CN101187986A - Face recognition method based on supervisory neighbour keeping inlaying and supporting vector machine - Google Patents

Abstract

The invention discloses a facial recognition method based on supervised neighborhood preserving embedding (SNPE) and a support vector machine (SVM). The method comprises a training process and a test process, and includes specifically the following procedures: a, a weight matrix for a given data sample group is constructed; b, according to the weight matrix obtained in the procedure a, a generalized characteristic vector problem about the data samples is solved and an embedding matrix projecting the data samples into a low-dimensional data space is searched; c, characteristic extraction is performed on the data samples with the embedding matrix to acquire characteristic data of the low-dimensional space; d, mode classification of the acquired characteristic data in the procedure c is conducted with the SVM to realize the type recognition of the data samples. The facial recognition method provided by the invention can solve the problems in the prior that the linear dimensionality reduction method cannot maintain well inter- and intra-type sample structures; the non-linear dimensionality reduction method has large computational amount; and the common classifier has over-learning and under-learning problems.

Description

Keep embedding face identification method with support vector machine based on the neighbour that supervision is arranged

Technical field

The present invention relates to a kind of face identification method, specifically, relate to a kind of face identification method that keeps embedding (SNPE) and support vector machine (SVM) based on the neighbour that supervision is arranged, belong to Flame Image Process and mode identification technology.

Background technology

Recognition of face is a pattern recognition problem, feature extraction is that the categorizing system in the pattern-recognition all needs an important step solving, its main target is to abandon irrelevant or less important information when obtaining optimum, notable feature, the dimension that reduces data is to lower the complicacy of categorizing system, and this just relates to so-called manifold learning problem.Technology classical in the manifold learning is linear dimension reduction method, and for example principal component analysis (PCA) (PCA) can realize flowing the linear of shape or approach linear embedding; When available category information, linear discriminant analysis method (LDA) can be sought the linear subspaces of an optimum to be used for classification.But these linear dimension reduction methods exist very big shortcoming: because the dimensionality reduction algorithm is linear transformation, so dimensionality reduction sample data behind the lower dimensional space might overlap, will produce bad influence for the classification in later stage.And the algorithm of most of linear dimensionality reductions can not keep in the class well and class between the composition of sample, influenced the effect of feature extraction.

In recent years, some Nonlinear Dimension Reduction algorithms were suggested.These algorithms mainly are the nonlinear organizations of finding in the stream shape, for example Laplce's characteristic matching method, local linear embed (LLE) and equidistant mapping (Isomap) etc.The Nonlinear Dimension Reduction method can keep the topological structure of sample, but calculated amount is big, and this method only is applicable to training sample, and how to make single test sample book dimensionality reduction is a difficult point always.Algorithm based on nuclear also is suggested, for example based on the principal component analysis (PCA) (KPCA) of nuclear, based on the linear discriminant analysis method (KLDA) etc. of nuclear.These algorithms can produce Nonlinear Mapping, but but do not consider the manifold structure of sample data, cause the dimensionality reduction effect not very good.

Traditional face identification method has nearest neighbor method, Euclidean distance method, mahalanobis distance method and neural network method etc.For its dimension, people's face sample is a small sample problem seldom.For this small sample problem of recognition of face, traditional sorting technique the study phenomenon occurred on the one hand easily, caused the algorithm generalization ability poor; On the other hand, traditional sorting technique learning performance is poor, can't be competent at people's face this non-linear very strong classification mode of classifying.SVM is for study and the classification that solves small sample problem proposes, it can overcome method such as neural network intrinsic mistake learn and owe problem concerning study, have very strong non-linear classification simultaneously again.

Summary of the invention

The invention provides and a kind ofly keep embedding face identification method with support vector machine, can solve linear dimension reduction method and can not keep in the class well and the composition of sample between class, Nonlinear Dimension Reduction method calculated amount is big and problem concerning study is learnt and owed to mistake that the general category device exists based on the neighbour that supervision is arranged.

