CN102819731A - Face identification based on Gabor characteristics and Fisherface - Google Patents

Abstract

The invention relates to face identification based on Gabor characteristics and Fisherface. By a method utilizing a Gabor base for feature extraction, the dimension of the base obtained by Gabor filtering is too high; dimension reduction is carried out by the Fisherface method, and judgment analysis is carried out on a method for solving training average of each user so as to obtain a class corresponding to each user, thereby classifying. According to the face identification based on Gabor characteristics and Fisherface, by adopting the identification method, better identification efficiency can be achieved, and the identification rate can be up to 100% when environmental change is not large.

Description

Recognition of face based on Gabor characteristic and Fisherface

Technical field

The present invention relates to the field of recognition of face, especially a kind of recognition of face based on Gabor characteristic and Fisherface.

Background technology

Biological identification technology is meant the technology of differentiating through to characteristics of human body's digitized measurement, comprises that characteristics such as fingerprint, people's face, sound, iris, palmmprint can be used to carry out customer identification.Face recognition technology is an important topic in the biological identification technology, is unusual active research direction at present.With utilize the other biological characteristic to carry out identification to compare, recognition of face has directly, makes things convenient for, friendly, non-offensive advantage, thereby has extremely application prospects.

Though the mankind can identify people's face and even expression like a dream, the machine recognition of people's face is a great problem of difficulty.At first people's face is the irregular surface of the non-rigid body of a three-dimensional; Secondly, people's face can change along with the variation of age, health and expression; Once more, when gathering facial image, different illumination, angle all can influence recognition of face ground accuracy.Because human brain can not know still that to the mechanism of recognition of face the machine recognition of face also is in the stage of groping and innovating, and relates to many-sided many knowledge such as computer vision, pattern-recognition, physiology and psychology.The recognition of face that all of these factors taken together all is becomes and has challenge, but very has a problem of value.

A typical face identification system mainly comprises training process and identifying.Training process mainly accomplish with the known person face position, the design of feature extraction and selection and sorter; Identifying is then accomplished unknown picture is handled, and finally identifies the classification and the decision-making of identity.Its general structure is as shown in Figure 1, and as can be seen from the figure, its main functional modules comprises following several sections: 1. Image Acquisition: the facial image data source comprises motion image sequence (video flowing) and rest image.Mainly can pass through scanner, digital camera, the first-class digital input equipment of making a video recording obtains; 2. people's face detection and location: this module is used for analyzing the image of input, judges whether people's face is wherein arranged, if having, then finds out the position of people's face, and from background image, separates facial image; 3. image pre-service: pretreated main effect is to make as much as possible that facial image is in same yardstick and standard, finally for subsequent treatment high-quality input picture is provided.Usually this part need accomplish functions such as yardstick normalization to abstract image, gray scale normalization, noise reduction, the photograph that delusters, white balance; 4. feature extraction and selection: the facial image for after handling extracts the characteristic that is used to discern according to certain strategy, and original face spatial mappings is arrived new feature space.In this step, not only pay attention to how extracting characteristic with good separation performance, also must consider the application indexes such as robustness and treatment effeciency of total algorithm; 5. training: the i.e. design of sorter.This process mainly generates the parameter that can be used for discerning.Usually, on existing sample training collection basis, confirm certain decision rule, make by this rule being identified the minimum or outcome expectancy maximum of error recognition rate that object is classified and caused; 6. identification: accomplish the classification and the differentiation of people's face through comparing unknown human face parameter that obtains and the parameter of training gained, provide recognition result.

Summary of the invention

The technical matters that the present invention will solve is: in order to overcome the problem that exists in above-mentioned, improves a kind of based on recognition of face based on Gabor characteristic and Fisherface, its discrimination height and degree of accuracy height.

The technical solution adopted for the present invention to solve the technical problems is: a kind of recognition of face based on Gabor characteristic and Fisherface; Make the method for feature extraction through utilizing the Gabor base; The dimension of the base that obtains after the Gabor filtering is too high; Utilize the Fisherface method to carry out dimensionality reduction and ask the average method of training to carry out discriminatory analysis, thereby obtain the corresponding class of each user, thereby classify each user.

