CN105868713A - Method for identifying parallel features integrating human face expression based on nucleus LDA - Google Patents

Abstract

The invention relates to a method for identifying parallel features integrating human face expression based on nucleus LDA. The method comprises the following steps: conducting parallel integration on two groups of characteristic vectors which are subject to different expressions by adopting complex number combination so as to constitute a complex characteristic vector, introducing the nucleus Fisher discriminating criterion to a complex space, which, on the basis of the complex space, addresses the defect of traditional LDA that only linearity problem can be analyzed. The method also includes redefining a within-class scatter matrix, and addresses the problem of small sampling and unbalanced characteristic matrix through adjustable parameters (as shown in the description). According to the invention, the method is improved compared with traditional parallel characteristic integration method and serial characteristic integration method, and not only addresses inability of processing non-linear characteristics in the fields of human face expression identification of traditional LDA, and also addresses the problem of small samples at a certain degree. The method obtains higher identification in experiments of a database of human face expression characteristics.

Description

A kind of Concurrent Feature based on core LDA merges facial expression recognizing method

Technical field

The invention belongs to data identification；Data represent；Record carrier；The technical field of the process of record carrier, relates to especially And a kind of solve small sample problem and eigenmatrix imbalance problem to improve face identification rate by regulatable parameter Concurrent Feature based on core LDA merges facial expression recognizing method.

Background technology

Feature extraction is one of link of most critical in Expression Recognition, extracts and has the feature of discriminating meaning to Accurate classification Human face expression, solving practical problems plays an important role, and feature extraction more puts in place, then Expression Recognition is the most accurate.Expression Recognition Technology may apply to all trades and professions, such as, be applied to police criminal detection field, and the micro-expression to destination object can be assisted to judge Make more structurally sound data supporting.

Along with deepening continuously of correlational study, Feature Fusion gradually receives concern in the industry.Feature Fusion Both merged manifold effective authentication information, eliminated again the information of major part redundancy, thus realized being effectively compressed of information, save The about real-time process of information storage space, beneficially quickening arithmetic speed and the information of carrying out.

The most conventional Feature fusion is mainly serial fusion method.Serial fusion method is first by two or more sets Characteristic vector generates an associating vector according to end to end mode, the most again this new characteristic vector is carried out feature and carries Taking, the method remains manifold authentication information, has certain advantage, but the dimension of new feature after simultaneously causing merging Number sharply increases, thus strengthens the difficulty of subsequent step such as feature extraction and identification, makes examination speed and accuracy rate be greatly reduced.

Thus, research worker carries out substantial amounts of linguistic term to traditional serial fusion method, and wherein, Yang Jian etc. studies and carries Having gone out a kind of method that Concurrent Feature merges, the principle of the method is to utilize complex vector that two or more sets on sample space are special Collection constitutes complex eigenvector space altogether, the feature of the real vector space will be extended to complex vector space.Concurrent Feature is melted Method linear discriminant criterion (LDA) closed extracts effective diagnostic characteristics, LDA be at present conventional feature extracting method it One, but substantially extract due to it is linear character, and nonlinear characteristic is processed Shortcomings.Therefore, related researcher Proposing LDA method based on core, i.e. based on Kernel discriminant analysis method (KDA), the method is high by sample is mapped to one Dimension space, utilizes Fisher method Extraction and discrimination feature at this higher dimensional space, obtains the nonlinear characteristic of original image, in practice it has proved that Core method of discrimination has significant advantage to solving nonlinear problem, but practice also shows, nuclear space during using KDA Dimension is commonly greater than the number of training sample, i.e. small sample problem, a lot, as to Fisher criterion to the improved method of small sample Local weighted, redefine class scatter matrix, solve small sample problem etc. with kernel, even, merge based on Concurrent Feature Characterization method owing to using Fisher to differentiate criterion, the most not only there is small sample problem, there is also fusion spy simultaneously Levying the unbalanced problem of matrix, in making class, Scatter Matrix is not only affected by small sample problem and is lost class inscattering information, also can Produce deviation and greater variance because eigenmatrix is uneven, affect experiment effect.

