, select the image of minimum average B configuration distance correspondence as enrollment figure Picture, other image is as test image；

Calculating sub module 112, for calculate same finger every width test image and its enrollment image between away from In obtaining class, mate mark, calculate the distance between each enrollment image and obtain mating between class mark；And according in class Mark, the receptance FAR of mistake in computation and the reject rate FRR of mistake is mated between partition number and class；

Judge submodule 113, for wrongheaded receptance FAR whether equal to the threshold value preset, if the connecing of mistake By rate FAR equal to this threshold value, send sort instructions to classification submodule 114；

Classification submodule 114, for the image and the correct image accepted that are labeled with False Rejects are classified, and to Mark submodule 115 sends the instruction of mark；

Mark submodule 115, for the image or the correct image accepted that are labeled with False Rejects are carried out quality annotation, And set corresponding quality tab.

Finger vein image quality based on convolutional neural networks assessment system and other finger venas that the present invention provides Image quality measure system is compared, it is possible to automatically finger vein image is labeled, thus decreases artificial mark and bring Hard work and error.And the present invention provides assessment system finger vein image to be labeled more accurately, carries High mark quality.

Finally it should be noted that above example is only in order to illustrate technical scheme and unrestricted, by this The technical scheme of invention is modified or equivalent, without deviating from the spirit and scope of technical solution of the present invention, and all should Contain in the middle of scope of the presently claimed invention.

Claims (9)

1. a finger vein image quality appraisal procedure based on convolutional neural networks, it is characterised in that described method includes Following steps:

S1: the quality of finger vena gray level image in data base is labeled, it is thus achieved that with the gray level image of quality tab, and Obtain the vein pattern of the gray level image with quality tab, after encoding, obtain bianry image；

S2: the bianry image training sample set with quality tab obtained in establishment step S1；

S3: with the gray level image training sample set of quality tab in establishment step S1；

S4: extract the convolutional neural networks model of gray level image depth characteristic；Described convolutional neural networks model includes: input Layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the 3rd convolutional layer, the first full articulamentum, second complete Articulamentum and output layer；

S5: extract the convolutional neural networks model of bianry image depth characteristic；Described convolutional neural networks model includes: input Layer, first volume lamination, the first pond layer, volume Two lamination, the second pond layer, the first full articulamentum, the second full articulamentum and defeated Go out layer；

S6: the training of convolutional neural networks model

The random number utilizing Gaussian distributed initializes each layer wave filter, and the initial value of side-play amount is arbitrary constant；Use with Convolutional neural networks is trained by machine gradient descent method；The bianry image training sample set that step S2 is set up and step The gray level image training sample set that S3 sets up is divided into different subclass, is separately input to step S5 and step S4 in batches In the convolutional neural networks model applied, when the image of all batches carries out a propagated forward at convolutional neural networks model After, calculate gradient and carry out back propagation to update wave filter power and side-play amount, by the searching wave filter and skew that iterates Optimal solution；

S7: after completing training, is input to prediction finger venous image in the convolutional neural networks model of step S4 and step S5, In selecting step S4 and S5 step in convolutional neural networks model the second full articulamentum be output as inputting a width gray level image and The depth characteristic vector of bianry image；Connect two depth characteristic vectors and form the associating of a width input prediction finger venous image Express vector；

S8: be input in support vector machine be trained by the Combined expression vector that step S7 is formed, uses probability to support vector Machine calculates the quality of prediction finger venous image.

Appraisal procedure the most according to claim 1, it is characterised in that described to finger vena gray level image in data base Quality is labeled method particularly includes:

S11: the selection of enrollment image

Select a finger appoints piece image, utilizes ripe recognizer method to extract and coupling two width finger vena figures Picture, and calculate this image and remaining width image averaging value distance；Select the image of minimum average B configuration distance correspondence as this finger Enrollment image, other image is as test image；

The mark of S12: picture quality

Calculate in the distance between every width test image and its enrollment image of same finger obtains class and mate mark；Calculate Distance between each enrollment image obtains mating between class mark；Mate mark between mark and class according to mating in class, calculate The receptance FAR of mistake and the reject rate FRR of mistake；Preset a threshold value, when receptance FAR is equal to the threshold value preset, then According to whether be labeled as the image of False Rejects or the image of correct acceptance to distinguish low quality gray level image or height by system Quality gray image.

Appraisal procedure the most according to claim 1, it is characterised in that described first volume lamination, volume Two lamination or the 3rd In convolutional layer, the characteristic image of l layerCalculate according to equation below:

$z_n^l=\underset{0<m\&le;M^{l-I}}{max}(w_{n,m}^l*x_m^l)+b_n^l$

Wherein,It is the input spectrum of l layer,Being the convolution kernel between m-th input and n output characteristic spectrum, * is convolution Operation, M^l-1It is the quantity of input feature vector spectrum,It it is the skew of the n-th output spectra.

