CN104700089A - Face identification method based on Gabor wavelet and SB2DLPP - Google Patents

Abstract

The invention discloses a face identification method based on Gabor wavelet and SB2DLPP. The face identification method based on the Gabor wavelet and SB2DLPP mainly includes pre-treatment, feature extraction, feature dimension reduction and classification identification, and to be specific, the face identification method includes that (1) pre-treating all the face images in a known face database, wherein the pre-treatment includes scale normalization and histogram equalization; (2) using the Gabor wavelet to extract features of the pre-treated face images; (3) leading in class information, and using a supervised bidirectional two-dimensional local preserving projection (SB2DLPP) algorithm to reduce the dimensions of the high-dimensional image features extracted through the step (2) to extract feature matrices mapped to a low-dimensional sub-space; (4) using a nearest neighbor classifier to perform classification identification. The face identification method based on the Gabor wavelet and SB2DLPP uses the Gabor wavelet and improved SB2DLPP algorithm to identify images, the problems that a traditional face identification method is easy to be influenced by light, expression and the like external factors are overcame, and the face identification rate is effectively improved.

Description

A kind of face identification method based on Gabor wavelet and SB2DLPP

Technical field

The invention belongs to image procossing, computer vision, mode identification technology, particularly a kind of face identification method based on Gabor wavelet and SB2DLPP.

Background technology

Face recognition technology extracts face characteristic by computing machine, and carry out a kind of technology of identification and checking by these features.Face, as an important biological characteristic, communicates very important information.Because the feature of its uniqueness, individual identity information can be passed on, so the face recognition technology biometrics identification technology that to be development in recent years the fastest.At present, recognition of face is all with a wide range of applications in fields such as credit card discriminating, safety check, monitoring.But due to the complicacy of human face structure, the reasons such as the face change that expression, age and attitude cause and the polytrope that face imaging process causes by illumination, shooting angle etc., it is a challenging research field that recognition of face is still recognized so far.

Gabor wavelet conversion coefficient mates as feature good visual characteristic and Biological background, and can overcome the impact that the factors such as illumination, yardstick, attitude produce through the image of Gabor filtering, therefore Gabor wavelet is widely used in recognition of face in recent years.But it is too high that Gabor wavelet converts the intrinsic dimensionality extracted, directly utilize Gabor characteristic to carry out discriminator, calculated amount can be very large, cannot meet the requirement of real-time of recognition of face, therefore need to carry out dimensionality reduction to Gabor characteristic.

Current dimension-reduction algorithm is mainly divided into linear subspaces method and manifold learning.Dimension-reduction algorithm based on linear subspaces mainly contains principal component analysis (PCA) (PCA), linear discriminant analysis (LDA) etc.But research in recent years has proved that face sample is distributed in one most probably and is embedded on the low-dimensional non-linearity manifold of higher dimensional space.Now, traditional linear method often cannot seek the nonlinear organization of high dimensional data inside, and manifold learning also produces in this context.Representative manifold learning has: local linear embeds (LLE), Isometric Maps (Isomap), laplacian eigenmaps (LE) etc.Due to LLE, the Method of Nonlinear Dimensionality Reductions such as LE just obtain good low-dimensional to training sample and embed, be difficult to the low dimension projective obtaining new sample point, the people such as He propose the linearized algorithm of LE and LLE: locality preserving projections (LPP) and neighbour keep embedding (NPE).

But these algorithms, all based on vector, need to be first one-dimensional vector by two dimensional image matrix conversion.This conversion can cause two problems usually: one is lose some important structural informations; But the limited sample of the vector sum of higher-dimension can face singular value problem.Therefore, some two-dimentional algorithms are suggested in recent years, such as: 2DPCA, 2DLPP etc.

Realizing in process of the present invention, inventor proposes a kind of SB2DLPP algorithm, utilizes classification information to strengthen the distinguishing ability of algorithm, carries out dimensionality reduction simultaneously to the face characteristic extracted by Gabor wavelet.The present invention has good robustness to changes such as illumination, and can obtain good recognition effect.

