CN103207993B - Differentiation random neighbor based on core embeds the face identification method analyzed - Google Patents

Abstract

A kind of differentiation random neighbor based on core of disclosure embeds the face identification method analyzed, it relates to area of pattern recognition, effectively to extract nonlinear discriminant information, obtain for the purpose of higher discrimination, including training process and test process, step is as follows: a) randomly selects l sample of each object and carries out model training, it is thus achieved that corresponding projection matrix B, and remaining data are all as test sample; B) all of training sample and test sample are projected to low dimensional manifold space; C) nearest neighbor classifier is adopted to be identified rate detection. Face identification method provided by the invention is effectively improved discrimination in existing technical foundation, maintains the composition of sample in class and between class well. The present invention can be used in machine learning and pattern recognition category, except recognition of face, it may also be used for the field such as image recognition and target recognition.

Description

Differentiation random neighbor based on core embeds the face identification method analyzed

Technical field

The present invention is a kind of face identification method, specifically, relates to a kind of differentiation random neighbor based on core and embeds the face identification method analyzed, can be used for recognition of face, image recognition, target recognition etc.

Background technology

In society, identity validation has highly important value. In recent years, the biological characteristic of the mankind is increasingly widely used in the identity of individual and authenticates, and compared to traditional method safety, reliable, feature is unique, stability is high, not easily stolen and crack. The intrinsic biological characteristic of the mankind mainly has: DNA, fingerprint, iris, voice, gait, palmmprint, face etc., based on people's cognition to independent personal feature, in conjunction with advanced computer technology and pattern recognition theory, such as DNA identifies that technology, fingerprint identification technology, face recognition technology etc. grow up one after another. For current research level, DNA identifies and fingerprint recognition has higher discrimination, and the most by force but the Condition of Strong Constraint of its use still limits the use of both approaches to reliability. Recognition of face has following powerful advantages compared to other biological feather recognition method: (1) too much participates in without user, contactless collection, without the property invaded; (2) user be there is no any obvious stimulation, it is simple to hide; (3) equipment cost is cheap, mainly adopts photographic head to collect face. Thus recognition of face is as a kind of special biometrics identification technology, have the applied environment of many uniquenesses, as criminal's arrest, automatically-controlled door access control system, customs pass by inspection, credit card confirmation etc.

Recognition of face becomes the study hotspot in Pattern recognition and image processing field already, and current main stream approach is based on the face recognition algorithms of subspace. Such as edge Fisher analytic process, local Fisher techniques of discriminant analysis, minimax distance analysis method and maximum spacing figure imbedding method etc.In recent years, for the input data of nonlinear Distribution structure, it has been proposed that the face identification method of many Nonlinear Dimension Reduction technology, wherein of greatest concern it is based on kernel method and two kinds of technology based on geometry. Such as Isometric Maps method, be locally linear embedding into, laplacian eigenmaps and local tangent space alignment etc. The method that the present invention proposes belongs to the recognition of face based on kernel method, and it can produce nonlinear mapping, represents the popular structure of sample data well, reaches more satisfactory dimensionality reduction effect.

Add up through patent consulting, the patent of existing many recognition of face aspects both at home and abroad: such as, keeps the face identification method (200710114882.4) of embedding and support vector machine based on the neighbour having supervision, based on the face identification method (200710300730.3) of general non-linear discriminating analysis, a kind of face identification method (200810030577.1) etc.

Summary of the invention

The invention solves the problems that the linear dimensionality reduction technology of existing technology can not process the technical problem of the input data of nonlinear Distribution structure very well and can not be effectively improved discrimination, maintain in class well and the shortcoming of the composition of sample between class, it is provided that a kind of differentiations random neighbor based on core embeds the face identification method of analysis.

The technical solution adopted for the present invention to solve the technical problems is:

A kind of differentiation random neighbor based on core embeds the face identification method analyzed, and comprises the following steps:

A) randomly selecting l sample of each object and carry out model training, it is thus achieved that corresponding projection matrix B, remaining data are all as test sample;

B) all of training sample and test sample are projected to low dimensional manifold space;

C) nearest neighbor classifier is adopted to be identified rate detection.

