CN104363981A - Face verification method and system - Google Patents

Abstract

The invention relates to a face verification method and system. The method comprises preprocessing high-dimensional face characteristic data by means of principal component analysis and linear discriminant analysis, arranging the data dimension after dimension reduction of the principal component analysis; establishing a discriminant high-order boltzmann machine and arranging the number of nodes in a hidden layer; reducing modeling parameters of the discriminant high-order boltzmann machine by means of a tensor diagonalization strategy; inputting coupled face data into the discriminant high-order boltzmann machine, maximizing conditional probability of relationship category by means of a stochastic gradient descent algorithm, optimizing the weight of the boltzmann machine iteratively, and obtaining the final discriminant high-order boltzmann machine; inputting the coupled face data to be verified into the discriminant high-order boltzmann machine model, and obtaining the corresponding verification result data. According to the utility model, data relation category information is introduced into the unsupervised boltzmann machine model, the discrimination capacity of the model is improved, and the face verification method is suitable for face verification with accuracy requirements.

Description

Face verification method and system

Technical field

The present invention relates to a kind of face verification method and system, particularly relate to a kind of face verification method and system based on discriminate high-order Boltzmann machine.

Background technology

It is the hot research problem of area of pattern recognition that face verification technology is applied widely due to it always, such as, be applied in customs and bank etc. as authentication.

Given a pair facial image, the target of face verification judges whether these two facial images represent same person.The face verification under constraint environment is mainly studied in work before, and the facial image namely in experiment is all gather under controlled environmental factor, comprises fixing illumination, attitude and expression.Such as, but in actual applications, these factors are all uncontrollable, and the facial image in video monitoring may comprise arbitrary illumination.So, nearest many researchers start then research unconstrained condition under face verification problem.Facial image under unconstrained condition factor has very large variance within clusters, deals with challenging.

Existing face verification problem comprises two steps: data characteristics represents measures with data relationship.Represent for data characteristics, people use some Feature Descriptors manually set to represent human face data usually, such as LBP, SIFT and Gabor etc.Obtaining after paired human face data represents, usually utilizing COS distance to measure the similarity size of often pair of image, and then can judging whether this represents same person to image.

Summary of the invention

The object of the invention is the defect for prior art, a kind of face verification method and system is provided, to realize the face verification being more suitable for having required precision.

For achieving the above object, the invention provides a kind of face verification method, described method comprises:

Step 1, principal component analysis and linear discriminant analysis is utilized to carry out pretreatment respectively, comprising the data dimension arranged after principal component analysis dimensionality reduction to higher-dimension face characteristic data;

Step 2, set up discriminate high-order Boltzmann machine, the nodes of hidden layer is set;

Step 3, utilize the diagonalizable strategy of tensor to reduce the model parameter of this discriminate high-order Boltzmann machine;

Step 4, paired human face data is input in discriminate high-order Boltzmann machine, utilize stochastic gradient descent algorithm to maximize other conditional probability of relation object, thus optimize the weight of this Boltzmann machine iteratively, thus obtain final discriminate high-order Boltzmann machine;

Step 5, to described discriminate high-order Boltzmann machine mode input paired human face data to be verified, obtain corresponding the result data.

Present invention also offers a kind of face verification system, described system comprises: network sets up module, network weight optimizes module and Data Verification module, wherein:

Described network sets up module, for setting up discriminate high-order Boltzmann machine, and arranges the nodes of network hidden layer;

Described network weight optimizes module, utilizes stochastic gradient descent algorithm to minimize the object function of this Boltzmann machine to obtain the Boltzmann machine weight after optimization, thus obtains final discriminate high-order Boltzmann machine model;

Described Data Verification module, for inputting paired human face data to be verified to described discriminate high-order Boltzmann machine, obtain classification layer two nodal value, the relative size comparing two numerical value can draw face verification result data.

Face verification method of the present invention and system, by introducing the other information of data relation object in without supervision Boltzmann machine model, making Model checking power strengthen, being more suitable for the face verification with required precision.

Accompanying drawing explanation

Fig. 1 is the flow chart of face verification method of the present invention;

Fig. 2 shows discriminate high-order Boltzmann machine structural representation used in the present invention;

Fig. 3 is the schematic diagram of tensor diagonalization;

Fig. 4 is the schematic diagram of face verification system of the present invention;

Fig. 5 be different input data dimensions and different implicit node quantity time experiment effect variation diagram.

Detailed description of the invention

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Invention face verification method and system mainly propose a more effective data relationship measure, can promote the precision of face verification largely.