Summary of the invention

For the defect existed in above-mentioned prior art or deficiency, the object of the invention is to, propose a kind of face identification method based on Gabor wavelet and SB2DLPP.First do not utilize this feature of classification information for unsupervised learning algorithm B2DLPP, the basis of B2DLPP is introduced classification information and B2DLPP is become have the B2DLPP of supervision (SB2DLPP).Secondly for the feature of facial image, choose one group of suitable Gabor wavelet kernel function, and by this group Gabor wavelet kernel function, Gabor wavelet conversion is carried out to image, extract the facial image feature external factor such as illumination, expression all to certain robustness.On this basis, application SB2DLPP carries out dimensionality reduction to the eigenmatrix that Gabor wavelet is extracted, and finally utilizes nearest neighbor classifier to carry out Classification and Identification to facial image.

Technical scheme of the present invention is, based on the face identification method of Gabor wavelet and SB2DLPP, it is characterized in that: comprise the steps:

1) image in known face database is divided into two parts, a part is for building training image data set, and remainder is for building test pattern data set;

2) pre-service is carried out for all original facial images of training image data centralization;

3) Gabor wavelet is adopted to step 2) pretreated training facial image carries out feature extraction;

4) keep at bidirectional two-dimensional partial projection classification information is introduced on the basis of (B2DLPP), to step 3) the higher-dimension facial image feature application that extracts has the bidirectional two-dimensional partial projection of supervision to keep (SB2DLPP) algorithm to carry out dimensionality reduction, thus obtains the low-dimensional characteristic coefficient of low-dimensional subspace projection matrix U and V and training facial image;

5) to step 1) all original facial images of test pattern data centralization that obtain carry out pre-service;

6) to step 5) pretreated test facial image adopts Gabor wavelet to carry out feature extraction;

7) by step 6) Gabor characteristic of test pattern that obtains represents and projects to low n-dimensional subspace n thus the low-dimensional characteristic coefficient obtaining testing facial image;

8) contrast step 7) in the low-dimensional characteristic coefficient of test facial image to be identified and step 4) in train the low-dimensional characteristic coefficient of facial image, employing nearest neighbor classifier carries out Classification and Identification, exports final discrimination.

Step 2) and step 5) comprise the steps:

A) carry out dimension normalization to original image, the size of every piece image adjusts to unified size m × n.

B) histogram equalization method is adopted to adjust the contrast of original facial image.Algorithm of histogram equalization comprises: the number of times that each gray level of statistic histogram occurs; Accumulative normalization histogram; The pixel value that computed image is new.

Step 3) and step 6) comprise the steps:

A) one group of two-dimensional Gabor kernel function is defined it is defined as:

Wherein, v and u represents scale factor and the direction of Gabor wavelet respectively, the position of z=(x, y) representation space territory pixel, k _{u, v}for the wave vector of small echo, be defined as: wherein, k _v=k _max/ f ^vand φ _u=π u/8, k _maxrepresent maximum frequency.

B) facial image and Gabor wavelet bank of filters are carried out convolution.A given width facial image A _i, then image A _igabor characteristic obtained by following Convolution Formula: wherein, z=(x, y), * represents convolution, represent Gabor kernel function, G _{u, v}z () represents the convolved image of the Gabor kernel function when being of a size of v and direction is u.Here only have chosen 2 yardsticks (v ∈ { 0,1}), 8 groups of Gabor filter that 4 directions (u ∈ { 0,2,4,6}) combines are for extracting feature.Application convolution theorem, G _{u, v}z () obtains by fast fourier transform.Image A _ithe feature extracted by Gabor wavelet form S set={ G _{u, v}(z), v ∈ { 0,1}, u ∈ { 0,2,4,6}}.