Specifically, in the face identification method of the present invention, specifically include training part and part of detecting two parts, wherein,

Described training process specifically includes following step:

A1 determines training sample matrix X=[x₁,x₂,…,x_N] and class label, it is determined that kernel function, set variance parameter λ and maximum iteration time Mt;

According in step a1, between sample matrix X calculating input sample, Euclidean distance, the Sample Similarity in former space and class label calculate joint probability p to a2 between two_ij:

$p_{\mathrm{ij}}=\left\{\begin{array}{cc}\frac{\exp (-(K_{\mathrm{ii}}+K_{\mathrm{jj}}-{2K}_{\mathrm{ij}})/2{\mathrm{\λ}}^2)}{{\mathrm{\Σ}}_{c_t=c_l}\exp (-(K_{\mathrm{tt}}+K_{\mathrm{ll}}-2K_{\mathrm{tl}})/2{\mathrm{\λ}}^2)}& \mathrm{if}{c}_i=c_j\\ \frac{\exp (-(K_{\mathrm{ii}}+K_{\mathrm{jj}}-{2K}_{\mathrm{ij}})/2{\mathrm{\λ}}^2)}{{\mathrm{\Σ}}_{c_t\≠c_m}\exp (-(K_{\mathrm{tt}}+K_{\mathrm{mm}}-2K_{\mathrm{tm}})/{2\mathrm{\λ}}^2)}& \mathrm{else}\end{array}\right.$

Joint probability p_ijIntroduce Gauss RBF kernel function κ (x, x ')=exp (-σ | | x-x ' | |₂ ²). The given n having class label ties up sample x₁ ¹,x₂ ¹,…,x_N1 ¹,x₁ ²,x₂ ²,…,x_N2 ²,…,x₁ ^C,x₂ ^C,…,x_NC ^C, whereinRepresenting the i-th sample of c class, the total classification number of sample is C, N_iIt is the sample number of the i-th class, K_i=[κ(x₁,x_i),...,κ(x_N,x_i)]^T, it is a column vector;

A3 initializes transformation matrix B⁰So that it is element meets (0,1) Gauss distribution;

A4 calculates joint probability q according to Sample Similarity and the class label of subspace_ij, keep the similarity between similar sample as much as possible by KL divergence and reduce the similarity between foreign peoples's sample, finally utilizing conjugate gradient method to update transformation matrix B^t:

A41 joint probability q_ijFor:

$q_{\mathrm{ij}}=\left\{\begin{array}{cc}\frac{{(1+{(K_i-K_j)}^T{B}^{T}B(K_i-K_j))}^{-1}}{{\mathrm{\Σ}}_{c_t=c_l}{(1+{(K_t-K_l)}^T{B}^{T}B(K_t-K_l))}^{-1}}& \mathrm{if}{c}_i=c_j\\ \frac{{(1+{(K_i-K_j)}^T{B}^{T}B(K_i-K_j))}^{-1}}{{\mathrm{\Σ}}_{c_t\≠c_m}{(1+{(K_t-K_m)}^T{B}^{T}B(K_t-K_m))}^{-1}}& \mathrm{else}\end{array}\right.$

A42 objective cost function is:

$\min C\left(A\right)=\underset{c_i=c_j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}+\underset{c_i\≠c_k}{\mathrm{\Σ}}{p}_{\mathrm{ik}}\log \frac{p_{\mathrm{ik}}}{q_{\mathrm{ik}}}$

A43, under this object function, updates transformation matrix B by classical conjugate gradient method^tIt is iterated solving, wherein carrys out parametrization cost functional by two kinds of methods:

A431. projection matrix B parameter cost functional is utilized:

$\frac{\mathrm{dC}\left(B\right)}{d\left(B\right)}=\underset{c_i=c_j}{\mathrm{\Σ}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}{\left(q_{\mathrm{ij}}\right)}^{\′}+\underset{c_i\≠c_t}{\mathrm{\Σ}}\frac{p_{\mathrm{it}}}{q_{\mathrm{it}}}{\left(q_{\mathrm{it}}\right)}^{\′}$

$=2B[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}(K_i-K_j){(K_i-K_j)}^T+\underset{c_i\≠c_t}{\mathrm{\Σ}}{u}_{\mathrm{it}}(K_i-K_t){(K_i-K_t)}^T]$

For making expression convenient, define following auxiliary variable:

w_ij=[1+(K_i-K_j)^ΤB^ΤB(K_i-K_j)]^-1

u_ij=(p_ij-q_ij)w_ij

$u_{\mathrm{ij}}^{\mathrm{in}}=\left\{\begin{array}{c}\\ \end{array}\right.\begin{array}{cc}{u}_{\mathrm{ij}}& \mathrm{if}{c}_i=c_j\\ 0& \mathrm{else}\end{array}$