Appraisal procedure the most according to claim 3, it is characterised in that described first volume lamination, volume Two lamination or the 3rd Using correction linear unit as excitation function in convolutional layer, it is defined as follows:

$y_n^l=max(z_n^l,0)$

Wherein,Represent the output spectra of l layer.

Appraisal procedure the most according to claim 1, it is characterised in that by described first pond layer, the second pond layer The output characteristic spectrum of one convolutional layer and volume Two lamination divides the region of non-overlapping copies, front p maximum in choosing each region Average as the representative value in this region, the output of first volume lamination or volume Two lamination is sampled；If I^kRepresent through the K convolution kernel carries out output spectra after convolution,Represent I^kAll elements in middle s × s regional areaThe set obtained after sorting from big to small, T=s × s represents the number of element；To I^kAdopt The output characteristic obtained after sampleCalculate according to equation below:

$R_{i,j}^k=\frac{\underset{t=0}{\overset{p-1}{\&Sigma;}}{c}_t^k}{p},(p\&le;T).$

Appraisal procedure the most according to claim 1, it is characterised in that described kth step wave filter power w_kMore new regulation be:

${\mathrm{\&Delta;}}_{k+1}=0.9\&CenterDot;{\mathrm{\&Delta;}}_k-0.004\&CenterDot;\mathrm{\&lambda;}\&CenterDot;w_k-\mathrm{\&lambda;}\&CenterDot;\frac{\&part;L}{\&part;w_k}$

w_k+1=Δ_k+1+w_k

Wherein Δ represents momentum, and λ is learning rate,For w_k.Gradient.

Appraisal procedure the most according to claim 1, it is characterised in that the probabilistic SVMs of described use is by connection Conjunction depth characteristic vector V and its quality tab q ∈ 0,1}, and probabilistic SVMs is trained, its output probability value is p

$p(q=1|\mathrm{\&xi;}\left(v\right))=\frac{1}{1+\exp (\mathrm{\&omega;}\&CenterDot;\mathrm{\&xi;}(v)+\mathrm{\&gamma;})}$

ξ (v) represents the output of traditional support vector machine, ω and γ represents that probabilistic SVMs trains two parameters obtained.

8. finger vein image quality based on a convolutional neural networks assessment system, it is characterised in that described assessment system Including assessing unit (1) and the data base (2) communicated with assessment unit (1), described data base (2) internal memory contains finger vena Gray level image, described assessment unit (1) including:

Quality annotation module (11), for being labeled the quality of finger vena gray level image, it is thus achieved that with quality tab Gray level image, and obtain the vein pattern of the gray level image with quality tab, obtain bianry image after encoding；

Training sample set builds formwork erection block (12) jointly, for the binary map with quality tab quality annotation module (11) obtained Picture and gray level image set up bianry image training sample set and gray level image training sample set respectively；

Model building module (13), extracts bianry image and the convolutional neural networks of gray level image depth characteristic for setting up respectively Model；

Convolution training module (14), for building the bianry image training sample set that formwork erection block (12) is set up jointly by training sample set It is divided into different subclass with gray level image training sample set, is separately input in batches extract bianry image and gray level image In the convolutional neural networks model that depth characteristic is corresponding, it is trained；

Connect processing module (15), for obtaining gray level image and bianry image in the convolutional neural networks model after training Depth characteristic vector；And form Combined expression vector for connecting two depth characteristic vectors；

Computing module (16), for being input in support vector machine be trained by Combined expression vector, calculates prediction finger The quality of vein image.

Assessment system the most according to claim 8, it is characterised in that described quality annotation module (11) including:

Image selects and feature extraction submodule (110), for optionally selecting a width gray scale from some images of same finger Image also carries out feature extraction and obtains bianry image；

Enrollment image selects submodule (111), is used for calculating the figure that image selects and feature extraction submodule (110) selects As being left the meansigma methods distance of image with same finger, select the image of minimum average B configuration distance correspondence as enrollment figure Picture, other image is as test image；

Calculating sub module (112), the distance between every width test image and its enrollment image calculating same finger Obtain coupling mark in class, calculate the distance between each enrollment image and obtain mating between class mark；And mate according in class Mark, the receptance FAR of mistake in computation and the reject rate FRR of mistake is mated between mark and class；

Judge submodule (113), for wrongheaded receptance FAR whether equal to the threshold value preset, if the acceptance of mistake Rate FAR is equal to this threshold value, sends sort instructions to classification submodule (114)；

Classification submodule (114), for classifying the image and the correct image accepted that are labeled with False Rejects, and to mark Note submodule (115) sends the instruction of mark；

Mark submodule (115), for the image or the correct image accepted that are labeled with False Rejects are carried out quality annotation, and Set corresponding quality tab.