The concrete grammar of classification is following: the a.Gabor wavelet basis: be used in and be based on it and biological aspect correlativity in the recognition of face; The receptive field space structure of mammal visual cortex simple cell can be described with the Gabor function on mathematics; Simple cell is reactionless to large-area diffused light; And have the edge of light and shade contrast relatively the time and with detection, and the position at edge and orientation there is the selectivity of strictness.The Gabor wavelet basis also has some good character to be selected such as the locus, and direction is selected, and frequency is selected, and therefore orthogonality etc. are suitable in the feature extraction of image, particularly are used in the identification of people's face; The b.Gabor base is done feature extraction; The dimensionality reduction of c.Gabor eigenvector and prejudgementing criteria analysis; The d.Fisherface method: be a criterion that is in daily use, its principle is to differentiate through distributing between the class after type interior distribution normalization; E. based on the classifying rules of Gabor characteristic.

The invention has the beneficial effects as follows, the recognition of face based on Gabor characteristic and Fisherface of the present invention, it adopts this kind RM, can reach better recognition efficient, and when environmental change was little, discrimination can reach 100%.

Description of drawings

Below in conjunction with accompanying drawing and embodiment the present invention is further specified.

Fig. 1 is the general structure synoptic diagram of typical face identification system of the present invention;

Fig. 2 is the synoptic diagram of the mould of basic real part of Gabor of the present invention and Gabor base;

Fig. 3 is the synoptic diagram of the Gabor characteristic of facial image of the present invention.

Embodiment

Combine accompanying drawing that the present invention is done further detailed explanation now.These accompanying drawings are the synoptic diagram of simplification, basic structure of the present invention only is described in a schematic way, so it only show the formation relevant with the present invention.

A kind of recognition of face based on Gabor characteristic and Fisherface; Make the method for feature extraction through utilizing the Gabor base; The dimension of the base that obtains after the Gabor filtering is too high; Utilize the Fisherface method to carry out dimensionality reduction and ask the average method of training to carry out discriminatory analysis, thereby obtain the corresponding class of each user, thereby classify each user.

The concrete grammar of classification is following:

The a.Gabor wavelet basis

Gabor Wavelets also has some good character and selects such as the locus, and direction is selected, and frequency is selected, and therefore orthogonality etc. are suitable for particularly being used in the identification of people's face in the feature extraction of image.The kernel function of Gabor wavelet basis can be described with following formula:

${\mathrm{\ψ}}_{\mathrm{\μ},v}=\frac{{\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2}{{\mathrm{\σ}}^2}{e}^{-({\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2{\left|\right|z\left|\right|}^2/2{\mathrm{\σ}}^2)}[e^{{\mathrm{izk}}_{\mathrm{\μ},v}}-e^{-({\mathrm{\σ}}^2/2)}]---\left(1\right)$

μ wherein, v is respectively direction and scale factor, z=(x y) is row vector, and x, y are two-dimensional coordinate, $k_{\mathrm{\μ},v}=\left(\begin{array}{c}{k}_v\mathrm{Cos}{\mathrm{\φ}}_{\mathrm{\μ}}\\ k_v\mathrm{Sin}{\mathrm{\φ}}_{\mathrm{\μ}}\end{array}\right),$ k _v=k _Max/ f ^v, φ _μ=π μ/8.(1) interior first of square bracket are alternating components in the formula, and second is DC compensation, when parameter σ is very big, can ignore for second.

General is such to choosing of parameter, get 5 different yardstick v ∈ 0,1,2,3,4} and 8 direction μ ∈ 0 ..., 7}, and get σ=2 π, k _Max=pi/2, If x, the scope of y coordinate is (63,64), has obtained the Gabor base of one group of plural number like this, and is as shown in Figure 2, the real part of Fig. 2 (a) expression Gabor base, the mould of Fig. 2 (b) expression Gabor base

The b.Gabor base is done feature extraction

(the Gabor base that obtains with (1) formula carries out feature extraction to it for x, y) expression one width of cloth gray level image, and this process is exactly the computing of a convolution, that is: to suppose I

O _μ，v(z)=I(z)*ψ _μ，v(z) （2）

In the formula, ψ _{μ, v}(z) the different parameter μ of expression, the Gabor base that v is corresponding, O _{μ, ν}(z) expression is organized the Gabor characteristic that obtains after the filtering of Gabor base through this, and z=(x, y).