$u_{\mathrm{ij}}^{\mathrm{ou}}=\left\{\begin{array}{cc}{u}_{\mathrm{ij}}& \mathrm{if}{c}_i\≠c_j\\ 0& \mathrm{else}\end{array}\right.$

By above-mentioned auxiliary variable, above-mentioned gradient formula can be reduced to:

$\frac{\mathrm{dC}\left(B\right)}{d\left(B\right)}=2B[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}(K_i-K_j){(K_i-K_j)}^T+\underset{c_i\≠c_j}{\mathrm{\Σ}}{u}_{\mathrm{it}}(K_i-K_t){(K_i-K_t)}^T]$

$=4B\left[K(D^{\mathrm{in}}-U^{\mathrm{in}}+D^{\mathrm{ou}}-U^{\mathrm{ou}})K^T\right]$

Wherein diagonal matrix Dⁱⁿ, D^ouIn element by corresponding UⁱⁿAnd U^ouEach column and composition (or each row and, due to UⁱⁿAnd U^ouIt is symmetrical matrix), namelyAnd

A432. utilizing the cost functional in projection matrix A parameter characteristic space, the linear projection transformation matrix A in feature space F can according to nonlinear mapping function It is expressed as (using belowSubstitute):

Wherein A=B Φ, B=[b⁽¹⁾,...,b^(r)]^TAnd

$\frac{{\mathrm{dC}}^F\left(A\right)}{d\left(A\right)}=\underset{c_i=c_j}{\mathrm{\Σ}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}{\left(q_{\mathrm{ij}}\right)}^{\′}+\underset{c_i\≠c_t}{\mathrm{\Σ}}\frac{p_{\mathrm{it}}}{q_{\mathrm{it}}}{\left(q_{\mathrm{it}}\right)}^{\′}$

$=2[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}{\mathrm{BQ}}_{\mathrm{ij}}^{(K_i-K_j)}+\underset{c_i\≠c_t}{\mathrm{\Σ}}{u}_{\mathrm{it}}{\mathrm{BQ}}_{\mathrm{it}}^{(K_i-K_t)}]\mathrm{\Φ}$

WhereinFor N N matrix, the i-th row of matrix are by vector K_i-K_jComposition, jth arranges by vector K_j-K_iComposition, all the other row are made up of null vector;

A5 exports final projection matrix B^t。

Described training process specifically includes following step:

A51. training sample matrix X=[x is determined₁,x₂,…,x_N] and class label;

A52. projection matrix B is utilized^tTraining sample is projected to low dimensional manifold space;

A53. projection matrix B is utilized^tTraining sample is projected to low dimensional manifold space.

The technology design of the present invention: to a kind of new Dimension Reduction Analysis method proposed by Zheng Jianwei etc. recently, is called that differentiation random neighbor embeds (discriminativestochasticneighborembedding, DSNE) and carries out the improvement based on core. DSNE embeds (stochasticneighborembedding in the random neighbor proposed such as Hinton, SNE) and t-distribution SNE(t-distributedstochasticneighborembedding, the t-SNE of improvement that propose such as Laurens) basis on introduce linear projective transformation thought and class label information. Euclidean distance between high dimensional data is converted into probability expression-form by SNE, its cost functional builds criterion calls subspace and has identical form of probability with the former input space, and t-SNE adopts and has symmetric joint probability and express the conditional probability form substituted in SNE, and in subspace, introduce t-distribution show similarity between sample between two. Owing to SNE and t-SNE broadly falls into non-linear unsupervised dimension reduction method, so there is " sample exterior problem " and being not suitable for the defect of pattern discrimination task. 2011, the random neighbor towards popular study proposed by Wu etc. projects (manifold-orientedstochasticneighborprojection, MSNP) " sample exterior problem " is solved well, but being linear unsupervised dimension reduction method based on MSNP, it is still not suitable for pattern recognition task. And linearly have the DSNE of supervision to solve the problem of these two aspects dexterously, but linear feature makes it cannot efficiently solve nonlinear feature extraction problem, and DSNE is for different classes of sample, its probability density still has much room for improvement. The present invention utilizes the thought of kernel method to propose a kind of differentiation random neighbor based on core and embeds the face identification method (kernelDSNE, KDSNE) analyzed, and overcomes the defect of DSNE well.

The invention have the advantage that the input data that can process nonlinear Distribution structure very well, be effectively improved discrimination, maintain in class well and the composition of sample between class.