For solving the problems of the technologies described above, the present invention is achieved by the following technical solutions:

A kind ofly keep embedding face identification method with support vector machine, may further comprise the steps based on the neighbour that supervision is arranged:

A, to weight matrix of given one group of data sample structure, be used for describing the relation between each data sample;

B, according to the weight matrix of step a gained, find the solution an extensive proper vector problem about data sample, search out the embedded matrix that data sample is mapped to the low-dimensional data space;

C, the embedded matrix that utilizes step b to obtain carry out feature extraction to the data sample, obtain the characteristic that data sample is mapped to lower dimensional space;

D, employing support vector machine are carried out pattern classification to the characteristic of the data sample that step c is obtained, and realize the type identification to the data sample.

Wherein, the structure weight matrix comprises following two steps among the described step a:

A1, utilize the sample number in known classification information and the class to determine K value, construct k nearest neighbor;

B1, calculate weight matrix: W representation value matrix, then W according to following process _IjBe i sample x _iTo j sample x _jWeights, when j sample do not belong to the k nearest neighbor of i sample, W _IjBe 0, by finding the solution the objective function that minimizes under the following constraint condition, calculate weight matrix W then:

$\min \underset{i}{\mathrm{\Σ}}{\left|\right|x_i-\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}{x}_j\left|\right|}^2,$

Constraint condition is

$\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}=1,j=\mathrm{1,2},\·\·\·,N.$

Further, the dimension of described embedded matrix lacks 1 than total classification number of described data sample.Described steps d can adopt two category support vector machines cascade models to realize the classification of many classification modes.

Further again, should carry out pre-service to described data sample earlier before the structure weight matrix.The preferred interpolation method of " cube convolution " that adopts carries out the convergent-divergent pre-service to described data sample.

Specifically, in face identification method of the present invention, specifically comprise training process and test process two parts, wherein,

Described training process specifically comprises the steps:

A2, at first the data sample is carried out pre-service, form training sample;

B2, to weight matrix of above-mentioned training sample structure, be used for describing the relation between each training sample, and utilize described weight matrix to find the solution the embedded matrix of training sample correspondence;

C2, the embedded matrix that utilizes step b2 to obtain carry out feature extraction to training sample, obtain the characteristic that training sample is mapped to lower dimensional space;

The characteristic of d2, the training sample that obtained based on step c2 is constructed two category support vector machines cascade models, sets kernel function type and penalty coefficient, carries out the supporting vector machine model training.

Wherein, the kernel function of described support vector machine is a kind of in linear kernel function, polynomial kernel function and the radially basic kernel function.The penalty coefficient of described support vector machine is set at 1-100.

Described test process specifically comprises the steps:

A3, at first the data sample is carried out pre-service, form test sample book;

B3, utilize the embedded matrix that obtains in the training process that test sample book is carried out feature extraction, obtain the characteristic that test sample book is mapped to lower dimensional space;

The characteristic of c3, test sample book that step b3 is obtained is input in the supporting vector machine model that has trained, carries out the type identification of test sample book.

Compared with prior art, advantage of the present invention and good effect are embodied in following two aspects:

1, adopted the SNPE algorithm in the feature extraction, overcome conventional linear dimension reduction methods such as PCA and be easy to generate the deficiency that sample data is overlapping, can not keep reaching in the class composition of sample between class well, and the big shortcoming of Nonlinear Dimension Reduction algorithm computation amount such as LLE, the SNPE algorithm has kept the local manifold structure of data set well, can accomplish optimum dimensionality reduction, reduce running time of algorithm again to a great extent.

2, adopt two classification SVM cascade models to realize the classification of many classification modes on the sorting algorithm, required support vector number is less, and in case add a new time-like, only need to get final product adding trained two classification SVM foremost, and needn't change or train again original SVM group, therefore, effectively reduced calculated amount.

