CN100561502C - A kind of method and apparatus of face authentication - Google Patents

Abstract

The invention discloses a kind of method of face authentication, this method comprises: utilize the Fei Sheer fisher eigenface method based on pivot analysis, calculate the Fei Sheer fisher coefficient of target to be certified; The similar matching degree that the phenotype Weak Classifier calculates described fisher coefficient correspondence is searched in utilization; More described similar matching degree and pre-set threshold draw authentication result.The present invention also discloses a kind of device of face authentication simultaneously, and this device comprises: coefficients calculation block, calculate the Fei Sheer fisher coefficient of target to be certified; The output computing module utilizes and searches the similar matching degree that the phenotype Weak Classifier calculates described fisher coefficient correspondence; Authentication module, more described similar matching degree and pre-set threshold draw authentication result.The method and apparatus of this face authentication of the present invention, by introduce than single threshold represent ability stronger search the phenotype Weak Classifier, can better classify to training sample, thereby improve the discrimination of face authentication, reduced false detection probability.

Description

A kind of method and apparatus of face authentication

Technical field

The present invention relates to mode identification technology, be specifically related to a kind of method and apparatus of face authentication.

Background technology

Along with computer technology rapid development, computer process ability is greatly improved, and many emerging technological means such as pattern-recognition, motion detection occurred.Wherein, people's face detects a kind of as mode identification technology, is one of convenient mode of man-machine interaction in the computer vision system.

People's face detects, be meant in image or image sequence information such as the position of determining everyone face, size, people's face detects the prerequisite of being not only technology such as face authentication, Expression Recognition, gesture recognition and people's face synthesize, and the detection of people's face originally has important practical value in each field such as intelligent human-machine interaction, intelligent video monitoring, video conference and picture and video frequency searching etc.Face authentication, the technology such as the detection of employing people face, face feature point location that are meant are obtained people's face positional information in the image, and serve as that the information that facial image was comprised is extracted on the basis with people's face positional information, and further adopt information that facial image comprises determine object whether by a kind of method of statement identity.

Face authentication at first needs to utilize people's to be certified face image that sorter is trained, and sets up this people's sorter model.

When someone carries out face authentication, usually state the identity information of oneself earlier, Verification System finds the sorter model of this information correspondence from validation database according to the information of statement, adopt this sorter model that this people is carried out face authentication, if the authentication result identity information of people's statement is therewith coincide, then the expression authentication is passed through, otherwise the expression authentication is not passed through.Described people to be certified, target hereinafter referred to as to be certified.

For each target to be certified of carrying out face authentication, all need to be this target P to be certified _nTrain a model, and for this target to be certified, all training samples can be divided into two classes: the first kind is this P _nCorresponding facial image, second class is that all are not this P _nThe facial image of people's correspondence.

Usually the method for the face authentication that adopts be based on pivot analysis (Principal ComponentsAnalysis, face identification method PCA), PCA are a kind of linear data dimensionality reduction and feature extracting methods, and this method is divided into to be trained and use two stages:

At first training obtains the PCA space: the width of the training image (being generally rectangular image) during the supposition training is l, highly is h, with this two dimension input picture according to each pixel from left to right, from top to bottom sequential deployment is a dimensional vector, obtain

$X=\left[\begin{array}{c}{x}_1\\ x_2\\ .\\ .\\ .\\ x_{M-1}\\ x_M\end{array}\right]$

Make M=l*h, then the length of this one dimension row image vector is M.The number of supposing input picture is N, and then N image column vector is respectively X ₁, X ₂... .X _N, the mean vector of N vector is $u=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{X}_n,$ Deduct mean vector respectively with described N image column vector, form the capable vector Z of an one dimension=[X ₁-u, X ₂-u ... .X _N-u], then the covariance matrix of image vector is S _t=ZZ ^TObviously, S _tBe symmetric matrix, carry out obtaining after the characteristic value decomposition: S _t=W ∧ W ^T, make Y=W ^TX, then YY ^T=∧ promptly passes through linear transformation W ^TAfter, eliminated between data redundant.

Suppose among the above-mentioned eigenvectors matrix W that obtains, total k of nonzero eigenvalue characteristic of correspondence vector, in order further to reduce the number of the proper vector that adopts in the algorithm, to reduce operand, preceding t proper vector can also further selecting the eigenwert maximum is as the final face characteristic vector (claiming eigenface again) that uses.

Then, face authentication is carried out in the PCA space of using training to obtain: a dimensional vector X who obtains the input picture correspondence, described column vector X to above-mentioned t the space projection that proper vector is formed that is obtained by training, is obtained a projection vector of being made up of t coefficient.Discerning as feature with this t dimension projection vector, is described face identification method based on PCA.

Wherein, for face authentication, described input picture is for extracting the facial image that human face region and scaling obtain after the fixed size according to face feature point positions such as eyes, faces.