Fig. 1 is the flow chart of face verification method of the present invention, and as shown in the figure, the present invention specifically comprises the steps:

Step 101, utilize principal component analysis and linear discriminant analysis to carry out pretreatment respectively to higher-dimension face characteristic data, wherein need the data dimension after principal component analysis dimensionality reduction is set;

Wherein, principal component analysis (Principal Component Analysis, PCA) is a kind of conventional Method of Data with Adding Windows, and this method needs to set the data dimension after dimensionality reduction, usually from hundreds of to several thousand not etc.Linear discriminant analysis (Linear Discriminant Analysis, LDA) is a kind of conventional sub-space learning method having supervision, and the identification of data can be made after process to strengthen.

Step 102, set up discriminate high-order Boltzmann machine, the nodes of hidden layer is set;

Wherein, described discriminate high-order Boltzmann machine is multitiered network structure, comprises two data input layers, a hidden layer and a classification output layer.Two input layers of described discriminate high-order Boltzmann machine are paired face training data, the character representation of such as facial image.In an embodiment of the present invention, require the size that all human face data keep identical, such as the vector of same length; Classification output layer is for being expressed as the result to face training data, and the result due to face verification only has two classes, and namely mate or do not mate, output layer only comprises two nodes here, respectively this two kind corresponding; This discriminate high-order Boltzmann machine has network weight, for obtaining next node layer value according to current layer nodal value.The described input layer of discriminate high-order Boltzmann machine and the nodes of output layer are fixing, but the nodes of its hidden layer needs manual adjustment to make the effect of this model optimum.

Fig. 2 shows the discriminate high-order Boltzmann machine structural representation used in one embodiment of the invention.As shown in Figure 2, this is the network of four layers, and the circle point in every layer represents network node.Far Left is two-layer represents two human face data input layers respectively in bottom, input layer input is is all the column vector of I dimension each dimension of vector represents with a node, and its value is all normalized to 0 average and variance is 1.Top is classification output layer, the classification of face verification is expressed as the vector of one 2 dimension the value of vector is [1,0] or [0,1], two kinds of situation Corresponding matchings or do not mate.Middle hidden layer is the vector of a K dimension every one dimension value of vector is 0 or 1.The value of h and z can be calculated by the vector value of their front layers, and wherein h and x and y layer are connected to each other:

$\begin{array}{c}{h}_k=\mathrm{\σ}(c_k+\underset{\mathrm{ij}}{\mathrm{\Σ}}{W}_{\mathrm{ijk}}{x}_i{y}_j+\underset{t}{\mathrm{\Σ}}{U}_{\mathrm{kt}}{z}_t)\\ z_t=\frac{e^{d_t+{\mathrm{\Σ}}_k{U}_{\mathrm{kt}}{h}_k}}{{\mathrm{\Σ}}_{t^{*}}{e}^{d_{t^{*}}+{\mathrm{\Σ}}_k{U}_{{\mathrm{kt}}^{*}}{h}_k}},\end{array}$

Wherein, σ (x)=1/ (1+e ^-x), x=(x _i) _{i ∈ I}, y=(y _j) _{j ∈ J}, h=(h _k) _{k ∈ K}, z=(z _t) _{t ∈ { 1,2}}.C=(c _k) _{k ∈ K}with d=(d _t) _{t ∈ { 1,2}}be respectively the offset parameter of hidden layer and classification layer.W=(W _ijk) _{i ∈ I, j ∈ J, k ∈ K}two connection weights between input layer and hidden layer, and U=(U _kt) _{k ∈ K, t ∈ { 1,2}}it is the connection weight between hidden layer and classification layer.

Step 103, utilize the diagonalizable strategy of tensor to reduce the model parameter of this discriminate high-order Boltzmann machine; Discriminate high-order Boltzmann machine contains a large amount of model parameter needs and learns, and when the data dimension inputted is hundreds of, three-dimensional weight tensor W just may containing weight parameter up to a million.Like this, when amount of training data is not very many, be easy to cause over-fitting.Therefore the diagonalizable method of tensor is adopted to reduce parameter here, as shown in Figure 3.On h axle, each h _kcorrespond to a weight matrix section:

W _..k＝(W _ijk) _i∈I,j∈J

Each weights W wherein in matrix _ijkall correspond to node x _iand y _jannexation.Implicit hypothesis be any one node in x all with y in so there is dependency relation in node.Such hypothesis is concerning improper human face data.Specifically, we utilize local description to extract the feature of face key point annex usually, such as, extract LBP feature at positions such as face eyes, nose and mouths, then the series connection of these features are obtained final feature.So each node only characterizes the regional area of face in x and y, as shown in the R1-R5 in Fig. 3.When carrying out face verification, people are usually it is of concern that whether the same position such as eyes in two width facial images are couplings, and whether the nose of the eyes and another one people of being indifferent to a people mates.So, only have the x that position is closed on _inode and y _jwhen the human face region that node represents is the same, just there is stronger correlation in them, and as the block diagonalization matrix in Fig. 3, wherein black representative has correlation, and white is not for having correlation.Therefore for any node x _i, the node that we define relative ε-neighborhood is { y _j, j ∈ [i-ε, i+ ε] }, then by each weight matrix section W _..kreduce to the weight only retained on its 2 ε+1 diagonal:

${\tilde{W}}_{..k}={\left(W_{\mathrm{ijk}}\right)}_{i\∈I,j\∈[i-\mathrm{\ϵ},i+\mathrm{\ϵ}]}$

Parameter ranges in whole like this weight tensor W is namely from O (n ³) be reduced to O (n ²).

Step 104, paired human face data is input in discriminate high-order Boltzmann machine, utilize stochastic gradient descent algorithm to maximize other conditional probability of relation object, thus optimize the weight of this Boltzmann machine iteratively, thus obtain final discriminate high-order Boltzmann machine;

Be described for the network in Fig. 2, two input layer data, middle hidden layer and classification output layer represent a discriminate high-order Boltzmann machine, and its energy function E (x, y, h, z) is defined as:

$\begin{array}{c}E(x,y,h,z)=-\underset{i,k,j\∈[i-\mathrm{\ϵ},i+\mathrm{\ϵ}]}{\mathrm{\Σ}}{W}_{\mathrm{ijk}}{x}_j{y}_j{h}_k-\underset{\mathrm{kt}}{\mathrm{\Σ}}{U}_{\mathrm{kt}}{h}_k{z}_t-\frac{1}{2}\underset{i}{\mathrm{\Σ}}{(x_i-a_i)}^2\\ -\frac{1}{2}\underset{j}{\mathrm{\Σ}}{(y_j-b_j)}^2-\underset{k}{\mathrm{\Σ}}{c}_k{h}_k-\underset{t}{\mathrm{\Σ}}{d}_t{z}_t\end{array}$

Wherein, a=(a _i) _{i ∈ I}with b=(b) _{j ∈ J}represent the bias term parameter of two input layers respectively.

The basis of energy function can obtain the probability distribution P (x, y, z) of input layer data x, y and its relation classification z:

$P(x,y,z)=\underset{h}{\mathrm{\Σ}}\frac{1}{{\mathrm{\Σ}}_{x,y,h,z}{e}^{-E(x,y,h,z)}}{e}^{-E(x,y,h,z)}$

In order to learning model parameter, common way utilizes stochastic gradient descent algorithm to minimize object function, namely minimizes the maximum likelihood function L of negative logarithm _gen:

$L_{\mathrm{gen}}=-\log \underset{\mathrm{\α}}{\mathrm{\Σ}}P(x^{\mathrm{\α}},y^{\mathrm{\α}},z^{\mathrm{\α}})$

Wherein α represents the index of training data.Object function L _gengradient about parameter W can calculate as follows:

$\frac{\&PartialD;L_{\mathrm{gen}}}{\&PartialD;W}=\underset{\mathrm{\&alpha;}}{\mathrm{\&Sigma;}}-<\frac{\&PartialD;E(x^{\mathrm{\&alpha;}},y^{\mathrm{\&alpha;}},h,z^{\mathrm{\&alpha;}})}{\&PartialD;W}{>}_h+<\frac{\&PartialD;E(x,y,h,z)}{\&PartialD;W}{>}_{x,y,h,z}$

Wherein <. > _qrepresent the expectation about variable q.But the Section 2 calculated in above formula is very difficult, because computational process needs all value states traveling through variable x, y, h and z.Use L _genas model learning process during object function as shown in Figure 6.In addition, in order to easy computational process, we have also been attempted the object function of a discriminate, other conditional probability of relation object namely under negative logarithm:

$L_{\mathrm{dis}}=-\log \underset{\mathrm{\&alpha;}}{\mathrm{\&Sigma;}}P\left(z^{\mathrm{\&alpha;}}\right|x^{\mathrm{\&alpha;}},y^{\mathrm{\&alpha;}})$

