CN103839033A - Face identification method based on fuzzy rule - Google Patents

Abstract

The invention discloses a face identification method based on a fuzzy rule. The face identification method is a novel feature extraction method and belongs to the field of face identification. The method comprises the steps that a face image is divided; sub-image variance and information entropy are calculated as the output of a fuzzy controller, so as to acquire the fuzzy weight of a corresponding sub-image; and LBP extraction is carried out on the sub-image to acquire a histogram vector and PCA dimensionality reduction is carried out on a sub-image vector after LBP extraction. The fuzzy weight has a significant feature for the image identification rate. Through PCA dimensionality reduction, the computation time of an algorithm is saved, and the timeliness of the algorithm is improved.

Description

A kind of face identification method based on fuzzy rule

Technical field

The invention belongs to recognition of face field, be specifically related to a kind of face identification method based on fuzzy rule.

Background technology

Recognition of face, refers in particular to utilize and analyzes the computer technology that relatively face visual signature information is carried out identity discriminating.It relates to the subjects knowledge such as pattern-recognition, image processing, computer vision, physiology, psychology, is current research focus.The actual a series of correlation techniques that build face identification system that comprise of recognition of face of broad sense, comprise that man face image acquiring, face location, recognition of face pre-service, identity validation and identity search etc.; And the recognition of face of narrow sense is refered in particular to by face and is carried out technology or the system that identity validation or identity are searched.System input is generally one or a series of facial image not determining one's identity that contains, and the facial image of some known identities in face database or coding accordingly, its output is a series of similarity scores, shows the identity of face to be identified.

As everyone knows, a commercial value has widely been created in recognition of face, such as crime identification, security system, credit card validation scene monitoring etc.Compared with the identification of other biological feature, recognition of face is because it is untouchable and use friendly to receive remarkable concern.Up to now, a lot of face identification methods are suggested, and these algorithms are broadly divided into two classes, the recognizer based on geometric properties and the recognizer based on template.Method based on geometric properties is extracted significant local feature (for example eyes, nose and face) exactly and their geometric relationship is classified, and comprises elastic graph matching process etc.Method core concept based on template is exactly eigenface, and eigenface is that all eigenwerts of view picture facial image and proper vector draw.The basis of the face identification method based on eigenface is that KL changes, and this is the optimum orthogonal transformation of one in compression of images.Under normal circumstances, the transformation matrix of KL conversion has the eigenvector of scatter matrix between training sample class to generate.Projection on the eigenvector of facial image above has compared with macro-energy, is called principal component; Projection on eigenvector below has less energy, is called component of degree n n.In the time giving up part component of degree n n, KL converts also referred to as principal component analysis (PCA) (PCA).Unfortunately PCA is due to its gray-scale statistical value based on image, and the difference that the image difference that external factor is brought and face itself bring cannot be distinguished.Therefore for illumination, can reduce its discrimination.

Face identification method based on PCA is very long one section main flow that becomes research period, and a lot of face identification methods all more or less have with it relation, such as linear discriminant analysis method (LDA), fisherface method and Bayes's recognition of face method etc. thereafter.Recent years, support vector machine (SVM), neural network and kernel method are also widely used in this field.

1996, Ojalad etc. have proposed local binary pattern (LBP), LBP operator is the texture measure in a kind of tonal range, it is derived from a kind of texture local neighbor definition, at first for the local contrast of complementary ground dimensioned plan picture proposes, people were expanded the application of LBP afterwards, were applied to recognition of face field.

Why LBP can be by extensive concern, is that it has following major advantage:

1, tagsort ability is strong.LBP has any dull unchangeability and image rotation unchangeability, be it for expression attitude etc., variation has compared with strong robustness, and it can also effectively describe some local tiny characteristic, such as bright spot, dim spot, grain details etc., and reflect their distribution situation.

2,, without training, there is good generalization.The opportunity statistics of main flow or the face identification method of learning strategy need training data for face modeling, therefore existing problems aspect popularization.And LBP method extraction histogram vectors does not need to train, thereby not there are not the problems referred to above.

Although LBP has above-mentioned advantage, from the angle of Classification and Identification, it has following deficiency: due in the time extracting LBP histogram, facial image need be divided, there will be a problem in the middle of partition process, divide scale problem.Divided block number too much can cause histogram vectors dimension excessive, is that computation complexity improves like this, and piecemeal words very little lose again statistical significance.The problem being derived by partition problem is that LBP is to put on an equal footing to zoning, but this does not meet general knowledge, and people generally believe that as marking areas such as eyes, nose, faces identification is had to larger contribution, therefore will strengthen its importance to these regions.