C) in order to comprise different spatial frequencys, space scale, positional information and set direction information, splice G by row _{u, v}a Gabor characteristic matrix Y strengthened is obtained after (z) _i.Before splicing, in order to reduce space dimensionality, first can to G _{u, v}z () employing down-sampling factor is the down-sampling process of ρ.If represent the eigenmatrix after down-sampling, then the Gabor characteristic matrix strengthened be defined as: $Y_i^{\mathrm{\ρ}}=\{G_{\mathrm{0,0}}^{\mathrm{\ρ}},G_{0,2}^{\mathrm{\ρ}},G_{\mathrm{0,4}}^{\mathrm{\ρ}},G_{\mathrm{0,6}}^{\mathrm{\ρ}},G_{\mathrm{1,0}}^{\mathrm{\ρ}},G_{\mathrm{1,2}}^{\mathrm{\ρ}},G_{\mathrm{1,4}}^{\mathrm{\ρ}},G_{\mathrm{1,6}}^{\mathrm{\ρ}}\}.$

Described step 4) comprise the steps:

A) supposition is through step 2) pretreated training image data set comprises the training image A that N number of size is m × n ₁, A ₂..., A _n, every piece image is through step 3) and the Gabor characteristic extracted after Gabor wavelet filtering is: because the characteristics of image dimension that Gabor wavelet is extracted is too high, so adopt SB2DLPP algorithm to carry out dimensionality reduction, be specially:

1. according to classification information structure weight matrix S.

Assuming that training image sample set can be divided into r class, the image pattern point set of l class is designated as C _l, the image pattern point number that l class comprises is then by | C _l| represent.Again the image in training set being carried out sequence makes mutually of a sort data sequence adjacent.If Y _iwith sample point Y _jbelong to same class, be endowed the weighted value of a non-zero between two sample points, otherwise weight is set to 0.Weight function is defined as:

2. low-dimensional maps, and solves optimum projection matrix U and V.

Objective function derive can obtain through simple algebraically:

$F(u,V)=\mathrm{tr}\left[U^{T}Q(L\⊗{\mathrm{VV}}^T)Q^{T}U\right]=\mathrm{tr}\left[V^T{P}^T(L\⊗{\mathrm{UU}}^T)\mathrm{PV}\right]$

Wherein, D is a diagonal matrix, the value on diagonal line be the row of weight matrix S or row sum namely matrix L=D-S, for the Kronecker Product Operator of matrix, tr (*) is the mark of square formation. for splicing all image arrays by ranks, Q=[Y ₁, Y ₂..., Y _n] for splice all image arrays by row.

In order to eliminate the impact of size factor in projecting direction, add constraint condition:

$\begin{array}{c}G(U,V)=\underset{i}{\mathrm{\Σ}}{D}_{\mathrm{ii}}{\left|\right|A_i\left|\right|}^2=1\⇒\mathrm{tr}\left(U^{T}Q(D\⊗{\mathrm{VV}}^T)Q^{T}U\right)=1\\ \⇒\mathrm{tr}\left(V^T{P}^T(D\⊗{\mathrm{UU}}^T)\mathrm{PV}\right)=1\end{array}$

Minimization problem is further converted to:

$\underset{U,V}{\arg \min }\frac{\mathrm{tr}\left(V^T{P}^T(L\⊗{\mathrm{UU}}^T)\mathrm{PV}\right)}{\mathrm{tr}\left(V^T{P}^T(D\⊗{\mathrm{UU}}^T)\mathrm{PV}\right)}$

$\underset{U,V}{\arg \min }\frac{\mathrm{tr}\left(U^{T}Q(L\⊗{\mathrm{VV}}^T)Q^{T}U\right)}{\mathrm{tr}\left(U^{T}Q(D\⊗{\mathrm{VV}}^T)Q^{T}U\right)}$

Therefore, the value of U and V solves generalized eigenvalue iteratively and proper vector obtains, and its computing formula is converted into by two formulas above:

$P^T(L\⊗{\mathrm{UU}}^T)\mathrm{Pv}=\mathrm{\γ}{P}^T(D\⊗{\mathrm{UU}}^T)\mathrm{Pv}$

$Q(L\⊗{\mathrm{VV}}^T)Q^{T}u=\mathrm{\λQ}(D\⊗{\mathrm{VV}}^T)Q^{T}u$

Wherein, iteration initialization condition setting is: U ₀=I _s.