Though the face authentication method based on PCA is a recognition of face field comparative maturity, the also more stable method of while performance, this method discrimination is not high, and flase drop takes place easily.Therefore, it is a lot of to carry out improved algorithm on the basis of PCA, such as the method for core pivot element analysis or the method for binary pivot analysis, but, most widely used PCA and Fei Sheer (Fisher) linear discriminant analysis (Linear Discriminant Analysis, face authentication method LDA) of being based on.

The following linear discriminant analysis method of introducing earlier based on PCA.

So-called Fisher linear discriminant analysis method is to seek a linear transformation sample is carried out conversion, makes the easier classification of sample characteristics after the conversion for example to have d dimension sample X ₁, X ₂... X _n, N wherein ₁The individual ω that belongs to ₁Class, N ₂The individual ω that belongs to ₂Class.Respectively each sample is done projection on direction W, obtain scalar y _n=W ^TX _n, N d dimension sample can obtain N scalar like this, then set { y ₁, y ₂... y _nCan be divided into two subclass Y ₁And Y ₂

The target of Fisher method is in order to seek classification direction W ^*Make Y ₁, Y ₂Classified best, for this reason: in the definition d dimension X space,

All kinds of sample averages $m_i=\frac{1}{N_i}\underset{X_j\∈{\mathrm{\ω}}_i}{\mathrm{\Σ}}{X}_j,i=\mathrm{1,2};$

Sample within class scatter matrix S _iFor: $S_i=\underset{X_i\∈{\mathrm{\ω}}_i}{\mathrm{\Σ}}(X_j-m_i)*{(X_j-m_i)}^T,i=\mathrm{1,2};$

Total within class scatter matrix S _wBe S _W=S ₁+ S ₂

Dispersion matrix S between the sample class _bBe S _b=(m ₁-m ₂) * (m ₁-m ₂) ^T

To one dimension Y space, the definition sample average ${m_i}^{\′}=\frac{1}{N_i}\underset{y\∈Y_i}{\mathrm{\Σ}}y,i=\mathrm{1,2};$

Dispersion in the sample class ${S_i}^{\′}=\underset{y\∈Y_i}{\mathrm{\Σ}}{(y-{m_i}^{\′})}^2,i=\mathrm{1,2};$

Dispersion S in total class _w'=S ₁'+S ₂'.

Further define the Fisher criterion function $J_F\left(w\right)=\frac{{({m_1}^{\′}-{m_2}^{\′})}^2}{{S_W}^{\′}};$

In order to make the data after the projection be more prone to classification, therefore require described Fisher criterion function J _F(w) get minimum value, thereby need the interior dispersion S of total class of the one-dimensional data after the projection _w' try one's best little and between class dispersion as far as possible big.

The Fisher criterion function can be deformed into $J_F\left(w\right)=\frac{W^T{S}_{b}W}{W^T{S}_{w}W},$ W then ^*=arg max (J _F(w)).Because J _F(w) be Generalized Rayleigh Quotient, can adopt Lagrange (Lagrange) multiplier method to find the solution.Obtain at last $w^{*}=\frac{R}{\mathrm{\λ}}*S_w^{-1}*(m_1-m_2),$ R=(m wherein ₁-m ₂) ^T* w ^*Be scalar, then w ^*Direction be S _w ^-1* (m ₁-m ₂).

Described face authentication method, flow process based on PCA and Fisher LDA as shown in Figure 1, comprising:

Step 101: adopt the training sample training to obtain the PCA space, calculate the proper vector in PCA space, get preceding t maximum eigenwert characteristic of correspondence vector constitutive characteristic face space.

Step 102:, calculate classification direction W according to described sample X according to the sample X in the eigenface SPATIAL CALCULATION fisher method ^*

According to described fisher method, need to calculate W ^*, so that to Y ₁, Y ₂Best classification is carried out in two set.Therefore in face authentication, for an image to be certified, the result of determination that need obtain is that the people's face in this image is the someone, to the result's of the yes or no that produces behind every image authentication to be certified set, is exactly described Y ₁, Y ₂Set.In order to reach the effect of best face authentication, just must be to Y ₁Set, Y ₂Correct classification is carried out in set, to avoid producing flase drop.

Therefore, the described method of step 102 is specially:

The facial image that training is used projects on the eigenface space, the projection vector that to obtain a length be t, the projection vector that with described length is t by this sample X, calculates classification direction S according to the formula that defines in the fisher method as the sample X in the fisher eigenface method _w ^-1* (m ₁-m ₂).

Because the facial image that training is used all is to have carried out the input picture of demarcating, that is, for the people of a known identities, whether the facial image that uses during training is that people's face of the people of this known identities is known.Therefore, the classification direction W that finally calculates according to the good input picture of described demarcation ^*Have suitable differentiation, can distinguish Y more accurately ₁Set and Y ₂Set, the accuracy when having improved face authentication.