When carrying out convolution algorithm through (2) formula, continuation need be done in the image border.We can come calculated product through following formula:

Where

and representing the fast Fourier transform and inverse transform.For one 128 * 128 facial image, we do the Gabor feature extraction can obtain result as shown in Figure 3.

Fig. 3 (b) has comprised most of characteristic of people's face, and the processing of back for ease can be earlier to O _{μ, v}(z) delivery is done following processing then 1. to O _{μ, v}(z) falling sampling, is 128 * 128 the picture that obtains after the filtering of Gabor base to be fallen be sampled into 16 * 16 picture when we test.2. with O _{μ, v}(z) being standardized as average is 0, and variance is 1 matrix.3. with matrix O _{μ, v}(z) be connected between row and the row, form the one dimension long vector.Because μ ∈ 0 ..., 7}, { 0,1,2,3,4} so we do filtering with 40 groups of Gabor bases respectively for same width of cloth image, can obtain 40 stack features vector O through top processing to v ∈ then _0,0, O _0,1, O _4,7, these eigenvectors are coupled together we can obtain the Gabor eigenvector:

x=(O _0，0 ^TO _0，1 ^T.......O _4,7 ^T) ^T（4）

In the formula, T representing matrix transposition, x ∈ R ⁿ

The dimensionality reduction of c.Gabor eigenvector and prejudgementing criteria analysis

The Gabor eigenvector that obtains from above is in the space that high dimension arranged.Yet psychologic research shows that for example the sensory activity of identical category judgement generally is to represent to carry out down in the low dimension of sensorial data.The low dimension of things is represented for study, to be even more important, because when the needed dimension of expression bottom principle increased, the number of the example that needs in order to effective study can go up by index.And when considering the problem of calculated amount, low dimension expression also is crucial.From the demand to low dimension expression, we need a kind of effective dimensionality reduction algorithm.And if with square error as standard, pivot analysis (PCA) is exactly a kind of good dimensionality reduction method for expressing so, its main target is that the visual stimulus of higher-dimension (for example facial image) is projected a lower dimensional space.

PCA is the decorrelation skill of a standard, and it is used for trying to achieve the projection base of one group of quadrature, makes dimension reduce.Suppose

Be proper vector x ^(ρ)Covariance matrix, so for a random vector x ^(ρ)PCA promptly be covariance matrix ∑ with it _{X (ρ)}Be decomposed into following form:

∑ _x(ρ)=ΦΛΦ ^t

（5）

Φ=[φ ₁φ ₂...φ _N]Λ=diag{λ ₁,λ ₂,...,λ _N}

Wherein

is an orthogonal characteristic vector matrix,

be a diagonal angle eigenvalue matrix of falling continuous arrangement.

The critical nature of PCA is exactly, and when only characterizing original signal with a sub-set of pivot, from the angle of least mean-square error, PCA has optimum signal reconstruction ability.According to this character, the most direct application of PCA is exactly a dimensionality reduction:

y ^(ρ)＝P ^tx ^(ρ) （6）

P=[φ wherein ₁φ ₂... φ _m], m<n,

Low dimensional vector y ^(ρ)Comprised raw data x ^(ρ)In the most representative characteristic.

Yet we are to be appreciated that the advantage of PCA more is embodied on the irrelevance of data compression and low order statistics.PCA does not consider identification or classificatory characteristics, so the advantage of PCA can not embody at aspects such as recognitions of face well.In order to solve this shortcoming of PCA, we need a kind of better method, and it can either obtain raw data ground and hang down the dimension expression, has high discrimination again.One of solution just is to use the linear criterion of Fisher (FLD) to obtain the high discrimination between the pattern.The application of this method in recognition of face is to be realized by the Fisherface method.