Note column vector u ₁, u ₂..., u _ll minimal eigenvalue characteristic of correspondence vector of first formula above, column vector v ₁, v ₂..., v _rr minimal eigenvalue characteristic of correspondence vector of second formula above, then optimum left projection matrix U=(u ₁, u ₂..., u _l) and the right projection matrix V=of optimum (v ₁, v ₂..., v _r).

B) calculation training view data concentrates each width to train facial image Y _iproject to the low-dimensional characteristic coefficient M of low n-dimensional subspace n _i, its computing formula is: M _i=U ^ty _iv.

Described step 7) be specially: assuming that test pattern data set is through step 5) result that obtains after pre-service is designated as then through step 6) the test facial image Gabor characteristic extracted after Gabor wavelet filtering represents and is designated as the Gabor characteristic of test pattern represents X _gproject to low n-dimensional subspace n thus obtain testing the low-dimensional characteristic coefficient T of facial image _gby formula: T _g=U ^tx _gv obtains.

Described step 8) comprise the steps:

A) the low-dimensional characteristic coefficient T of each width test pattern is calculated _g(g=1,2 ..., H) with the low-dimensional characteristic coefficient M of all training images _i(i=1,2 ..., N) between distance and be designated as dist _gi, distance computing formula is:

$\mathrm{dist}(T_g,M_i)={\mathrm{\Σ}}_{x=1}^x{\mathrm{\Σ}}_{y=1}^y[T_g^{(x,y)}-M_i^{(x,y)}]$

To each width test pattern T _gdistance set { the dist obtained _gi, i=1,2 ..., the element in N} sorts from small to large, thus the low-dimensional characteristic coefficient obtaining training image corresponding to minor increment is M _i.If M _icorresponding training image A _iclassification be l, so T _gcorresponding test pattern B _gthen be identified l class.

B) the test face picture number correctly identified is added up.Correct identification test pattern number is designated as num, and initial value is 0.Assuming that test pattern B _ggeneric is e, if e=l, and so num=num+1; Otherwise num value is constant.

C) export discrimination, discrimination computing formula is: accuracy=num/H × 100%, and wherein, num is the test face picture number correctly identified, and H is the sum for the facial image tested.

Compared with prior art, its remarkable advantage: (1) makes full use of the advantage of Gabor wavelet, makes the inventive method still can keep good recognition performance when facial image is subject to the interference of the external factor such as illumination, expression in the present invention.(2) adopt SB2DLPP algorithm in the dimensionality reduction stage of high dimensional feature, take into full account classification information, strengthen the distinguishing ability of algorithm, improve discrimination.

Accompanying drawing explanation

Fig. 1 is a kind of main flow chart of the face identification method based on Gabor wavelet and SB2DLPP, and Fig. 2 is the part face image graph of pretreated training image data centralization on Yale storehouse.

Embodiment

For enabling above-mentioned purpose of the present invention, feature and advantage become apparent more, below in conjunction with the drawings and specific embodiments, the present invention is further detailed explanation, these accompanying drawings are the schematic diagram of simplification, the basic structure of this aspect is only described in a schematic way, and therefore it only shows the formation relevant with the present invention.

Experiment porch of the present invention is MATLAB 2012, and main frame configures: the central processing unit of Inter (R) Core (TM) I5-3470,4GB internal memory.This experiment porch the inventive method processes view data in Yale face database, finally obtains discrimination.

Embodiment: the embodiment of the present invention uses disclosed Yale face database to carry out identifying operation.The Yale storehouse that the embodiment of the present invention uses contains the 165 width facial images of 15 people, the facial image that everyone has 11 width different, these images have different expression (sad, happy, normal, sleepiness, surprised, nictation), different illumination conditions (left light source, front light source, right light source) and different face detail (wear/do not wear glasses).