Description of drawings

Fig. 1 is sample face (PCAfaces) behind the PCA dimensionality reduction and sample face (SNPEfaces) synoptic diagram behind the SNPE dimensionality reduction;

Fig. 2 realizes polytypic synoptic diagram by two classification SVM cascade models;

Fig. 3 is the groups of people's face image pattern in the ORL face database.

Embodiment

Below in conjunction with accompanying drawing embodiments of the present invention are described in further detail.

At first introduce basic design philosophy of the present invention:

1, in the feature extraction, given one group of data sample in the space is around at first constructed a weight matrix and is used for describing relation between the data sample.For each data sample point, represent that with the linear combination of its neighbour's data sample combination coefficient has just constituted weight matrix.Then, seek optimum embedding and make this neighbour's structure also can remain in the lower dimensional space, and utilize known classification information and the interior sample number of class to determine the K value.So just avoid traditional neighbour to keep embedding the first step in (NPE) algorithm, can guarantee the accuracy that K value is selected, accomplished the dimensionality reduction of optimum, can reduce running time of algorithm to a great extent again.

2, on the sorting algorithm, adopt two classification SVM cascade models to realize many classification mode classification.The advantage of this sorting algorithm is that required support vector number is less, and in case add a new time-like, only need to get final product adding trained two classification SVM foremost, and needn't change or train again original SVM group, thereby can reduce calculated amount.

Below the specific implementation method is described in detail.

Fig. 1 is the contrast synoptic diagram of the sample face (SNPEfaces) behind sample face (PCAfaces) and the SNPE dimensionality reduction behind the PCA dimensionality reduction.The PCA method is that the linear combination by feature comes dimensionality reduction, data projection to the low-dimensional linear subspaces.But facial image is not the linear combination of feature, by PCAfaces as can be seen, though the result after the feature extraction has kept facial contour, still has much noise to exist, and local linear structure does not keep, and some sample points have produced overlapping.

And SNPE is non-linear dimension reduction method, by SNPEfaces as can be seen, facial image noise after the feature extraction seldom (has been removed The noise such as illumination) effectively, and the low-dimensional manifold structure of sample is local linear, keep the unchangeability in stream shape field, be very suitable for machine recognition.

Referring to Fig. 2 and shown in Figure 3, be that sample specifies implementation step of the present invention with the ORL face database.

At first, from the ORL face database, read 40 classifications, five samples of every class totally 200 people's face sample data x ₁, x ₂..., x ₂₀₀, X ∈ R ^DSample data as training process.

To each width of cloth facial image x _iAfter the interpolation method of employing " cube convolution " carried out the convergent-divergent pre-service, each width of cloth facial image can be stacked into 32 * 32 dimensional vectors, and then generate 1024 * 1 dimensional vectors, and 200 sample datas can be formed training sample Xtrain _{1024 * 200}

Then to training sample Xtrain _{1024 * 200}Carry out the SNPE conversion:

At first, structure k nearest neighbor (KNN).

K nearest neighbor method: directly find out with respect to i sample point x _iK nearest neighbor sample point x _j, utilize the interior sample number of known classification information and class to determine the K value.Because chosen five samples in every class people's face, five sample numbers are contained in promptly every class the inside, so, choose K=4, directly find out with respect to i sample point x _iK nearest neighbor sample point x _j

Then, calculate weight matrix.

W representation value matrix, then W _IjBe i sample x _iTo j sample x _jWeights, when j sample do not belong to the k nearest neighbor of i sample, W _IjBe 0, by finding the solution the objective function that minimizes under the following constraint condition, can calculate weight matrix W then:

$\min \underset{i}{\mathrm{\Σ}}{\left|\right|x_i-\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}{x}_j\left|\right|}^2,$

Constraint condition is

$\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}=1,j=\mathrm{1,2},\·\·\·,200;$

Next step calculates embedded matrix.