Summary of the invention

Present invention solves the technical problem that and be, in prior art, though serial fusion method remains manifold discriminating Information, has certain advantage, but after causing merging, the dimension of new feature sharply increases simultaneously, thus strengthen subsequent step such as Feature extraction and the difficulty of identification, make examination speed and accuracy rate be greatly reduced, and uses Concurrent Feature fusion method with linear During differentiating that criterion (LDA) extracts effective diagnostic characteristics, the dimension of nuclear space is commonly greater than the number of training sample, i.e. Small sample problem, even, owing to the method uses Fisher to differentiate criterion, therefore to there is fusion feature matrix uneven the most simultaneously The problem of weighing apparatus, in making class, Scatter Matrix is not only affected by small sample problem and is lost class inscattering information, also can be because of eigenmatrix Uneven and produce deviation and greater variance, the problem having a strong impact on experiment effect, and then provide a kind of optimization based on core The Concurrent Feature of LDA merges facial expression recognizing method.

The technical solution adopted in the present invention is, a kind of Concurrent Feature based on core LDA merges facial expression recognizing method, Said method comprising the steps of:

Step 1.1: use Gabor filter to extract face characteristic from arbitrary human face expression feature database, obtain several Global characteristics vector β on direction；Use PCA algorithm that the local feature of face is extracted, obtain local feature vectors α, By α and β by Concurrent Feature information fusion, obtain matrix X；

Step 1.2: when carrying out Feature Fusion, when two stack features vectors of same sample exist bigger in quantitative relation During difference, by discrete matrix S in class_ωRedefine and solve small sample problem, i.e.

$S_{\mathrm{\ω}}=\frac{1}{c}\underset{i=1}{\overset{c}{\Σ}}{S}_i$

Wherein, S_i=S_i+ kI, k are for controlling parameter, 0≤k≤1, S_iIt it is the covariance matrix of single sample class；Control k's Value is to increase S_ωLittle characteristic vector value, reduce big characteristic vector value so that S_ωDeviation minimum；

Step 1.3: by matrix X in nonlinear mapping Φ transforms to feature space F, i.e. Φ: x_i∈X→Φ(x_i)∈ F；In feature space F, linear Fisher Discrimination Functions is

Wherein, ω ∈ F, and

WithScatter matrix between scatter matrix and class in class corresponding in respectively feature space F,Represent the sample average in the i-th classification in feature space F,Represent spy Levy the average of all samples in the F of space；

Step 1.4: formula (II) and formula (III) are introduced complex space, obtains the class scatter matrix of complex spaceWith in class Scatter Matrix

Wherein,P(ω_i) it is the prior probability of the i-th class training sample；

Step 1.5: by theory of reproducing kernel space, solution vector ω can be launched by all training sample data in feature space F,

Wherein, core discriminant vector ζ=(ζ₁,ζ₂,…,ζ_n)^T, Φ=(Φ (x₁),Φ(x₂),…,Φ(x_n)), ζ is in Φ The optimal core discriminant vectors of discriminant vectors ω；

Step 1.6: after formula (IV), formula (V) and formula (VI) are substituted into formula (I), through matrixing, obtain

Wherein, K () is inner product kernel function,

μ₀=E [Φ (x₁)^HΦ(x_k),...,Φ(x_n)^HΦ(x_k)|ω_i]^H, k=1,2 ..., n；P is core class scatter square Battle array, Q is Scatter Matrix in core class；

Step 1.7: after formula (VII) and formula (VIII) are substituted into formula (I), obtains the linear Fisher mirror in feature space F Other function is

Step 1.8: by formula (VIII) and

Thus

Step 1.9: seek J (ζ) the acquirement maximum when what value of ζ is rightUse Lagrange Algorithm for Solving, Obtain P ζ=λ Q' ζ, try to achieve one group of base characteristic vector, obtain best projection direction ζ, i.e. when taking best projection direction ζ, J (ζ) Obtain maximum；

Step 1.10: utilize best projection direction ζ that matrix X is projected to corresponding feature space, obtain all samples Optimal classification characteristic Y: y_i=ζ^Hx_i, y_i∈Y；

Step 1.11: with y_iIt is characterized value and identifies the face of arbitrary human face expression feature database.

Preferably, in described step 1.1, the local feature of the face of PCA extraction is used to include that mouth feature, eye are special Levy, nose feature.

Preferably, in described step 1.1, X=α+i β or X=β+i α.

Preferably, in described step 1.1, global characteristics vector β is generally the vector on 4～8 directions.