As shown in Figure 1, a kind of face identification method based on Gabor wavelet and SB2DLPP, is characterized in that comprising following steps:

Step S1, from Yale database, everyone chooses 6 width images totally 90 width facial images is to build training image data set, and remaining totally 75 width images are for building test pattern data set.

Step S2, carries out pre-service to all original facial images of training image data centralization.

A) carry out dimension normalization to training image data set, the size of every piece image adjusts to unified size 100 × 100.

B) histogram equalization method is adopted to adjust the contrast of original facial image.

Through the pretreated training image data set of step S2 comprise 90 width sizes be 100 × 100 training image be designated as A ₁, A ₂..., A ₉₀.The part facial image of pretreated training image data centralization as shown in Figure 2.

Step S3, adopts Gabor wavelet to carry out feature extraction to the pretreated training facial image of step S2.

A) one group of two-dimensional Gabor kernel function is defined it is defined as:

Wherein, v and u represents scale factor and the direction of Gabor wavelet respectively, the position of z=(x, y) representation space territory pixel, k _{u, v}for the wave vector of small echo, be defined as: wherein, k _v=k _max/ f ^vand φ _u=π u/8, k _maxrepresent maximum frequency.In case study on implementation process, optimum configurations is: k _max=pi/2, f=2, σ=2 π.

B) facial image and Gabor wavelet bank of filters are carried out convolution.A given width facial image A _i, then image A _igabor characteristic obtained by following Convolution Formula: wherein, z=(x, y), * represents convolution, represent Gabor kernel function, G _{u, v}z () represents the convolved image of the Gabor kernel function when being of a size of v and direction is u.Here only have chosen 2 yardsticks (v ∈ { 0,1}), 8 groups of Gabor filter that 4 directions (u ∈ { 0,2,4,6}) combines are for extracting feature.Application convolution theorem, G _{u, v}z () obtains by fast fourier transform.Image A _ithe feature extracted by Gabor wavelet form S set={ G _{u, v}(z), v ∈ { 0,1}, u ∈ { 0,2,4,6}}.

C) first to G _{u, v}z () employing down-sampling factor is the down-sampling process of ρ, in case study on implementation process, the value of decimation factor ρ is set to 4, then splices the G after sampling by row _{u, v}z () obtains a Gabor characteristic matrix Y strengthened _i.If represent the eigenmatrix after down-sampling, then the Gabor characteristic matrix strengthened be defined as:

$Y_i^4=\{G_{\mathrm{0,0}}^4,G_{0,2}^4,G_{\mathrm{0,4}}^4,G_{\mathrm{0,6}}^4,G_{\mathrm{1,0}}^4,G_{\mathrm{1,2}}^4,G_{\mathrm{1,4}}^4,G_{\mathrm{1,6}}^4\}$

The Gabor characteristic that the every piece image of training image data centralization extracts after step S3Gabor wavelet filtering is: ${\left\{Y_i\right\}}_{i=1}^{90}\∈R^{25\×200}.$

Step S4, classification information is introduced on the basis of B2DLPP, dimensionality reduction is carried out to the higher-dimension facial image feature application SB2DLPP algorithm that step S3 extracts, thus obtains the low-dimensional characteristic coefficient of low n-dimensional subspace n left projection matrix U and right projection matrix V and training facial image.Detailed process is as follows:

A) because training image is 25 × 200 through the characteristics of image dimension that Gabor wavelet is extracted on Yale storehouse, this dimension is too high, so adopt SB2DLPP algorithm to carry out dimensionality reduction, is specially:

1. according to classification information structure weight matrix S.

Assuming that training image sample set can be divided into 15 classes, the image pattern point set of l class is designated as C _l, the image pattern point number that l class comprises is then by | C _l| represent.Again the image in training set being carried out sequence makes mutually of a sort data sequence adjacent.If Y _iwith sample point Y _jbelong to same class, be endowed the weighted value of a non-zero between two sample points, otherwise weight is set to 0.Weight function is decided to be:

Therefore, weight matrix S can be rewritten as:

Wherein, l _r(r=1,2 ..., 15) be the diagonal blocks of weight matrix S, it is a symmetrical submatrix (L _ij=L _ji).