Condition distribution P (z|x, y) wherein about z can be expressed as:

$p(z_t=1|x,y)=\frac{e^{d_t}{\mathrm{\&Pi;}}_k(1+e^{c_k+{\mathrm{\&Sigma;}}_{\mathrm{ij}}{W}_{\mathrm{ijk}}{x}_i{y}_j+U_{\mathrm{kt}}})}{{\mathrm{\&Sigma;}}_{t^{*}}{e}^{d_{t^{*}}}{\mathrm{\&Pi;}}_k(1+e^{c_k+{\mathrm{\&Sigma;}}_{\mathrm{ij}}{W}_{\mathrm{ijk}}{x}_i{y}_j+U_{{\mathrm{kt}}^{\text{*}}}})}$

Be different from object function L _gen, object function L in this case _disgradient about parameter W can be calculated accurately:

$\frac{\&PartialD;L_{\mathrm{dis}}}{\&PartialD;W}=\underset{k}{\mathrm{\&Sigma;}}\mathrm{\&sigma;}\left(M_{\mathrm{tk}}\right)\frac{\&PartialD;M_{\mathrm{tk}}}{\&PartialD;\mathrm{\&theta;}}-\underset{k,t^{*}}{\mathrm{\&Sigma;}}\mathrm{\&sigma;}\left(M_{t^{*}k}\right)P\left(z_{t^{*}}\right|x,y)\frac{\&PartialD;M_{t^{*}k}}{\&PartialD;\mathrm{\&theta;}}$

$M_{\mathrm{tk}}=c_k+\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{W}_{\mathrm{ijk}}{x}_i{y}_j+U_{\mathrm{kt}}$

After obtaining gradient, we in an iterative manner parameter to adjusting:

$W\&LeftArrow;W-\mathrm{\&epsiv;}\frac{\&PartialD;L_{\mathrm{dis}}}{\&PartialD;W},$

Wherein, ε represents a constant learning rate.

Step 105, to the described discriminate high-order Boltzmann machine mode input trained paired human face data to be verified, such as x and y, obtains output layer nodal value z, can be judged the result of face verification by the relative size of wherein two numerical value.If z ₀> z ₁then do not mate, otherwise then mate.

Fig. 4 is the schematic diagram of face verification system of the present invention, and as shown in the figure, the present invention specifically comprises: network sets up module 1, network weight optimizes module 2 and Data Verification module 3, wherein:

Concrete, network sets up module 1 for setting up discriminate high-order Boltzmann machine, and arranges the nodes of network hidden layer; Discriminate high-order Boltzmann machine is multitiered network structure, comprises two inputs data Layer, a hidden layer and classification output layers.Input layer is be trained to right human face data, and output layer represents the result be trained to human face data; The discriminate high-order Boltzmann machine of discriminate has network weight, for obtaining next node layer value according to current layer nodal value.

Network weight is optimized module 2 and is utilized stochastic gradient descent algorithm to minimize the object function of this Boltzmann machine to obtain the Boltzmann machine weight after optimization, thus obtains final discriminate high-order Boltzmann machine model.

Data Verification module 3 is for inputting paired human face data to be verified to discriminate high-order Boltzmann machine, and obtain classification layer two nodal value, the relative size comparing two numerical value can draw face verification result.

In order to describe the specific embodiment of the present invention in detail, illustrate for two image data sets (LFW and WDRef) here.LFW data set comprises the facial image that 13233 come from 5749 people.Wherein 1680 people have 1 image, and the image having more than two of all the other people.WDRef data set comprises 99773 facial images of throwing the net on network, and the human face data wherein in WDRef and LFW does not have overlap.Face images is all directly collect from Internet, has very large change in illumination, posture etc.

We carry out face verification experiment under arranging at two kinds:

1) supervised learning under initial data, the data by LFW are divided into 10 parts, get wherein 9 points train, remain 1 part and test, so repeat 10 times.

2) supervised learning under external data.Utilize WDRef data to train, LFW database is tested.

Concrete steps are as follows:

Step S1, for all data, utilizes principal component analysis that original higher-dimension face LBP feature is dropped to 2000 dimensions, and recycling linear discriminant analysis processes.

Step S2, use the discriminate high-order Boltzmann machine model of four layers, two input layer, hidden layer and output layer comprise 2000,2000,1000 and 2 nodes respectively.

Step S3, utilizes the network weight of the diagonalizable tactful cancellation of tensor except diagonal, weight quantity is reduced to O (n ²).