Preferably, in described step 1.2, increase and control parameter k to control the value of k to increase S_ωLittle eigenvalue, reduction Big eigenvalue so that S_ωDeviation minimum.

The Concurrent Feature based on core LDA that the invention provides a kind of optimization merges facial expression recognizing method, by inciting somebody to action Two groups of characteristic vectors expressed through difference use the combining form Parallel Fusion of plural number, constitute complex eigenvector, and by core Fisher differentiate criterion introduce complex space, thus on the basis of complex space solve tradition LDA can only analytical line sex chromosome mosaicism lack Fall into, Scatter Matrix in class is redefined simultaneously, solve small sample problem by regulatable parameter k and eigenmatrix is uneven Problem.The method of the present invention has improvement in varying degrees than Traditional parallel Feature fusion and serial nature fusion method, Not only solving the problem that nonlinear characteristic cannot be processed by tradition LDA in fields such as expression recognition, method is simultaneously one Determining to solve in degree small sample problem, the experiment on the data base of human face expression feature database achieves higher discrimination.

Accompanying drawing explanation

Fig. 1 is present invention classification and recognition when selecting different k under JAFFE data base；

Fig. 2 is present invention classification and recognition when selecting different k under Yale data base.

Detailed description of the invention

Below in conjunction with embodiment, the present invention is described in further detail, but protection scope of the present invention is not limited to This.

The present invention relates to a kind of Concurrent Feature based on core LDA merge facial expression recognizing method, described method include with Lower step:

Step 1.1: use Gabor filter to extract face characteristic from arbitrary human face expression feature database, obtain several Global characteristics vector β on direction；Use PCA algorithm that the local feature of face is extracted, obtain local feature vectors α, By α and β by Concurrent Feature information fusion, obtain matrix X；

Step 1.2: when carrying out Feature Fusion, when two stack features vectors of same sample exist bigger in quantitative relation During difference, by discrete matrix S in class_ωRedefine and solve small sample problem, i.e.

$S_{\mathrm{\ω}}=\frac{1}{c}\underset{i=1}{\overset{c}{\Σ}}{S}_i$

Wherein, S_i=S_i+ kI, k are for controlling parameter, 0≤k≤1, S_iIt it is the covariance matrix of single sample class；Control k's Value is to increase S_ωLittle characteristic vector value, reduce big characteristic vector value so that S_ωDeviation minimum；

Step 1.3: by matrix X in nonlinear mapping Φ transforms to feature space F, i.e. Φ: x_i∈X→Φ(x_i)∈ F；In feature space F, linear Fisher Discrimination Functions is

Wherein, ω ∈ F, and

WithScatter matrix between scatter matrix and class in class corresponding in respectively feature space F,Represent the sample average in the i-th classification in feature space F,Represent spy Levy the average of all samples in the F of space；

Step 1.4: formula (II) and formula (III) are introduced complex space, obtains the class scatter matrix of complex spaceWith in class Scatter Matrix

Wherein,m₀ ^Φ=E{ Φ (x_i)}；P(ω_i) it is the prior probability of the i-th class training sample；

Step 1.5: by theory of reproducing kernel space, solution vector ω can be launched by all training sample data in feature space F,

Wherein, core discriminant vector ζ=(ζ₁,ζ₂,…,ζ_n)^T, Φ=(Φ (x₁),Φ(x₂),…,Φ(x_n)), ζ is in Φ The optimal core discriminant vectors of discriminant vectors ω；

Step 1.6: after formula (IV), formula (V) and formula (VI) are substituted into formula (I), through matrixing, obtain

Wherein,

K(·,·) For inner product kernel function,μ₀=E [Φ (x₁)^HΦ (x_k),...,Φ(x_n)^HΦ(x_k)|ω_i]^H, k=1,2 ..., n；P is core class scatter matrix, and Q is Scatter Matrix in core class；

Step 1.7: after formula (VII) and formula (VIII) are substituted into formula (I), obtains the linear Fisher mirror in feature space F Other function is

Step 1.8: by formula (VIII) and

Thus

Step 1.9: seek J (ζ) the acquirement maximum when what value of ζ is rightUse Lagrange Algorithm for Solving, Obtain P ζ=λ Q' ζ, try to achieve one group of base characteristic vector, obtain best projection direction ζ, i.e. when taking best projection direction ζ, J (ζ) Obtain maximum；

Step 1.10: utilize best projection direction ζ that matrix X is projected to corresponding feature space, obtain all samples Optimal classification characteristic Y: y_i=ζ^Hx_i, y_i∈Y；

Step 1.11: with y_iIt is characterized value and identifies the face of arbitrary human face expression feature database.