Accompanying drawing explanation

Fig. 1 is the part face image pattern in ORL face database;

Fig. 2 is the part face image pattern in Yale face database;

Fig. 3 is the change of the discrimination under different subspace dimension in ORL face database;

Fig. 4 is the change of the discrimination under different subspace dimension in Yale face database;

Fig. 5 is the flow chart of the present invention.

Detailed description of the invention

The invention will be further described below. With reference to accompanying drawing 1-4:

A kind of differentiation random neighbor based on core embeds the face identification method analyzed, and comprises the following steps:

A) randomly selecting l sample of each object and carry out model training, it is thus achieved that corresponding projection matrix B, remaining data are all as test sample;

B) all of training sample and test sample are projected to low dimensional manifold space;

C) nearest neighbor classifier is adopted to be identified rate detection.

Specifically, in the face identification method of the present invention, specifically include training part and part of detecting two parts, wherein,

Described training process specifically includes following step:

A1 determines training sample matrix X=[x₁,x₂,…,x_N] and class label, it is determined that kernel function, set variance parameter λ and maximum iteration time Mt;

According in step a1, between sample matrix X calculating input sample, Euclidean distance, the Sample Similarity in former space and class label calculate joint probability p to a2 between two_ij:

$p_{\mathrm{ij}}=\left\{\begin{array}{cc}\frac{\exp (-(K_{\mathrm{ii}}+K_{\mathrm{jj}}-{2K}_{\mathrm{ij}})/2{\mathrm{\λ}}^2)}{{\mathrm{\Σ}}_{c_t=c_l}\exp (-(K_{\mathrm{tt}}+K_{\mathrm{ll}}-2K_{\mathrm{tl}})/2{\mathrm{\λ}}^2)}& \mathrm{if}{c}_i=c_j\\ \frac{\exp (-(K_{\mathrm{ii}}+K_{\mathrm{jj}}-{2K}_{\mathrm{ij}})/2{\mathrm{\λ}}^2)}{{\mathrm{\Σ}}_{c_t\≠c_m}\exp (-(K_{\mathrm{tt}}+K_{\mathrm{mm}}-2K_{\mathrm{tm}})/{2\mathrm{\λ}}^2)}& \mathrm{else}\end{array}\right.$

Joint probability p_ijIntroduce Gauss RBF kernel functionThe given n having class label ties up sample x₁ ¹,x₂ ¹,…,x_N1 ¹,x₁ ²,x₂ ²,…,x_N2 ²,…,x₁ ^C,x₂ ^C,…,x_NC ^C, whereinRepresenting the i-th sample of c class, the total classification number of sample is C, N_iIt is the sample number of the i-th class, K_i=[κ(x₁,x_i),...,κ(x_N,x_i)]^T, it is a column vector;

A3 initializes transformation matrix B⁰So that it is element meets (0,1) Gauss distribution;

A4 calculates joint probability q according to Sample Similarity and the class label of subspace_ij, keep the similarity between similar sample as much as possible by KL divergence and reduce the similarity between foreign peoples's sample, finally utilizing conjugate gradient method to update transformation matrix B^t:

A41 joint probability q_ijFor:

$q_{\mathrm{ij}}=\left\{\begin{array}{cc}\frac{{(1+{(K_i-K_j)}^T{B}^{T}B(K_i-K_j))}^{-1}}{{\mathrm{\Σ}}_{c_t=c_l}{(1+{(K_t-K_l)}^T{B}^{T}B(K_t-K_l))}^{-1}}& \mathrm{if}{c}_i=c_j\\ \frac{{(1+{(K_i-K_j)}^T{B}^{T}B(K_i-K_j))}^{-1}}{{\mathrm{\Σ}}_{c_t\≠c_m}{(1+{(K_t-K_m)}^T{B}^{T}B(K_t-K_m))}^{-1}}& \mathrm{else}\end{array}\right.$

A42 objective cost function is:

$\min C\left(A\right)=\underset{c_i=c_j}{\mathrm{\Σ}}{p}_{\mathrm{ij}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}+\underset{c_i\≠c_k}{\mathrm{\Σ}}{p}_{\mathrm{ik}}\log \frac{p_{\mathrm{ik}}}{q_{\mathrm{ik}}}$

A43, under this object function, updates transformation matrix B by classical conjugate gradient method^tIt is iterated solving, wherein carrys out parametrization cost functional by two kinds of methods:

A431. projection matrix B parameter cost functional is utilized:

$\frac{\mathrm{dC}\left(B\right)}{d\left(B\right)}=\underset{c_i=c_j}{\mathrm{\Σ}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}{\left(q_{\mathrm{ij}}\right)}^{\′}+\underset{c_i\≠c_t}{\mathrm{\Σ}}\frac{p_{\mathrm{it}}}{q_{\mathrm{it}}}{\left(q_{\mathrm{it}}\right)}^{\′}$

$=2B[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}(K_i-K_j){(K_i-K_j)}^T+\underset{c_i\≠c_t}{\mathrm{\Σ}}{u}_{\mathrm{it}}(K_i-K_t){(K_i-K_t)}^T]$

For making expression convenient, define following auxiliary variable:

w_ij=[1+(K_i-K_j)^ΤB^ΤB(K_i-K_j)]^-1

u_ij=(p_ij-q_ij)w_ij

$u_{\mathrm{ij}}^{\mathrm{in}}=\left\{\begin{array}{c}\\ \end{array}\right.\begin{array}{cc}{u}_{\mathrm{ij}}& \mathrm{if}{c}_i=c_j\\ 0& \mathrm{else}\end{array}$

$u_{\mathrm{ij}}^{\mathrm{ou}}=\left\{\begin{array}{cc}{u}_{\mathrm{ij}}& \mathrm{if}{c}_i\≠c_j\\ 0& \mathrm{else}\end{array}\right.$

By above-mentioned auxiliary variable, above-mentioned gradient formula can be reduced to:

$\frac{\mathrm{dC}\left(B\right)}{d\left(B\right)}=2B[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}(K_i-K_j){(K_i-K_j)}^T+\underset{c_i\≠c_t}{\mathrm{\Σ}}{u}_{\mathrm{it}}(K_i-K_t){(K_i-K_t)}^T]$

$=4B\left[K(D^{\mathrm{in}}-U^{\mathrm{in}}+D^{\mathrm{ou}}-U^{\mathrm{ou}})K^T\right]$

Wherein diagonal matrix Dⁱⁿ, D^ouIn element by corresponding UⁱⁿAnd U^ouEach column and composition (or each row and, due to UⁱⁿAnd U^ouIt is symmetrical matrix), namelyAndA432. utilizing the cost functional in projection matrix A parameter characteristic space, the linear projection transformation matrix A in feature space F can according to nonlinear mapping function It is expressed as (using belowSubstitute):

Wherein A=B Φ, B=[b⁽¹⁾,...,b^(r)]^TAnd

$\frac{{\mathrm{dC}}^F\left(A\right)}{d\left(A\right)}=\underset{c_i=c_j}{\mathrm{\Σ}}\frac{p_{\mathrm{ij}}}{q_{\mathrm{ij}}}{\left(q_{\mathrm{ij}}\right)}^{\′}+\underset{c_i\≠c_t}{\mathrm{\Σ}}\frac{p_{\mathrm{it}}}{q_{\mathrm{it}}}{\left(q_{\mathrm{it}}\right)}^{\′}$

$=2[\underset{c_i=c_j}{\mathrm{\Σ}}{u}_{\mathrm{ij}}{\mathrm{BQ}}_{\mathrm{ij}}^{(K_i-K_j)}+\underset{c_i\≠c_t}{\mathrm{\Σ}}{u}_{\mathrm{it}}{\mathrm{BQ}}_{\mathrm{it}}^{(K_i-K_t)}]\mathrm{\Φ}$

WhereinFor N N matrix, the i-th row of matrix are by vector K_i-K_jComposition, jth arranges by vector K_j-K_iComposition, all the other row are made up of null vector;

A5 exports final projection matrix B^t。

Described training process specifically includes following step:

A1 determines training sample matrix X=[x₁,x₂,…,x_N] and class label;

A2 utilizes projection matrix B^tTraining sample is projected to low dimensional manifold space;

A3 utilizes projection matrix B^tTraining sample is projected to low dimensional manifold space.

Two classical face databases of ORL and Yale are adopted to be identified rate detection. In experiment, above-mentioned face database is adjusted to 32 × 32 pixels by unification, and the gray value of every pixel is within 0-255 scope. ORL face database randomly chooses 3 and 5 samples of every class and is identified rate detection, Yale face database then randomly chooses 4 and 6 samples of every class. The present invention adopts two kinds of linear dimension-reduction algorithms of DSNE and MSNP to carry out contrast test, and the random neighbor towards popular study that wherein Wu etc. propose projects and has been verified that in paper that MSNP algorithm is better than the general dimension-reduction algorithm such as SNE, t-SNE, LLTSA, LPP on identification ability. The concrete configuration parameter of various algorithms is as follows: in KDSNE1, KDSNE2 and DSNE, variance parameter λ=0.1 and maximum iteration time are 300; Sample degree of freedom γ=4 and maximum iteration time that in MSNP, Cauchy is distributed are 1000.