Find the solution following extensive proper vector problem:

XMX ^Tα＝λXX ^Tα(*)；

Wherein,

X＝(x ₁，x ₂，…x _N)，M＝(I-W) ^T(I-W)，I＝diag(1，…，1)。

Make column vector α ₀, α ₁..., α _D-1For formula (*) corresponding to separating λ after the eigenwert ordering ₀≤ λ ₁≤ ... λ _D-1, then embedded matrix is:

x _i→y _i＝A ^Tx _i

A＝(α ₀，α ₁…α _d-1)

In the following formula, d is the dimension of embedded space.The value of d is very big to the dimensionality reduction influential effect, if d is too little, promptly the dimensionality reduction space dimensionality is too low, and data may overlap; If d is too big, will increase unnecessary noise during dimensionality reduction.The desirable dimension d of embedded space should lack 1 than total classification number, so, having 40 classifications in the present embodiment, dimension d is chosen as 39, and then the embedded matrix A of present embodiment is the matrix of 1024 * 39 dimensions.

Utilize above-mentioned embedded matrix of trying to achieve that training sample is carried out feature extraction, obtain the characteristic that training sample is mapped to lower dimensional space, i.e. Xtrain ' _{39 * 200}=A ^TXtrain.

At last, based on Xtrain ' _{39 * 200}Structure two classification SVM cascade models promptly adopt the classification mode training SVM model of " 1 pair 1 ".Be that the 40 class is carried out the design of first sorter SVM1 as two classes for example, second class and the 40 class are carried out the training of second sorter SVM2 as two classes the first kind and last class.Carry out above-mentioned classification based training process one by one, design 39 SVM sub-classifiers at last, all sub-classifiers have constituted the sorter in the present embodiment recognition of face.In the above-mentioned SVM model training process, the kernel function type is selected linear kernel function, polynomial kernel function and radially basic kernel function, penalty coefficient C=10 respectively.

After training process finishes, the test process that carries out recognition of face again.

From the ORL face database, read 40 classifications, five samples of every class totally 200 people's face sample datas as the sample data of test process.

According to the processing mode identical with training data, at first sample data is carried out pre-service, obtain the test sample book data; Utilize the embedded matrix that obtains in the training process that the test sample book data are carried out feature extraction then, obtain the test sample book data behind the dimensionality reduction; At last the test sample book data behind the dimensionality reduction are input in the svm classifier model that trains in the training process and carry out Classification and Identification.

If sample class increases, only need to add trained two classification SVM sub-classifiers foremost and get final product, and needn't change or train again original SVM sub-classifier at sorter, therefore can reduce calculated amount effectively.

Below each tabular gone out to select for use respectively different feature extracting methods and SVM to be incorporated into the experimental data of pedestrian's face identification.Wherein, the experimental data of table one for carrying out recognition of face, the experimental data of table two, the experimental data of table three for carrying out recognition of face based on SNPE and SVM for carrying out recognition of face based on LLE and SVM based on PCA and SVM.

Table one

The SVM type	The kernel function parameter	The reading images time (s)	The PCA time (s)	The SVM time (s)	Discrimination (%)
						LINEAR	Do not have	8.781	0.297	17.453	80.5
RBF	σ 2＝0.85	8.719	0.297	11.563	81
						POLY	q＝2	8.821	0.281	12.594	80

Table two

The SVM type	The kernel function parameter	The reading images time (s)	The LLE time (s)	The SVM time (s)	Discrimination (%)
						LINEAR	Do not have	8.875	1.344	34.531	82
RBF	σ 2＝0.85	8.922	1.344	13.469	82
						POLY	q＝3	8.922	1.344	12.438	78.5

Table three

The SVM type	The kernel function parameter	The reading images time (s)	The SNPE time (s)	The SVM time (s)	Discrimination (%)
						LINEAR	Do not have	8.875	0.547	6.047	79
RBF	σ 2＝0.85	8.813	0.531	6.688	82.5
						POLY	q＝2	7.938	0.453	9.922	100

Respectively show COMPARISON OF CALCULATED RESULTS WITH EXPERIMENTAL DATA as can be seen from top: the face identification method working time based on SNPE and SVM is short, discrimination is higher, can reach 100% discrimination when selecting the polynomial kernel function.