In the present invention, when carrying out Concurrent Feature and merging, two stack features values of same sample may be deposited in quantitative relation In bigger difference: big eigenvalue is bigger than normal, and little eigenvalue is less than normal, and eigenmatrix may be made after fusion unbalance.For this reason, it may be necessary to By to discrete matrix S in class_ωRedefine and solve small sample problem, i.e.

In the present invention, two stack features collection A, the B being provided with on sample space Ω, A characteristic of correspondence vector is that α ∈ A, B are corresponding Characteristic vector be β ∈ B.γ=α+i β represents the combination of characteristic vector, and wherein i is imaginary unit, i.e. on sample space Ω Through combination feature space may be defined as C={ α+i β | α ∈ A, β ∈ B}, i.e. matrix X in the present invention, if two stack features to The dimension of amount α Yu β, the then characteristic vector of low-dimensional spot patch foot, this sample space is that n ties up complex vector space, n=max {dimA,dimB}.Definition inner product (X, Y)=X^HY, wherein, X, Y ∈ C, H are conjugate transpose symbol, claim the multiple sky defining inner product Between be the unitary space.Correspondingly, it is assumed that have L known mode class, Scatter Matrix and overall in class scatter matrix, class in the unitary space Scatter Matrix is expressed as:

$S_b=\underset{i=1}{\overset{L}{\Σ}}P\left|{\mathrm{\ω}}_i\right|\left|m_i-m_0\right|{\left|m_i-m_0\right|}^H$

$S_{\mathrm{\ω}}=\underset{i=1}{\overset{L}{\Σ}}P\left|{\mathrm{\ω}}_i\right|E\left\{\left|X-m_i\right|{\left|X-m_i\right|}^H\right|{\mathrm{\ω}}_i\}$

S_t=S_b+S_ω=E{ (X-m₀)(X-m₀)^H}

Wherein, P | ω_i| it is the prior probability of the i-th class training sample, m_i=E | X | ω_i| it is the equal of the i-th class training sample Value,Average for all training samples.

In the present invention, byAnd X, Y ∈ C, C={ α+i β | α ∈ A, β ∈ B}, analogizes discrete matrix S in the class understood in step 1.2_ωCovariance matrix S with single sample class_iIt is and local feature The rectangle that vector α is relevant with global characteristics vector β.

In the present invention, thus, it is possible to increase within class scatter matrix S by adjusting parameter k_ωLittle eigenvalue, reduction Its big eigenvalue suppresses deviation, thus reaches to improve the purpose of discrimination.

In the present invention, x_iIt it is the sample vector in matrix X.

In the present invention, Q is nonnegative definite matrix, and parameter k and the long-pending of unit matrix I are positive definite, then Q+kI is exactly positive definite, So formulaMiddle ζ has solution, and unrelated with the singularity of Scatter Matrix Q in core class, and the singularity solving Q is asked Topic, thus the problem of small sample is resolved, at suppression S_ωAlso spy that may be present in fusion is balanced while deviation Levy vectorial unequal problem.

In the present invention, in order to give full expression to expression information, use PCA algorithm to extract local message, use Gabor filter Extract the Global Information of human face expression, PCA algorithm is first carried out and obtains local feature vectors α, change by α with via Gabor To global characteristics vector β obtain matrix X by Concurrent Feature information fusion, analyze dimension size and the training sample of α and β Number, the value of adjusted controllable parameter k, in order to redefine Scatter Matrix in class, it is calculated divergence in the core class of complex space Matrix and core class scatter matrix, solve Generalized Characteristic Equation P ζ=λ Q' ζ, find one group of base characteristic vector, obtain best projection Direction ζ.X is projected to a t dimension space, obtains the optimal classification feature of all samples: Y_i=ζ^HX_i。

In described step 1.1, the local feature of the face of PCA extraction is used to include that mouth feature, eye feature, nose are special Levy.

In described step 1.1, X=α+i β or X=β+i α.