2. low-dimensional maps, and solves optimum left projection matrix U and right projection matrix V.

The value of U and V solves generalized eigenvalue iteratively and proper vector obtains, and computing formula is:

$P^T(L\⊗{\mathrm{UU}}^T)\mathrm{Pv}=\mathrm{\γ}{P}^T(D\⊗{\mathrm{UU}}^T)\mathrm{Pv}$

$Q(L\⊗{\mathrm{VV}}^T)Q^{T}u=\mathrm{\λQ}(D\⊗{\mathrm{VV}}^T)Q^{T}u$

Wherein, D is a diagonal matrix (size is 90 × 90), the value on diagonal line be the row of weight matrix S or row sum and D _u=∑ _js _ji, matrix L=D-S (size is 90 × 90), for the Kronecker Product Operator of matrix, q=[Y ₁, Y ₂..., Y ₉₀].

Initialization condition is set to: U ₀=I ₂₅, the value of iterations gets 2.

Note column vector u ₁, u ₂..., u ₂₅25 minimal eigenvalue characteristic of correspondence vectors of first formula above, column vector v ₁, v ₂..., v ₂₅25 minimal eigenvalue characteristic of correspondence vectors of second formula above, then optimum left projection matrix U=(u ₁, u ₂..., u ₂₅) and the right projection matrix V=of optimum (v ₁, v ₂..., v ₂₅).

B) calculation training view data concentrates each width to train facial image Y _iproject to the low-dimensional characteristic coefficient M of low n-dimensional subspace n _i, its computing formula is: M _i=U ^ty _iv=(u ₁, u ₂..., u ₂₅) ^ty _i(v ₁, v ₂..., v ₂₅).

Step S5, carries out pre-service to all original facial images of test pattern data centralization that step S1 obtains.Through the pretreated test pattern data set of step S5 comprise 75 width sizes be 100 × 100 test pattern be designated as

Step S6, adopts Gabor wavelet to carry out feature extraction to the pretreated test facial image of step S5.The Gabor characteristic that the every piece image of test pattern data centralization extracts after step S6Gabor wavelet filtering is:

Step S7, the Gabor characteristic of the test pattern obtained by step S6 represents and projects to low n-dimensional subspace n thus obtain testing the low-dimensional characteristic coefficient of facial image.The Gabor characteristic of test pattern represents x _gproject to low n-dimensional subspace n thus obtain testing the low-dimensional characteristic coefficient T of facial image _gby formula: T _g=U ^tx _giv=(u ₁, u ₂..., u ₂₅) ^tx _g(v ₁, v ₂..., v ₂₅) obtain.

Step S8, in contrast step S7 test facial image to be identified low-dimensional characteristic coefficient and step S4 in train the low-dimensional characteristic coefficient of facial image, employing nearest neighbor classifier carries out Classification and Identification, exports final discrimination.Be specially:

A) the low-dimensional characteristic coefficient T of each width test pattern is calculated _g(g=1,2 ..., 75) with the low-dimensional characteristic coefficient M of all training images _i(i=1,2 ..., 90) between distance and be designated as dist _gi, distance computing formula is:

$\mathrm{dist}(T_g,M_i)={\mathrm{\Σ}}_{x=1}^{25}{\mathrm{\Σ}}_{y=1}^{47}[T_g^{(x,y)}-M_i^{(x,y)}]$

To each width test pattern T _gdistance set { the dist obtained _gi, i=1,2 ..., the element in 90} sorts from small to large, thus the low-dimensional characteristic coefficient obtaining training image corresponding to minor increment is M _i.If M _icorresponding training image A _iclassification be l, so T _gcorresponding test pattern B _gthen be identified l class.

B) the test face picture number correctly identified is added up.Correct identification test pattern number is designated as num, and initial value is 0.Assuming that test pattern B _g(g=1,2 ..., 75) and generic is e, if e=l, so num=num+1; Otherwise num value is constant.The end value of num is 68 in this embodiment.