Step 103: obtain facial image to be certified, it is projected to the eigenface space obtain the projection vector that length is t, and then this projection vector is projected to the classification direction W that step 102 calculates ^*On, obtain a scalar (being called the Fisher coefficient).

Step 104: the magnitude relationship according to described fisher coefficient and the threshold value T that sets in advance draws authentication result.

A kind of method of described definite threshold value T is: the two good class facial images of demarcation that use during for training, obtain their projections respectively in the classification direction, and write down the Fisher coefficient that obtains after all projections; Suppose that the Fisher coefficient after the two class facial image projections meets Gaussian distribution, then data after the projection are carried out Gauss curve fitting after, select position that two class posterior probability equate as threshold value T.

The authentication method of PCA+Fisher shown in Figure 1 is than simple good based on the method effect of PCA, but, in the process of described definite threshold value T, need to suppose that the Fisher coefficient after the two class facial image projections meets Gaussian distribution, this can not be met in actual applications, has therefore influenced the scope of application of this method; In addition, the quantity of information that single threshold value T can represent only relies on threshold value T often can not exactly two class facial images be made a distinction very little, has therefore influenced the accuracy of this authentication method, occurs flase drop easily.

Summary of the invention

The embodiment of the invention provides a kind of method and apparatus of face authentication, and better face authentication effect can be provided.

For achieving the above object, technical scheme of the present invention specifically is achieved in that

A kind of method of face authentication, this method comprises:

Utilization is calculated the fisher coefficient of target to be certified based on the face authentication method of pivot analysis and Fei Sheer fisher linear discriminant analysis;

Calculating is used to train the training sample of the target to be certified of described sorter to concentrate, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described interval is divided into the sub-range of predetermined number, and determines the sub-range at the fisher coefficient place of described target to be certified;

Calculate the histogram of the training sample distribution of target to be certified in determined this sub-range;

According to described histogram, calculate the similar matching degree of searching the phenotype Weak Classifier;

More described similar matching degree and pre-set threshold draw authentication result.

The respectively corresponding sequence number in described each sub-range;

The method in the sub-range at the fisher coefficient place of described definite described target to be certified is:

Determine the sub-range corresponding sequence number at the fisher coefficient place of described target to be certified according to the following equation:

Utilize $j=\min (\mathrm{HNUM}-1,\mathrm{FLOOR}\left(\frac{(F_{\mathrm{coeff}}-F_{\min })*\mathrm{HNUM}}{F_{\max }-F_{\min }}\right)),$ Wherein min () is for getting minimum operation, and FLOOR () is downward rounding operation, and j is the sub-range corresponding sequence number at the fisher coefficient place of described target to be certified, F _CoeffBe the fisher coefficient of described target to be certified, HNUM is the quantity in described sub-range.

The histogrammic method that the training sample of target to be certified distributes in determined this sub-range of described calculating is:

Add up in the sub-range at fisher coefficient place of described target to be certified, the number of the fisher coefficient of two classes target training sample to be certified, obtain the histogram that the fisher coefficient of described two class training samples distributes in this sub-range, described two class training samples comprise: first kind training sample corresponding with target to be certified and the second corresponding class training sample of all non-described targets.

Described according to described histogram, the method that calculates the similar matching degree of searching the phenotype Weak Classifier is:

Utilize $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)}{{\mathrm{Hist}}_2\left(j\right)},$ Or $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}},$ Or $h\left(j\right)=\arctan \left(\ln \left(\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}}\right)\right)$ The similar matching degree of phenotype Weak Classifier is searched in calculating, and wherein σ is the positive number greater than 0, and ln is the natural logarithm computing, and arctan is an arctangent cp cp operation, Hist ₁(j) the expression sequence number is the histogram of first kind training sample in the sub-range of j, Hist ₂(j) the expression sequence number is the histogram of the second class training sample in the sub-range of j, and h (j) is the described similar matching degree that calculates.

The method of setting described threshold value is: the similar matching degree of the sorter of the fisher coefficient correspondence in the calculation training sample set, the similar matching degree of the sorter of the fisher coefficient correspondence of concentrating according to described training sample is determined interval [TF _Min, TF _Max], described TF _MinBe the minimum value of the similar matching degree of the sorter of fisher coefficient correspondence in the described training sample, TF _MaxMaximal value for the similar matching degree of the sorter of fisher coefficient correspondence;

From interval [TF _Min, TF _Max] in select the point of predetermined number, calculate the classification error rate of described each point respectively; Described classification error rate is to be included into the number of training purpose summation of error category;

The value of setting the point that training sample classification error rate is minimum in the described each point is described threshold value.