Table 1 is all algorithms best identified rate in two face databases and respective subspace dimension (bracket inner digital), and wherein thickened portion is the highest discrimination under identical training sample. As seen from Table 1, KDSNE1 has the discrimination of optimum in ORL face database, and KDSNE2 has the discrimination of optimum in Yale face database, it can be seen that, the height of KDSNE1 and KDSNE2 discrimination is also different with the difference of data base. As for KDSNE1 and the KDSNE2 comparison compared with other algorithms, can learn from figure and table, KDSNE2 in Yale relatively the DSNE of discrimination suboptimum improve > 3%, and although KDSNE2 discrimination promotes average less than 2% in ORL, and DSNE is in close proximity to KDSNE2 in l=5 tests, but DSNE has but used higher subspace dimension just to reach the discrimination being closer to, and is basically inferior to KDSNE1 and KDSNE2 from recognition of face.

Best identified rate that the various algorithm of table 1 obtains in ORL and Yale data base and respective dimensions

Claims (5)

1. embed the face identification method analyzed based on the differentiation random neighbor of core, including training process and test process, it is characterised in that comprise the following steps:

A) randomly select l sample of each object and carry out model training, it is thus achieved that corresponding projection matrix B ∈ R^r×N, wherein N is training sample quantity, and r is sample dimension after projection, and remaining data are all as test sample; In described step a), randomly select l sample of each object and carry out model training and include following five steps:

A1 determines sample matrix X=[x₁,x₂,…,x_N] and class label, it is determined that kernel function, set variance parameter λ and maximum iteration time Mt, wherein x_i∈R^d×N, it is i-th input sample, λ is the variance parameter of corresponding Gaussian function, and Mt is maximum iteration time;

According in step a1, between sample matrix X calculating input sample, Euclidean distance, the Sample Similarity in former space and class label calculate joint probability p to a2 between two_ij;

A3 initializes transformation matrix B⁰So that it is element meets (0,1) Gauss distribution;

A4 calculates joint probability q according to Sample Similarity and the class label of subspace_ij, keep the similarity between similar sample as much as possible by KL divergence and reduce the similarity between foreign peoples's sample, finally utilizing conjugate gradient method to update transformation matrix B^t;

A5 exports final projection matrix B^t;

B) all of training sample and test sample are projected to low dimensional manifold space;

C) nearest neighbor classifier is adopted to be identified rate detection.

2. face identification method according to claim 1, it is characterised in that calculate joint probability p in described step a2_ijTime introduce Gauss RBF kernel function κ (x, x ')=exp (-λ | | x-x ' | |₂ ²), the given n having class label ties up sample x₁ ¹,x₂ ¹,…,x_N1 ¹,x₁ ²,x₂ ²,…,x_N2 ²,…,x₁ ^C,x₂ ^C,…,x_NC ^C, wherein x_i ^cRepresenting the i-th sample of c class, the total classification number of sample is C, N_iBeing the sample number of the i-th class, after introducing kernel function, the joint probability of the sample in former space is:

Wherein K_i=[κ (x₁,x_i),...,κ(x_N,x_i)]^T, it is a column vector being made up of kernel function value.

3. face identification method according to claim 2, it is characterised in that the joint probability q of Operators Space of falling into a trap at described step a4_ijTime have also been introduced Gauss RBF kernel function κ (x, x ')=exp (-λ | | x-x ' | |₂ ²), it may be assumed that

。

4. face identification method according to claim 3, it is characterised in that by minimizing in similar sample and between foreign peoples's sample, respective KL divergence obtains objective cost function in described step a4:

。

5. face identification method according to claim 4, it is characterised in that under described object function, carrys out parametrization cost functional by two kinds of methods:

A41. projection matrix B parameter cost functional is utilized:

A42. projection matrix A ∈ R is utilized^r×dThe cost functional in parameter characteristic space, the linear projection transformation matrix A in feature space can according to nonlinear mapping functionIt is expressed as (using belowSubstitute

Wherein A=B Φ, B=[b⁽¹⁾,...,b^(r)]^TAnd

WhereinFor N N matrix, the i-th row of matrix are by vector K_i-K_jComposition, jth arranges by vector K_j-K_iComposition, all the other row are made up of null vector.