C) export discrimination, discrimination computing formula is:

accuracy＝num/H×100％＝68/75×100％＝90.67％

By above-mentioned specific embodiment, the present invention is based on the face identification method of Gabor wavelet and SB2DLPP, good robustness can be had to the change such as illumination, expression, better discrimination can be obtained compared to additive method.

Above-mentioned explanation fully discloses the specific embodiment of the present invention.It is pointed out that the scope be familiar with person skilled in art and any change that the specific embodiment of the present invention is done all do not departed to claims of the present invention.Correspondingly, the scope of claim of the present invention is also not limited only to described embodiment.

Claims (6)

1. based on a face identification method of Gabor and SB2DLPP, it is characterized in that, described method comprises:

1) image in known face database is divided into two parts, a part is for building training image data set, and remainder is for building test pattern data set;

2) pre-service is carried out for all original facial images of training image data centralization;

3) Gabor wavelet is adopted to step 2) pretreated training facial image carries out feature extraction;

4) keep at bidirectional two-dimensional partial projection classification information is introduced on the basis of (B2DLPP) algorithm, to step 3) the higher-dimension facial image feature application that extracts has the bidirectional two-dimensional partial projection of supervision to keep (SB2DLPP) algorithm to carry out dimensionality reduction, thus obtains the left projection matrix U of low n-dimensional subspace n and the low-dimensional characteristic coefficient of right projection matrix V and training facial image;

5) to step 1) all original facial images of test pattern data centralization that obtain carry out pre-service;

6) to step 5) pretreated test facial image adopts Gabor wavelet to carry out feature extraction;

7) by step 6) Gabor characteristic of test pattern that obtains represents and projects to low n-dimensional subspace n thus the low-dimensional characteristic coefficient obtaining testing facial image;

8) contrast step 7) in the low-dimensional characteristic coefficient of test facial image to be identified and step 4) in train the low-dimensional characteristic coefficient of facial image, employing nearest neighbor classifier carries out Classification and Identification, exports final discrimination.

2. the face identification method based on Gabor and SB2DLPP according to claim 1, is characterized in that, described step 2) and step 5) comprise the steps:

A) carry out dimension normalization to original image, the size of every piece image adjusts to unified size m × n.

B) histogram equalization method is adopted to adjust the contrast of original facial image.

3. the face identification method based on Gabor and SB2DLPP according to claim 1, is characterized in that, described step 3) and step 6) comprise the steps:

A) facial image and Gabor wavelet bank of filters are carried out convolution.A given width facial image A _i, then image A _igabor characteristic obtained by following Convolution Formula:

Wherein, z=(x, y), * represents convolution, represent Gabor kernel function, G _{u, v}z () represents the convolved image of the Gabor kernel function when being of a size of v and direction is u.Here only have chosen 2 yardsticks (v ∈ { 0,1}), 8 groups of Gabor filter that 4 directions (u ∈ { 0,2,4,6}) combines are for extracting feature.

B) G is spliced by row _{u, v}z () obtains a Gabor characteristic matrix Y strengthened _i.Before splicing, in order to reduce space dimensionality, first can to G _{u, v}z () employing down-sampling factor is the down-sampling process of ρ.If represent the eigenmatrix after down-sampling, then the Gabor characteristic matrix strengthened be defined as:

$Y_i^{\mathrm{\ρ}}=\{G_{\mathrm{0,0}}^{\mathrm{\ρ}},G_{\mathrm{0,2}}^{\mathrm{\ρ}},G_{\mathrm{0,4}}^{\mathrm{\ρ}},G_{\mathrm{0,6}}^{\mathrm{\ρ}},G_{\mathrm{1,0}}^{\mathrm{\ρ}},G_{\mathrm{1,2}}^{\mathrm{\ρ}},G_{\mathrm{1,4}}^{\mathrm{\ρ}},G_{\mathrm{1,6}}^{\mathrm{\ρ}}\}$

4. the face identification method based on Gabor and SB2DLPP according to claim 1, is characterized in that, described step 4) comprise the steps:

A) supposition is through step 2) pretreated training image data set comprises the training image A that N number of size is m × n ₁, A ₂..., A _n, every piece image is through step 3) and the Gabor characteristic extracted after Gabor wavelet filtering is: because the characteristics of image dimension that Gabor wavelet is extracted is too high, so adopt SB2DLPP algorithm to carry out dimensionality reduction.The objective function of SB2DLPP is: $\underset{U,V}{\min }F(U,V)=\underset{U,V}{\min }\underset{i<j}{\mathrm{\&Sigma;}}{\left|\right|Y_i-Y_j\left|\right|}^2{s}_{\mathrm{ij}},$ Concrete steps are as follows:

1. according to classification information structure weight matrix S.

Assuming that training image sample set can be divided into r class, the image pattern point set of l class is designated as C _l, the image pattern point number that l class comprises is then by | C _l| represent.Again the image in training set being carried out sequence makes mutually of a sort data sequence adjacent.If Y _iwith sample point Y _jbelong to same class, be endowed the weighted value of a non-zero between two sample points, otherwise weighted value is set to 0.Weight function is defined as:

Or

Therefore, weight matrix S can be rewritten as:

Wherein, L _r(r=1,2 ..., c) be the diagonal blocks of weight matrix S, it is a symmetrical submatrix (L _ij=L _ji).

2. low-dimensional maps, and solves optimum projection matrix U and V.

Higher-dimension human face data is projected to the left projection matrix of optimum of low n-dimensional subspace n and right projection matrix by U and V, and its value solves generalized eigenvalue iteratively and proper vector obtains, and computing formula is:

$P^T(L\&CircleTimes;{\mathrm{UU}}^T)\mathrm{Pv}=\mathrm{\&gamma;}{P}^T(D\&CircleTimes;{\mathrm{UU}}^T)\mathrm{Pv}$

$Q(L\&CircleTimes;{\mathrm{VV}}^T)Q^{T}u=\mathrm{\&lambda;Q}(D\&CircleTimes;{\mathrm{VV}}^T)Q^{T}u$

Wherein, D is a diagonal matrix, the value on diagonal line be the row of weight matrix S or row sum and D _ji=∑ _js _ji, matrix L=D-S, for the Kronecker Product Operator of matrix, q=[Y ₁, Y ₂..., Y _n].Note column vector u ₁, u ₂..., u _ll minimal eigenvalue characteristic of correspondence vector of first formula above, column vector v ₁, v ₂..., v _rr minimal eigenvalue characteristic of correspondence vector of second formula above, then optimum left projection matrix U=(u ₁, u ₂..., u _l) and the right projection matrix V=of optimum (v ₁, v ₂..., v _r).

B) calculation training view data concentrates each width to train facial image Y _iproject to the low-dimensional characteristic coefficient M of low n-dimensional subspace n _i, its computing formula is: M _i=U ^ty _iv.

5. the face identification method based on Gabor and SB2DLPP according to claim 1, is characterized in that, described step 7) be specially: assuming that test pattern data set is through step 5) result that obtains after pre-service is designated as then through step 6) the test facial image Gabor characteristic extracted after Gabor wavelet filtering represents and is designated as the Gabor characteristic of test pattern represents X _gproject to low n-dimensional subspace n thus obtain testing the low-dimensional characteristic coefficient T of facial image _gby formula: T _g=U ^tx _gv obtains.

6. the face identification method based on Gabor and SB2DLPP according to claim 1, is characterized in that, described step 8) be specially: adopt nearest neighbor classifier to carry out Classification and Identification.First the low-dimensional characteristic coefficient T of each width test pattern is calculated _g(g=1,2 ..., H) with the low-dimensional characteristic coefficient M of all training images _i(i=1,2 ..., N) between distance and be designated as dist _gi; If secondly test sample book T _gwith the training sample M being under the jurisdiction of l class _inearest, so test sample book T _gbe identified as l class; If last test sample book T _goriginally just belong to l class, then identify correct, otherwise mistake, add up correct recognition rata simultaneously and export final discrimination.