Described target to be certified comprises target human face region and target eyes zone;

The method that described similar matching degree and pre-set threshold draw authentication result is:

The similar matching degree of the fisher coefficient correspondence of target human face region is made as h _f, the similar matching degree of the fisher coefficient correspondence in target eyes zone is made as h _e, set h _fAnd h _eEach self-corresponding weights calculates the result after their weighted, the result after the described weighted and pre-set threshold is compared draw authentication result.

A kind of device of face authentication, this device comprises:

Coefficients calculation block, output computing module and authentication module; And described output computing module comprises: the interval is provided with unit, sub-range determining unit, histogram calculation unit and sorter output computing unit;

Described coefficients calculation block is utilized the face authentication method based on pivot analysis and Fei Sheer fisher linear discriminant analysis, calculates the fisher coefficient of target to be certified;

Described interval dispensing unit calculates the training sample of the target to be certified be used to train described sorter and concentrates, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described sub-range determining unit is divided into the sub-range of predetermined number with the described interval of determining, and determines the sub-range at the fisher coefficient place of described target to be certified;

Described histogram calculation unit calculates the histogram of the training sample distribution of target to be certified in determined this sub-range;

Described sorter output computing unit draws the similar matching degree of searching the phenotype Weak Classifier according to described histogram calculation;

Described authentication module, more described similar matching degree and pre-set threshold draw authentication result.

Described authentication module comprises: threshold value is provided with the unit;

Described threshold value is provided with the unit, the similar matching degree of the sorter of the fisher coefficient correspondence in the calculation training sample set, and the similar matching degree of the sorter of the fisher coefficient correspondence of concentrating according to described training sample is determined interval [TF _Min, TF _Max], described TF _MinBe the minimum value of the similar matching degree of the sorter of fisher coefficient correspondence in the described training sample, TF _MaxMaximal value for the similar matching degree of the sorter of fisher coefficient correspondence; From interval [TF _Min, TF _Max] in select the point of predetermined number, calculate the classification error rate of described each point respectively; Described classification error rate is to be included into the number of training purpose summation of error category;

The value of setting the point that training sample classification error rate is minimum in the described each point is described threshold value.

Described coefficients calculation block is calculated the Fei Sheer fisher coefficient of target to be certified, and described target to be certified comprises target human face region and target eyes zone;

Described output computing module utilizes and to search the phenotype Weak Classifier and calculate the similar matching degree of fisher coefficient correspondence of described target human face region and the similar matching degree of the fisher coefficient correspondence in described target eyes zone respectively.

Further comprise authentication ' unit in the described authentication module;

Described authentication ' unit, preestablish each self-corresponding weights of similar matching degree of the fisher coefficient correspondence in the similar matching degree of fisher coefficient correspondence of target human face region and target eyes zone, calculate the result after their weighted, the result after the described weighted and described threshold value are compared draw authentication result.

As seen from the above technical solutions, the method and apparatus of this face authentication of the embodiment of the invention, by introduce than single threshold represent ability stronger search the phenotype Weak Classifier, can better classify to training sample, thereby improved the discrimination of face authentication, reduced false detection probability.

Description of drawings

Fig. 1 is the synoptic diagram of prior art PCA+Fisher eigenface method.

Fig. 2 is the schematic flow sheet of the method for face authentication in the embodiment of the invention.

Fig. 3 is the composition structural representation of the device of face authentication in the embodiment of the invention.

Embodiment

For making purpose of the present invention, technical scheme and advantage clearer, below with reference to the accompanying drawing embodiment that develops simultaneously, the present invention is described in more detail.

On the basis of method shown in Figure 1, the embodiment of the invention proposes a kind of method of face authentication, and this method is utilized PCA+Fisher and in conjunction with searching the phenotype Weak Classifier, flow process as shown in Figure 2, comprising:

Step 201: utilize face authentication method, calculate the fisher coefficient of target to be certified based on pivot analysis and Fei Sheer fisher linear discriminant analysis;

The embodiment of the invention authenticates the target human face region, but because in the target human face region, the eyes zone is the zone that comprises the quantity of information maximum, therefore in order further to improve discrimination, the embodiment of the invention also further authenticates target eyes zone.Therefore, described target to be certified both can be the target human face region image that extracts from input picture, also can further comprise the target eyes area image that extracts from described input picture.

The method of the fisher coefficient of described calculating target to be certified is: the PCA coefficient of target to be certified is projected to the eigenface space, obtain again this projection vector being projected to fisher classification direction behind the projection vector, obtain the fisher coefficient of described target to be certified, concrete grammar is same as the prior art, repeats no more herein.