In the present invention, if two assemblage characteristic spaces on sample space Ω are respectively defined as C₁=α+i β | α ∈ A, β ∈ B}, C₂=β+i α | α ∈ A, β ∈ B}, if matrix H (α, β)=(α+i β) (α+i β)^H, H (α, β)=(β+i α) (β+i α)^H,Wherein, α, β are the real vector of n dimension, if, α=(α₁,...,α_n)^T, β=(b₁,...,b_n)^T, then

${\[H(\mathrm{\α},\mathrm{\β})\]}_{kl}=(a_k+{\mathrm{ib}}_k)\stackrel{\‾}{(a_l+{\mathrm{ib}}_l)}=(a_k{a}_l+b_k{b}_l)+i(a_l{b}_k-a_k{b}_l)$

${\[H(\mathrm{\α},\mathrm{\β})\]}_{kl}=(b_k+{\mathrm{ia}}_k)\stackrel{\‾}{(b_l+{\mathrm{ia}}_l)}=(a_k{a}_l+b_k{b}_l)-i(a_l{b}_k-a_k{b}_l)$

Therefore

Therefore,

Therefore, obtain X=α+i β or X=β+i α.

In described step 1.1, global characteristics vector β is generally the vector on 4～8 directions.

In the present invention, global characteristics vector β is generally the vector on 4～8 directions, and direction is divided equally, as with Gabor Extracting the global characteristics on six direction, six direction is respectively 0, π/6,2 π/6,3 π/6,4 π/6,5 π/6.

In described step 1.2, increase and control parameter k to control the value of k to increase S_ωLittle eigenvalue, reduce big feature Value so that S_ωDeviation minimum.

In the present invention, select JAFFE and Yale two expression storehouse；For without loss of generality, grader uses KNN (near based on K Adjacent rule).Wherein, JAFFE Facial expression database is made up of the 213 width images of 10 people, and everyone shows 7 kinds of expressions；And Yale Expression storehouse comprises 4 kinds of expressions, totally 165 width images of 15 people, is 8 gray level images of 320 × 243.

In the present invention, from JAFFE, select secondary 60 pair altogether of everyone 6 kinds expressions each, from Yale, select 4 kinds of tables of 10 people Feelings each one are secondary amounts to 40 width, and using every kind of front M width image expressed one's feelings as training sample, rear (10-M) width is as test sample.This Sample, training sample and test sample form typical high dimensional and small sample size problem.Circulate 5 times, take the meansigma methods conduct of institute's espressiove Experimental result.

K is changed between [0,1], can be consecutive variations can also be Discrete Change.From Fig. 1 and Fig. 2 it can be seen that When M >=4 owing to there being abundant training sample, k value is the least also can guarantee that variance and deviation balance, works as M=2, when 3, due to Lack of training samples, discrete matrix S in class_ωThere will be great number variance.Therefore, it is necessary for increasing and controls parameter to increase the little of it Eigenvalue, reduce big eigenvalue and suppress its deviation, thus control the variance of kernel, can be only achieved reasonable identification Rate.Experiment understands along with the increase discrimination of k value constantly increases simultaneously, obtains peak value, and identifying after increasing to certain value Near rate peak, discrimination change is slowly.

Present invention demonstrates that different expression storehouse or right with discrimination peak institute in the case of sample values different in the same storehouse The k that answers also differs, and the present invention is more suitable in small sample operation, when sample is less, uses regulatable parameter k to solve Certainly small sample problem and eigenmatrix imbalance problem effect are more preferable.

In the present invention, two data bases of JAFFE and Yale take a facial expression image conduct in a people every kind expression Test sample, remaining is as training sample.Circulating 5 times, average as discrimination, k used under JAFFE data base is K value used under 0.9, Yale data base is 0.85, obtains following the result:

The discrimination (%) of distinct methods under table 1 JAFFE data base

	Angry	Happily	Frightened	Sad	Surprised	Detest	Average recognition rate
								Serial nature merges	84.5	85.3	84.6	85.3	85.1	84.7	84.9
Concurrent Feature merges	89.5	90.7	90.2	90.4	90.8	89.7	90.2
								The present invention	93.3	93.7	93.1	93.6	93.5	93.0	93.4

The nicety of grading (%) of distinct methods grader under table 2 Yale data base

	Happily	Neutral	Sad	Surprised	Average recognition rate
						Serial nature merges	89.9	87.6	88.6	89.8	88.9
Concurrent Feature merges	92.3	91.5	92.7	91.5	92.1
						The present invention	95.6	95.1	96.3	96.8	95.9