Step S4, utilizes the object function L of gradient descent algorithm optimized network _dis, wherein gradient can carry out accurate Calculation and non-approximated.Optimization carries out in an iterative manner, arrange here maximum iteration time be 30 can ensure convergence.

Step S6, is input to the paired human face data of test in the model trained, and the result after output layer exports checking.The result exported is the vector of a bidimensional, and the relative size comparing two values can judge coupling or not mate.

We compared for the certain methods that some face verification precision are the highest now.Can find under two kinds of Setup Experiments by the contrast of ROC curve, our method all achieves best experimental result.In addition, when we test and use different input data dimensions and different implicit node quantity, experiment effect situation of change of the present invention.As shown in Figure 5, transverse axis represents the quantity (number of hidden units) of implicit node, the longitudinal axis represents recognition of face precision (accuracy), and in figure, each curve represents the face verification precision situation of change of a kind of data dimension when difference implies node.Input data dimension changes to 1000 dimensions from 200 dimensions, and implicit node quantity changes to 1200 from 200.As can be seen from the figure, the quantity of fixing implicit node, along with using the dimension of input data to increase, precision rises.For the dimension of arbitrary input data, when implicit node is greater than 400, namely precision keeps stable and no longer rises.

Professional should recognize further, in conjunction with unit and the algorithm steps of each example of embodiment disclosed herein description, can realize with electronic hardware, computer software or the combination of the two, in order to the interchangeability of hardware and software is clearly described, generally describe composition and the step of each example in the above description according to function.These functions perform with hardware or software mode actually, depend on application-specific and the design constraint of technical scheme.Professional and technical personnel can use distinct methods to realize described function to each specifically should being used for, but this realization should not thought and exceeds scope of the present invention.

The software module that the method described in conjunction with embodiment disclosed herein or the step of algorithm can use hardware, processor to perform, or the combination of the two is implemented.Software module can be placed in the storage medium of other form any known in random access memory (RAM), internal memory, read-only storage (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technical field.

Above-described detailed description of the invention; object of the present invention, technical scheme and beneficial effect are further described; be understood that; the foregoing is only the specific embodiment of the present invention; the protection domain be not intended to limit the present invention; within the spirit and principles in the present invention all, any amendment made, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (6)

1. a face verification method, is characterized in that, described method comprises:

Step 1, principal component analysis and linear discriminant analysis is utilized to carry out pretreatment respectively, comprising the data dimension arranged after principal component analysis dimensionality reduction to higher-dimension face characteristic data;

Step 2, set up discriminate high-order Boltzmann machine, the nodes of hidden layer is set;

Step 3, utilize the diagonalizable strategy of tensor to reduce the model parameter of this discriminate high-order Boltzmann machine;

Step 4, paired human face data is input in discriminate high-order Boltzmann machine, utilize stochastic gradient descent algorithm to maximize other conditional probability of relation object, thus optimize the weight of this Boltzmann machine iteratively, thus obtain final discriminate high-order Boltzmann machine;

Step 5, to described discriminate high-order Boltzmann machine mode input paired human face data to be verified, obtain corresponding the result data.

2. method according to claim 1, is characterized in that, the model of the described discriminate high-order Boltzmann machine in described step 2 comprises two data input layers, the four-layer network network structure of a hidden layer and a classification output layer.

3. method according to claim 2, is characterized in that, described two data input layers are paired training human face data, and described classification output layer represents face verification result; And described discriminate high-order Boltzmann machine has network weight, to obtain next node layer value according to current layer nodal value.

4. method according to claim 3, is characterized in that, the diagonalizable strategy of the tensor in described step 3 is the weight matrix three dimensional weight cubes between described two data input layers and hidden layer being treated to multiple two dimension.

5. method according to claim 1, is characterized in that, the object function of described discriminate high-order Boltzmann machine is other conditional probability of relation object of negative logarithm.

6. a face verification system, is characterized in that, described system comprises: network sets up module, network weight optimizes module and Data Verification module, wherein:

Described network sets up module, for setting up discriminate high-order Boltzmann machine, and arranges the nodes of network hidden layer;

Described network weight optimizes module, utilizes stochastic gradient descent algorithm to minimize the object function of this Boltzmann machine to obtain the Boltzmann machine weight after optimization, thus obtains final discriminate high-order Boltzmann machine model;

Described Data Verification module, for inputting paired human face data to be verified to described discriminate high-order Boltzmann machine, obtain classification layer two nodal value, the relative size comparing two numerical value can draw face verification result data.