Step 202: utilize and search the similar matching degree that the phenotype Weak Classifier calculates the fisher coefficient correspondence of described target to be certified;

Describedly search the sorter that the phenotype Weak Classifier is to use the training sample training in advance to obtain, when carrying out face authentication, when comprising target human face region and target eyes zone in the described target to be certified simultaneously, naturally need target human face region training sample and two groups of training samples of target eyes regional training sample during training in advance, described target eyes regional training sample both can obtain separately in advance, also can from described target human face region training sample, extract, training obtains two and searches the phenotype Weak Classifier respectively, though described two sorters structure is also inequality, the authentication method of their training method when carrying out face authentication is identical.

It is to be noted simultaneously, though described target human face region is identical with the disposal route in target eyes zone, but when actual utilization searching the phenotype Weak Classifier and handle separately, these two processes are separate, do not have any relation between mutually.

Step 203: more described similar matching degree and pre-set threshold draw authentication result.

Should select to make the minimum threshold value of training sample classification error rate when determining threshold value S, the classification error rate of training sample be defined as all under this threshold value by the number of samples sum of misclassification, promptly be divided into first kind sample the second class sample number be divided into the number sum of the first kind sample of the second class sample.Because target to be certified may comprise target human face region and target eyes zone simultaneously, therefore the classification error rate of described training sample is in two groups of training samples, by the summation of the wrong number of samples of sorting out.

The described method of obtaining described threshold value is: the similar matching degree of the sorter of the fisher coefficient correspondence in the calculation training sample set, the similar matching degree of the sorter of the fisher coefficient correspondence of concentrating according to described training sample is determined interval [TF _Min, TF _Max], described TF _MinBe the minimum value of the similar matching degree of the sorter of fisher coefficient correspondence in the described training sample, TF _MaxMaximal value for the similar matching degree of the sorter of fisher coefficient correspondence;

Since comprise infinite a plurality of point in the interval, in order to reduce calculated amount, can be from interval [TF _Min, TF _Max] in select a certain number of point, calculate the classification error rate of described each point respectively, the value of setting the point that training sample classification error rate is minimum in the described each point is described threshold value.

For example, a kind of feasible scheme of obtaining threshold value is as follows:

With interval [TF _Min, TF _Max] be divided into FN sub-range, then total FN+1 end points, the value of each end points is respectively ${\mathrm{TF}}_{\min }+\frac{({\mathrm{TF}}_{\max }-{\mathrm{TF}}_{\min })*\mathrm{fn}}{\mathrm{FN}},$ Fn=0,1...FN, wherein fn is the end points sequence number;

Calculate the classification error rate of described FN+1 end points correspondence respectively, and the endpoint value of record sort error rate minimum, this endpoint value is made as threshold value.

Understand easily, this method for for example, is not the establishing method that is used to limit described threshold value only.

Correspondingly, the described method that draws authentication result according to described similar matching degree and pre-set threshold is: the similar matching degree of establishing the fisher coefficient correspondence of target human face region is h _f, the similar matching degree of the fisher coefficient correspondence in target eyes zone is h _e, set h _fAnd h _eEach self-corresponding weights calculates the result after their weighted, the result after the described weighted and preset threshold is compared draw authentication result.

Understand easily, when target to be certified only comprises the target human face region, the similar matching degree h of the fisher coefficient in target eyes zone _eTherefore be 0, to the did not influence as a result of last weighted.

Following mask body place of matchmakers states in step 202～203, and the method that the phenotype Weak Classifier calculates the similar matching degree of described coefficient correspondence and draws authentication result is searched in a kind of possible utilization:

One, calculating is used to train the target training sample to be certified of described sorter to concentrate, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described interval is divided into the sub-range of predetermined number, for example described interval is divided into H _NUMIndividual sub-range.Described H _NUMValue rule of thumb select usually, the embodiment of the invention is not done concrete qualification, according to interval [F _Min, F _Max] the length value selecting to be of moderate size get final product, for example can get H usually _NUM=32,48 or 64 etc.

When target to be certified comprised target human face region and target eyes zone, the independent respectively fisher coefficient that calculates separately correspondingly obtain two class intervals, and the H that the sub-range is divided was carried out in two class intervals _NUMValue also separate.Also be same situation in each follow-up step, each step all is applicable to the treatment scheme in described target human face region and target eyes zone, therefore hereinafter no longer statement.

Two, determine the sub-range at the fisher coefficient place of described target to be certified;

The method in the sub-range at the fisher coefficient place of described definite described target to be certified is: the fisher coefficient corresponding sequence number j in described interval that calculates target to be certified.