From Table 1 and Table 2, the discrimination that the method that serial nature merges obtains in three kinds of methods is minimum, this It is to extract due to Gabor characteristic to produce bigger characteristic dimension, in the case of there is no dimensionality reduction, merges through serial nature and will make spy Levy dimension to sharply increase.The average recognition rate that traditional LDA Concurrent Feature merges on two Sample Storehouses is respectively 90.2% He 92.1%, and method proposed by the invention, respectively obtained the discrimination of 93.4% and 95.90%, demonstrated institute of the present invention The effectiveness of Parallel Fusion method is proposed.

In the present invention, again take method of testing unrelated with people, test and average for 5 times.K value used in this experiment is same Sample is respectively 0.9 and 0.85.

The discrimination (%) of different characteristic extracting method under table 3 JAFFE data base

	PCA	Gabor	Traditional parallel merges	The present invention
					Angry	84.6	87.4	90.2	92.7
Happily	84.3	89.3	90.6	92.7
					Frightened	82.7	87.4	89.5	92.6
Sad	84.7	86.7	91.5	91.8
					Surprised	82.9	86.9	90.4	93.4
Detest	83.7	87.2	89.6	93.9
					Average recognition rate	83.8	84.2	90.3	92.9

The nicety of grading (%) of different characteristic extracting method under table 4 Yale data base

	PCA	Gabor	Traditional parallel merges	The present invention
					Happily	86.6	90.4	93.6	95.3
Neutral	85.4	89.3	92.5	95.7
					Sad	86.8	89.5	92.4	94.8
Surprised	86.9	90.4	91.8	95.3
					Average recognition rate	86.4	89.9	92.6	95.3

Can draw from table 3 and table 4, compared with single human face expression PCA feature and Gabor characteristic recognition result, will Two kinds of features use the Parallel Fusion strategy mentioned to merge, and improve discrimination, and this is due to Parallel Fusion both sides Method contains local feature and the global feature of human face expression, also prevent information superfluous while remaining its effective authentication information Remaining.And the identification that the method that the present invention proposes is owing to fully having processed the nonlinear characteristic of facial expression image, on two data bases Rate reaches 92.9% and 95.3%, has been respectively increased 2.6% and 2.7% than traditional method, it was demonstrated that in nuclear space Concurrent Feature The effectiveness of convergence strategy.

The present invention solves in prior art, though serial fusion method remains manifold authentication information, has one Fixed advantage, but after causing merging, the dimension of new feature sharply increases simultaneously, thus strengthen subsequent step such as feature extraction with The difficulty identified, makes examination speed and accuracy rate be greatly reduced, and uses Concurrent Feature fusion method linear discriminant criterion (LDA), during extracting effective diagnostic characteristics, the dimension of nuclear space is asked commonly greater than the number of training sample, i.e. small sample Topic, even, owing to the method uses Fisher to differentiate criterion, therefore exists that fusion feature matrix is unbalanced asks the most simultaneously Topic, in making class, Scatter Matrix is not only affected by small sample problem and is lost class inscattering information, also can be uneven because of eigenmatrix And producing deviation and greater variance, the problem having a strong impact on experiment effect, by by two groups of characteristic vectors expressed through difference Use the combining form Parallel Fusion of plural number, constitute complex eigenvector, and core Fisher is differentiated, and criterion introduces complex space, thus Solve on the basis of complex space tradition LDA can only the defect of analytical line sex chromosome mosaicism, Scatter Matrix in class is redefined simultaneously, logical Cross regulatable parameter k to solve small sample problem and eigenmatrix imbalance problem.The method of the present invention is more special than Traditional parallel Levy fusion method and serial nature fusion method has improvement in varying degrees, not only solve tradition LDA and know at human face expression The problem that not etc. nonlinear characteristic cannot not processed by field, method to some extent solves small sample problem simultaneously, people Experiment on the data base in face expressive features storehouse achieves higher discrimination.