The d.Fisherface method

FLD is a criterion that is in daily use, and its principle is to differentiate through distributing between the class after type interior distribution normalization.Suppose ω ₁, ω ₂..., ω _LAnd N ₁, N ₂..., N _LType of being respectively, and the picture number in each type.Make M again ₁, M ₂..., M _LWith M be the average of each type, and total average.Type in and type between the distribution matrix ∑ _ωAnd ∑ _bDefine by following formula:

${\mathrm{\&Sigma;}}_{\mathrm{\&omega;}}=\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)\mathrm{\&epsiv;}\left\{(y^{\left(\mathrm{\&rho;}\right)}-M_i){(y^{\left(\mathrm{\&rho;}\right)}-M_i)}^t\right|{\mathrm{\&omega;}}_i\}---\left(7\right)$

${\mathrm{\&Sigma;}}_b=\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)(M_i-M){(M_i-M)}^t---\left(8\right)$

P (ω wherein _i) be prior probability, ∑ _ω,

The sum of L type of being.

FLD obtains a projection matrix Ψ, makes | ψ _t∑ _bΨ |/| Ψ ^t∑ _ωΨ | ratio reach maximum.The condition that this ratio is obtained maximum value is that Ψ is made up of the proper vector of matrix , that is:

${\mathrm{\&Sigma;}}_{\mathrm{\&omega;}}^{-1}{\mathrm{\&Sigma;}}_b\mathrm{\&Psi;}=\mathrm{\&Psi;\&Delta;}---\left(9\right)$

Ψ wherein,

is respectively proper vector and the eigenvalue matrix of matrix

.

Certainly FLD also has its shortcoming, one of them be exactly that it needs a large amount of training samples just can better be summarized.If this condition can not be satisfied, FLD will seem very poor for the discriminating power of new test data.

E. based on the classifying rules of Gabor characteristic

After a width of cloth picture entering system, the method that we at first do to introduce in the feature extraction by the Gabor base obtains the Gabor proper vector of this picture; The characteristic y of low dimension ^(ρ)Then obtain by (6).And then to (low dimension) Gabor proper vector y ^(ρ)Carry out Fisherface.Definition K=1,2 .., L is certain type of ω _kThe mean value of the training sample that obtains behind the process Fisherface.GFC then uses similarity criterion δ so, seeks (with respect to average) overall arest neighbors as sorting technique.

$\mathrm{\&delta;}(u^{\left(\mathrm{\&rho;}\right)},M_k^0)=\underset{j}{\min }\mathrm{\&delta;}(u^{\left(\mathrm{\&rho;}\right)},M_j^0)\&RightArrow;u^{\left(\mathrm{\&rho;}\right)}\&Element;{\mathrm{\&omega;}}_k---\left(10\right)$

When using similarity criterion δ, the proper vector u of image ^(ρ), be considered to belong to that nearest type of range averaging value

The similarity method of discrimination that native system adopts is a cosine similarity criterion, is defined by following formula:

${\mathrm{\&delta;}}_{\cos }(x,y)=\frac{{-x}^{t}y}{\left|\right|x\left|\right|\left|\right|y\left|\right|}---\left(11\right)$

It should be noted that when using cosine similarity criterion that a negative sign is arranged on the molecule, this be because nearest neighbor method with minor increment, rather than ultimate range is as differentiation.

With above-mentioned foundation desirable embodiment of the present invention is enlightenment, and through above-mentioned description, the related work personnel can carry out various change and modification fully in the scope that does not depart from this invention technological thought.The technical scope of this invention is not limited to the content on the instructions, must confirm its technical scope according to the claim scope.

Claims (2)

1. recognition of face based on Gabor characteristic and Fisherface; It is characterized in that: make the method for feature extraction through utilizing the Gabor base; The dimension of the base that obtains after the Gabor filtering is too high; Utilize the Fisherface method to carry out dimensionality reduction and ask the average method of training to carry out discriminatory analysis, thereby obtain the corresponding class of each user, thereby classify each user.

2. the recognition of face based on Gabor characteristic and Fisherface according to claim 1 is characterized in that: the concrete grammar of classification is following: the a.Gabor wavelet basis; The b.Gabor base is done feature extraction; The dimensionality reduction of c.Gabor eigenvector and prejudgementing criteria analysis; The d.Fisherface method; E. based on the classifying rules of Gabor characteristic.

Images