Interval [F _Min, F _Max] be divided into H _NUMIndividual sub-range, each section corresponding sequence number are j (0≤j≤H _NUM-1, j is an integer), the fisher coefficient corresponding sequence number j that calculates target to be certified can adopt following formula:

$j=\min (\mathrm{HNUM}-1,\mathrm{FLOOR}\left(\frac{(F_{\mathrm{coeff}}-F_{\min })*\mathrm{HNUM}}{F_{\max }-F_{\min }}\right)),$ Wherein min () is for getting minimum operation, and FLOOR () is downward rounding operation, as FLOOR (1.9)=1, and FLOOR (2.2)=2.Need to prove, calculate described fisher coefficient corresponding sequence number j and can also adopt alternate manner, do not enumerate one by one that therefore, the formula of aforementioned calculation sequence number j for for example, is not to be used for limiting only herein.

Three, calculate on determined this sub-range the histogram Hist that two class training samples of described target to be certified distribute ₁(j) and Hist ₂(j).

The method of described compute histograms is: in all first kind training samples, the number of samples that the corresponding sequence number of fisher coefficient is j equals Hist ₁(j); And similarly, in the second all class training samples, the number of samples that the corresponding sequence number of fisher coefficient is j equals Hist ₂(j).Because described interval is divided into H _NUMTherefore section understands Hist easily ₁And Hist ₂Be H _NUMDimensional vector.

Four,, calculate and search the phenotype Weak Classifier according to described histogram $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)}{{\mathrm{Hist}}_2\left(j\right)}.$

As shown in the above description, in the step 202, described utilization is searched the method that the phenotype Weak Classifier calculates the similar matching degree of described coefficient correspondence and is:

Step 202a (not shown): maximum, minimum value according to target training sample fisher coefficient to be certified obtain [F _Min, F _Max] interval, according to the H that sets _NUM, calculate described coefficient corresponding sequence number j in described sub-range;

Step 202b (not shown): calculate in the sub-range of sequence number j the histogram that described two class training samples distribute.

Need to prove that when the practical application of people's face checking, the Fisher coefficient of some image to be verified may exceed interval [F _Min, F _Max] scope, when described Fisher coefficient exceeds [F _Min, F _Max] scope the time, think that promptly described image to be verified is not this people.

Further, the histogram Hist on the interval sequence number j ₂(j) be zero, a kind of preferable definition mode of searching the phenotype Weak Classifier can further be set: $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}},$ Wherein σ be greater than zero on the occasion of, as get σ=0.006, but the present invention do not limit, equally also can get other values.

Perhaps, in order to obtain the expression mode that DATA DISTRIBUTION is more optimized, can also define to search the phenotype Weak Classifier be $h\left(j\right)=\arctan \left(\ln \left(\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}}\right)\right),$ This moment h (j) span be (0, π).Wherein ln is the natural logarithm computing, and arctan is an arctangent cp cp operation.

Preferably, the histogram Hist that can further obtain to statistics ₁And Hist ₂Adopt wave filter to carry out smoothing processing, for example adopt mean filter that described histogram is handled.

Equally, can also adopt mean filter to carry out smothing filtering to searching the phenotype Weak Classifier.

As seen, the method for this face authentication of the embodiment of the invention combines with PCA+Fisher by searching the phenotype Weak Classifier, makes the present invention go for the input picture that various sample probabilities distribute; Simultaneously, introducing than single threshold represent ability stronger search the phenotype Weak Classifier, can improve the discrimination of face authentication, reduce flase drop, in addition, when target to be certified comprised target human face region and target eyes zone simultaneously, preferred embodiment of the present invention can also further improve the discrimination of face authentication.

The present invention also provides a kind of device of face authentication, as shown in Figure 3, comprising: coefficients calculation block 310, output computing module 320 and authentication module 330;

Described coefficients calculation block 310 is calculated the Fei Sheer fisher coefficient of target to be certified;

Described output computing module 320 utilizes and searches the similar matching degree that the phenotype Weak Classifier calculates described fisher coefficient correspondence;

Described authentication module 330, more described similar matching degree and pre-set threshold draw authentication result.

Described output computing module 320 comprises: the interval is provided with unit 321, sub-range determining unit 322, histogram calculation unit 323 and sorter output computing unit 324;

Described interval is provided with unit 321, calculates the training sample of the target to be certified be used to train described sorter and concentrates, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described sub-range determining unit 322 is divided into the sub-range of predetermined number with the described interval of determining, and determines the sub-range corresponding sequence number at the fisher coefficient place of described target to be certified according to the following equation:

Utilize $j=\min (\mathrm{HNUM}-1,\mathrm{FLOOR}\left(\frac{(F_{\mathrm{coeff}}-F_{\min })*\mathrm{HNUM}}{F_{\max }-F_{\min }}\right)),$ Wherein min () is for getting minimum operation, and FLOOR () is downward rounding operation, and j is described sequence number, and the respectively corresponding sequence number in described each sub-range;