Summary of the invention

In order to solve LBP algorithm to the putting on an equal footing and high-dimensional problem of each region, the present invention proposes a kind of tax with zones of different weight and finally use method---a kind of face identification method based on fuzzy rule of PCA dimensionality reduction.This method is from the different of existing face identification method maximum: used fuzzy controller to obtain weight, average pixel variance and information entropy, as the input of fuzzy controller, are output as weight.When group graph region has larger help to identification, obtain weight and compose with this region by calculating, in the time that it is very little for identification impact, just composing with less weight.By being that the tax of different subgraphs region is too high with the proper vector possibility dimension of different weights serial connection, we adopt PCA to carry out dimensionality reduction to it like this, to improve recognition speed.

Technical scheme of the present invention is:

Based on a face identification method for fuzzy rule, comprise the steps:

1) gather facial image and carry out pre-service; Pre-service comprises the processing such as face detection, denoising and size normalization;

2) pretreated image is carried out to image division, image is divided into k ²individual subgraph; Such as 3 × 3 or 4 × 4 etc.;

3) the LBP histogram vectors of extraction subgraph, LBP histogram vectors is as the textural characteristics of subgraph;

4) calculate respectively the information entropy E of subgraph _j, subgraph pixel variances sigma _j;

5) fuzzy controller that adopts dual input list to export; The wherein information entropy E of subgraph _j, subgraph pixel variances sigma _jas input, weight w _jas output;

6) the textural characteristics subarea vector of subgraph is carried out to PCA dimensionality reduction, obtain PCA dimensionality reduction vector;

7) carry out recognition of face according to the textural characteristics of each subgraph and the weights of fuzzy controller output.

Described step 3) extract the LBP histogram vectors of subgraph, the statistics of the pixel i of subgraph m employing following formula calculates,

$H_i^m=\underset{x,y}{\mathrm{\Σ}}g\{f(x,y)=i\},i=0,...,n-1$

In formula, m represents subgraph, and n, i represent gray level,

Described step 4) calculate the information entropy E of subgraph _jfirst adopt formula

calculate the statistical probability of subgraph j pixel i

wherein k ²for subgraph number, f (x, y) is image pixel; Adopt again formula

calculate the information entropy E of subgraph j _j;

Calculate subgraph pixel variances sigma _jfirst adopt

calculate the mean pixel μ of subgraph j _j, wherein a represents that subgraph is long, and b represents that subgraph is wide, and f (x, y) is image pixel; Adopt again formula ${{\mathrm{\σ}}^2}_j=\frac{1}{\mathrm{ab}}\underset{x=1}{\overset{a}{\mathrm{\Σ}}}\underset{y=1}{\overset{b}{\mathrm{\Σ}}}{(f(x,y)-{\mathrm{\μ}}_j)}^2$ Calculate subgraph j pixel variances sigma _j.

Described step 5) the fuzzy rule rice triangular membership functions of fuzzy controller, the wherein information entropy E of subgraph _j, subgraph pixel variances sigma _jas input, weight w _jas output; To the information entropy E of subgraph _j, subgraph pixel variances sigma _jbe divided into little (S), in (M), large (B), obtain nine fuzzy rules, the i rule of fuzzy system is expressed as follows:

Rule?i：Ifσ _jis?A _1，i?and?E _j?is?A _2，i?Then?w _j?is?B _i。

Wherein A _{1, i}and A _{2, i}represent respectively σ _jand E _jthe degree of membership value of getting, B _irepresent the weighted value obtaining, A _{1, i}, A _{2, i}and B _ithe scope of winning the confidence is large, medium and small.

Described step 7) adopt sorting algorithm, by formula

calculate; Wherein k ²to divide subgraph number, w _jrespective regions weight, N _pCAthat major component is selected number, S _{i, j}and M _{i, j}the statistics of region j pixel i, the number that pixel i occurs at region j, wherein S _{i, j}represent sample, M _{i, j}represent template, D (S, M) represents histogram intersection distance.

Described step 3) and step 4) order replacement, described step 5) and step 6) order replacement.

The present invention proposes one and can respectively divide part according to facial image and give different weights and strengthen knowing method for distinguishing for identification importance is different, the method has stronger robustness and discrimination, can apply to better recognition of face.