Should be understood that; above-mentioned explanation is not to be limitation of the present invention; the present invention also is not limited in above-mentioned giving an example, and modification, distortion, interpolation or replacement that those skilled in the art are made in essential scope of the present invention also should belong to protection scope of the present invention.

Claims (10)

1. one kind keeps embedding face identification method with support vector machine based on the neighbour that supervision is arranged, and comprises training process and test process, it is characterized in that, may further comprise the steps:

A, to weight matrix of given one group of data sample structure, be used for describing the relation between each data sample;

B, according to the weight matrix of step a gained, find the solution an extensive proper vector problem about data sample, search out the embedded matrix that data sample is mapped to the low-dimensional data space;

C, the embedded matrix that utilizes step b to obtain carry out feature extraction to the data sample, obtain the characteristic that data sample is mapped to lower dimensional space;

D, employing support vector machine are carried out pattern classification to the characteristic of the data sample that step c is obtained, and realize the type identification to the data sample.

2. face identification method according to claim 1 is characterized in that, in described step a, the structure weight matrix comprises following two steps:

A1, utilize the sample number in known classification information and the class to determine K value, construct k nearest neighbor;

B1, calculate weight matrix: W representation value matrix, then W according to following process _IjBe i sample x _iTo j sample x _jWeights, when j sample do not belong to the k nearest neighbor of i sample, W _IjBe 0, by finding the solution the objective function that minimizes under the following constraint condition, calculate weight matrix W then:

$\min \underset{i}{\mathrm{\Σ}}{\left|\right|x_i-\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}{x}_j\left|\right|}^2,$

Constraint condition is

$\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}=1,j=\mathrm{1,2},\·\·\·,N.$

3. face identification method according to claim 1 is characterized in that, the dimension of described embedded matrix lacks 1 than total classification number of described data sample.

4. face identification method according to claim 1 is characterized in that, described steps d adopts two category support vector machines cascade models to realize the classification of many classification modes.

5. according to each described face identification method among the claim 1-4, it is characterized in that, earlier described data sample is carried out pre-service before the structure weight matrix.

6. face identification method according to claim 5 is characterized in that, adopts the interpolation method of " cube convolution " that described data sample is carried out the convergent-divergent pre-service.

7. face identification method according to claim 1 is characterized in that described training process comprises the steps:

A2, at first the data sample is carried out pre-service, form training sample;

B2, to weight matrix of above-mentioned training sample structure, be used for describing the relation between each training sample, and utilize described weight matrix to find the solution the embedded matrix of training sample correspondence;

C2, the embedded matrix that utilizes step b2 to obtain carry out feature extraction to training sample, obtain the characteristic that training sample is mapped to lower dimensional space;

The characteristic of d2, the training sample that obtained based on step c2 is constructed two category support vector machines cascade models, sets kernel function type and penalty coefficient, carries out the supporting vector machine model training.

8. face identification method according to claim 7 is characterized in that, the kernel function of described support vector machine is a kind of in linear kernel function, polynomial kernel function and the radially basic kernel function.

9. face identification method according to claim 7 is characterized in that the penalty coefficient of described support vector machine is set at 1-100.

10. according to each described face identification method among the claim 7-9, it is characterized in that described test process comprises the steps:

A3, at first the data sample is carried out pre-service, form test sample book;

B3, utilize the embedded matrix that obtains in the training process that test sample book is carried out feature extraction, obtain the characteristic that test sample book is mapped to lower dimensional space;

The characteristic of c3, test sample book that step b3 is obtained is input in the supporting vector machine model that has trained, carries out the type identification of test sample book.