Described histogram calculation unit 323, fisher coefficient according to each training sample, add up in the sub-range at fisher coefficient place of described target to be certified, the number of the fisher coefficient of two classes target training sample to be certified, obtain the histogram that the fisher coefficient of described two class training samples distributes in this sub-range, described two class training samples comprise: first kind training sample corresponding with target to be certified and the second corresponding class training sample of all non-described targets;

Described sorter output computing unit 324 utilizes $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)}{{\mathrm{Hist}}_2\left(j\right)},$ Or $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}},$ Or $h\left(j\right)=\arctan \left(\ln \left(\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}}\right)\right)$ The similar matching degree of phenotype Weak Classifier is searched in calculating, and wherein σ is the positive number greater than 0, and ln is the natural logarithm computing, and arctan is an arctangent cp cp operation, Hist ₁(j) the expression sequence number is the histogram of first kind training sample in the sub-range of j, Hist ₂(j) the expression sequence number is the histogram of the second class training sample in the sub-range of j.

When described target to be certified comprises target human face region and target eyes zone simultaneously, at this moment:

Described coefficients calculation block 310 is used to calculate the Fei Sheer fisher coefficient of target to be certified, and described target to be certified comprises target human face region and target eyes zone;

Correspondingly, described output computing module 320 utilized and to search the phenotype Weak Classifier and calculate the similar matching degree of fisher coefficient correspondence of described target human face region and the similar matching degree of the fisher coefficient correspondence in described target eyes zone respectively this moment.

Described authentication module 330 comprises: threshold value is provided with unit 331 and authentication determination unit 332;

Described threshold value is provided with unit 331, makes the minimum value of target training sample classification error rate to be certified be set to described threshold value, and described training sample classification error rate is to be included into the number of training purpose summation of error category.

Described authentication determination unit 332, preestablish each self-corresponding weights of similar matching degree of the fisher coefficient correspondence in the similar matching degree of fisher coefficient correspondence of target human face region and target eyes zone, calculate the result after their weighted, the result after the described weighted and described threshold value are compared draw authentication result.

As seen, the device of this face authentication of the embodiment of the invention combines with PCA+Fisher by searching the phenotype Weak Classifier, makes the present invention go for the input picture that various sample probabilities distribute; Simultaneously, introducing than single threshold represent ability stronger search the phenotype Weak Classifier, can improve the discrimination of face authentication, reduce flase drop, in addition, when target to be certified comprised target human face region and target eyes zone simultaneously, preferred embodiment of the present invention can also further improve the discrimination of face authentication.

Simultaneously; understand easily, the above is preferred embodiment of the present invention only, is not to be used to limit spirit of the present invention and protection domain; equivalent variations that any those of ordinary skill in the art made or replacement all should be considered as being encompassed within protection scope of the present invention.

Claims (10)

1, a kind of method of face authentication is characterized in that, this method comprises:

Utilization is calculated the fisher coefficient of target to be certified based on the face authentication method of pivot analysis and Fei Sheer fisher linear discriminant analysis;

Calculating is used to train the training sample of the target to be certified of described sorter to concentrate, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described interval is divided into the sub-range of predetermined number, and determines the sub-range at the fisher coefficient place of described target to be certified;

Calculate the histogram of the training sample distribution of target to be certified in determined this sub-range;

According to described histogram, calculate the similar matching degree of searching the phenotype Weak Classifier;

More described similar matching degree and pre-set threshold draw authentication result.

2, method according to claim 1 is characterized in that, the respectively corresponding sequence number in described each sub-range;

The method in the sub-range at the fisher coefficient place of described definite described target to be certified is:

Determine the sub-range corresponding sequence number at the fisher coefficient place of described target to be certified according to the following equation:

Utilize $j=\min (\mathrm{HNUM}-1,\mathrm{FLOOR}\left(\frac{(F_{\mathrm{coeff}}-F_{\min })*\mathrm{HNUM}}{F_{\max }-F_{\min }}\right)),$ Wherein min () is for getting minimum operation, and FLOOR () is downward rounding operation, and j is the sub-range corresponding sequence number at the fisher coefficient place of described target to be certified, F _CoeffBe the fisher coefficient of described target to be certified, HNUM is the quantity in described sub-range.

3, method according to claim 2 is characterized in that, the histogrammic method that the training sample of target to be certified distributes in determined this sub-range of described calculating is:

Add up in the sub-range at fisher coefficient place of described target to be certified, the number of the fisher coefficient of two classes target training sample to be certified, obtain the histogram that the fisher coefficient of described two class training samples distributes in this sub-range, described two class training samples comprise: first kind training sample corresponding with target to be certified and the second corresponding class training sample of all non-described targets.