Claims (5)

1. a Concurrent Feature based on core LDA merges facial expression recognizing method, it is characterised in that: described method includes following Step:

Step 1.1: use Gabor filter to extract face characteristic from arbitrary human face expression feature database, obtain several directions On global characteristics vector β；Use PCA algorithm that the local feature of face is extracted, obtain local feature vectors α, by α and β passes through Concurrent Feature information fusion, obtains matrix X；

Step 1.2: when carrying out Feature Fusion, when two stack features vectors of same sample exist bigger difference in quantitative relation Time, by discrete matrix S in class_ωRedefine and solve small sample problem, i.e.

$S_{\mathrm{\ω}}=\frac{1}{c}\underset{i=1}{\overset{c}{\Σ}}{S}_i$

Wherein, S_i=S_i+ kI, k are for controlling parameter, 0≤k≤1, S_iIt it is the covariance matrix of single sample class；Control k value with Increase S_ωLittle characteristic vector value, reduce big characteristic vector value so that S_ωDeviation minimum；

Step 1.3: by matrix X in nonlinear mapping Φ transforms to feature space F, i.e. Φ: x_i∈X→Φ(x_i)∈F；? In feature space F, linear Fisher Discrimination Functions is

Wherein, ω ∈ F, and

WithScatter matrix between scatter matrix and class in class corresponding in respectively feature space F,Table Show the sample average in the i-th classification in feature space F,Represent owning in feature space F The average of sample；

Step 1.4: formula (II) and formula (III) are introduced complex space, obtains the class scatter matrix of complex spaceWith divergence in class Matrix

Wherein,P(ω_i) it is the prior probability of the i-th class training sample；

Step 1.5: by theory of reproducing kernel space, solution vector ω can be launched by all training sample data in feature space F,

Wherein, core discriminant vector ζ=(ζ₁,ζ₂,…,ζ_n)^T, Φ=(Φ (x₁),Φ(x₂),…,Φ(x_n)), ζ be Φ differentiates to The optimal core discriminant vectors of amount ω；

Step 1.6: after formula (IV), formula (V) and formula (VI) are substituted into formula (I), through matrixing, obtain

Wherein,

$Q=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}P\left({\mathrm{\ω}}_i\right)E\left\{\left({\mathrm{\η}}_{x_i^j}-{\mathrm{\μ}}_i\right){\left({\mathrm{\η}}_{x_i^j}-{\mathrm{\μ}}_i\right)}^H|{\mathrm{\ω}}_i\right\},{\mathrm{\η}}_{x_i^j}={\left(K\left(x_1,x_i^j\right),...,K\left(x_n,x_i^j\right)\right)}^H,$

K () is inner product kernel function,μ₀=E [Φ (x₁)^H Φ(x_k),...,Φ(x_n)^HΦ(x_k)|ω_i]^H, k=1,2 ..., n；P is core class scatter matrix, and Q is Scatter Matrix in core class；

Step 1.7: after formula (VII) and formula (VIII) are substituted into formula (I), obtains the linear Fisher in feature space F and differentiates letter Number is

Step 1.8: by formula (VIII) and

Thus

Step 1.9: seek J (ζ) the acquirement maximum when what value of ζ is rightUse Lagrange Algorithm for Solving, obtain P ζ=λ Q' ζ, tries to achieve one group of base characteristic vector, obtains best projection direction ζ, and i.e. when taking best projection direction ζ, J (ζ) obtains Maximum；

Step 1.10: utilize best projection direction ζ that matrix X is projected to corresponding feature space, obtain the optimal of all samples Characteristic of division Y:y_i=ζ^Hx_i, y_i∈Y；

Step 1.11: with y_iIt is characterized value and identifies the face of arbitrary human face expression feature database.

A kind of Concurrent Feature based on core LDA the most according to claim 1 merges facial expression recognizing method, and its feature exists In: in described step 1.1, use the local feature of the face of PCA extraction to include mouth feature, eye feature, nose feature.

A kind of Concurrent Feature based on core LDA the most according to claim 1 merges facial expression recognizing method, and its feature exists In: in described step 1.1, X=α+i β or X=β+i α.

A kind of Concurrent Feature based on core LDA the most according to claim 1 merges facial expression recognizing method, and its feature exists In: in described step 1.1, global characteristics vector β is generally the vector on 4～8 directions.

A kind of Concurrent Feature based on core LDA the most according to claim 1 merges facial expression recognizing method, and its feature exists In: in described step 1.2, increase and control parameter k to control the value of k to increase S_ωLittle eigenvalue, reduce big eigenvalue so that S_ωDeviation minimum.