4, method according to claim 3 is characterized in that, and is described according to described histogram, and the method that calculates the similar matching degree of searching the phenotype Weak Classifier is:

Utilize $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)}{{\mathrm{Hist}}_2\left(j\right)},$ Or $h\left(j\right)=\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}},$ Or $h\left(j\right)=\arctan \left(\ln \left(\frac{{\mathrm{Hist}}_1\left(j\right)+\mathrm{\σ}}{{\mathrm{Hist}}_2\left(j\right)+\mathrm{\σ}}\right)\right)$ The similar matching degree of phenotype Weak Classifier is searched in calculating, and wherein σ is the positive number greater than 0, and 1n is the natural logarithm computing, and arctan is an arctangent cp cp operation, Hist ₁(j) the expression sequence number is the histogram of first kind training sample in the sub-range of j, Hist ₂(j) the expression sequence number is the histogram of the second class training sample in the sub-range of j, and h (j) is the described similar matching degree that calculates.

5, method according to claim 4 is characterized in that, the method for setting described threshold value is:

The similar matching degree of the sorter of the fisher coefficient correspondence in the calculation training sample set, the similar matching degree of the sorter of the fisher coefficient correspondence of concentrating according to described training sample is determined interval [TF _Min, TF _Max], described TF _MinBe the minimum value of the similar matching degree of the sorter of fisher coefficient correspondence in the described training sample, TF _MaxMaximal value for the similar matching degree of the sorter of fisher coefficient correspondence;

From interval [TF _Min, TF _Max] in select the point of predetermined number, calculate the classification error rate of described each point respectively; Described classification error rate is to be included into the number of training purpose summation of error category;

The value of setting the point that training sample classification error rate is minimum in the described each point is described threshold value.

According to each described method in the claim 1 to 5, it is characterized in that 6, described target to be certified comprises target human face region and target eyes zone;

The method that described similar matching degree and pre-set threshold draw authentication result is:

The similar matching degree of the fisher coefficient correspondence of target human face region is made as h _f, the similar matching degree of the fisher coefficient correspondence in target eyes zone is made as h _e, set h _fAnd h _eEach self-corresponding weights calculates the result after their weighted, the result after the described weighted and pre-set threshold is compared draw authentication result.

7, a kind of device of face authentication is characterized in that, this device comprises: coefficients calculation block, output computing module and authentication module; And described output computing module comprises: the interval is provided with unit, sub-range determining unit, histogram calculation unit and sorter output computing unit;

Described coefficients calculation block is utilized the face authentication method based on pivot analysis and Fei Sheer fisher linear discriminant analysis, calculates the fisher coefficient of target to be certified;

Described interval dispensing unit calculates the training sample of the target to be certified be used to train described sorter and concentrates, and the fisher coefficient of each training sample is determined interval [F according to the fisher coefficient that described training sample is concentrated _Min, F _Max], described F _MinBe the minimum value of fisher coefficient in the described training sample, F _MaxMaximal value for fisher coefficient in the described training sample;

Described sub-range determining unit is divided into the sub-range of predetermined number with the described interval of determining, and determines the sub-range at the fisher coefficient place of described target to be certified;

Described histogram calculation unit calculates the histogram of the training sample distribution of target to be certified in determined this sub-range;

Described sorter output computing unit draws the similar matching degree of searching the phenotype Weak Classifier according to described histogram calculation;

Described authentication module, more described similar matching degree and pre-set threshold draw authentication result.

8, device according to claim 7 is characterized in that, described authentication module comprises: threshold value is provided with the unit;

Described threshold value is provided with the unit, the similar matching degree of the sorter of the fisher coefficient correspondence in the calculation training sample set, and the similar matching degree of the sorter of the fisher coefficient correspondence of concentrating according to described training sample is determined interval [TF _Min, TF _Max], described TF _MinBe the minimum value of the similar matching degree of the sorter of fisher coefficient correspondence in the described training sample, TF _MaxMaximal value for the similar matching degree of the sorter of fisher coefficient correspondence;

From interval [TF _Min, TF _Max] in select the point of predetermined number, calculate the classification error rate of described each point respectively; Described classification error rate is to be included into the number of training purpose summation of error category;

The value of setting the point that training sample classification error rate is minimum in the described each point is described threshold value.

9, device according to claim 8 is characterized in that, described coefficients calculation block is calculated the Fei Sheer fisher coefficient of target to be certified, and described target to be certified comprises target human face region and target eyes zone;

Described output computing module utilizes and to search the phenotype Weak Classifier and calculate the similar matching degree of fisher coefficient correspondence of described target human face region and the similar matching degree of the fisher coefficient correspondence in described target eyes zone respectively.

10, device according to claim 9 is characterized in that, further comprises authentication ' unit in the described authentication module;

Described authentication ' unit, preestablish each self-corresponding weights of similar matching degree of the fisher coefficient correspondence in the similar matching degree of fisher coefficient correspondence of target human face region and target eyes zone, calculate the result after their weighted, the result after the described weighted and described threshold value are compared draw authentication result.

Images