Brief description of the drawings

Fig. 1 is common LBP feature extraction face recognition algorithms.

Fig. 2 is face recognition process.

Fig. 3 is method flow diagram of the present invention.

Fig. 4 is design of Fuzzy Controller and rule thereof.

Embodiment

According to the document of existing LBP feature extraction algorithm, we can show that the control block diagram of LBP feature extraction algorithm is as Fig. 1.Existing LBP feature extraction algorithm is a kind of experience weight assignment method as seen from Figure 1, very not objective.In order to strengthen the robustness of LBP feature extraction and to be more conducive to identification, redesign for experience weight assignment in LBP feature extraction algorithm, adopt fuzzy if-then rules device to obtain weight, allow algorithm have more objectivity and ubiquity, even if also can obtain discrimination accurately in the time there is no correlation experience.

For a better understanding of the present invention, below in conjunction with accompanying drawing, algorithm of the present invention is elaborated: the present embodiment is being contemplated that under prerequisite and is implementing with the present invention, providing detailed embodiment and specific operation process.Face recognition process frame diagram, as Fig. 2, first carries out image acquisition, and then face detects, if end of identification do not detected, otherwise, further carry out face location, then pretreated image is carried out to feature extraction and be beneficial to recognition of face.

The present invention is under windows7 environment, adopt this software of matlab7.11 to test, the data that experiment adopts are ORL facial image databases, and ORL has 40 people, 400 width images, adopt 200 images as training sample, image size is 112 pixel × 92 pixels.

The FB(flow block) of this method is as Fig. 3, and the step of its running is as follows:

1, image pre-service

Image pre-service refers to the processing of carrying out before image is carried out feature extraction, cuts apart and mated, and its fundamental purpose is information irrelevant in removal of images, and replying Useful Information and enhancing has the detectability of individual information and reduced data to greatest extent.Therefore, we will carry out a series of processing to gathering image, finally obtain standard picture.Due to this experiment adopt be ORL image library, be standard picture, therefore, need not carry out pre-service.

2, feature extraction

According to existing feature extraction algorithm, feature extraction is mainly divided into five steps, and idiographic flow is as follows:

21) divide facial image

Carry out piecemeal to processing image, piecemeal is according to k ²divide.While extracting histogram vectors, need to carry out piecemeal to image, but piecemeal can cause histogram vectors dimension excessive too much, cause computation complexity to improve, piecemeal can lose again the meaning of statistics very little.Not only can get rid of some redundancy components and effectively extract the component that is conducive to classification in histogram vectors, reduce and disturb, and can greatly reduce dimension, improve matching speed, be divided into 3 × 3 or 4 × 4 etc.

22) each subgraph is carried out to LBP computing, extract region histogram vector

LBP is a kind of texture description operator, and it has stronger classification capacity, and has unchangeability for dull grey scale change.LBP operator class is similar to the template operation in filtering.Progressive scanning picture, for each pixel in image, using the gray scale of this point as threshold value, it is carried out to binaryzation in 3 × 38 fields around, in a certain order by 8 bits of result composition of binaryzation, the response using the value of this binary number as this point.Finally adopt the histogram of response image as its texture description.Image after cutting apart is carried out to LBP computing, add up the histogram data of each several part, so just increased the dimension of proper vector, improved recognition correct rate simultaneously.

Here the histogram of subimage response image can define like this:

$H_i^m=\underset{x,y}{\mathrm{\Σ}}g\{f(x,y)=i\},i=0,...,n-1$

Wherein, m represents subgraph, and i represents gray level

23) each subimage is selected to corresponding weights

231) first ask for each subgraph information entropy and average pixel variance

In fact, we carry out recognition of face and just identify for some outstanding features, for example, and eyes, nose, face etc.That is to say that different people appearance co-located sub-block has different similarities.We can directly compose weights 1 to subgraph in the ordinary course of things, but in the present invention, because being utilizes face prominent feature to identify, therefore, those have the subgraph of outstanding contributions to carry out larger weights assignment to it to identification.We can be according to expertise, and antithetical phrase graph region is composed with different weights, and this is a kind of more subjective weight assignment, also can select information entropy or the average pixel variance in corresponding subgraph region, comparatively speaking, so more objective.

Determine weights by information entropy, first ask for the statistical probability of subgraph j pixel i, calculate by following formula

$P_j^i=\frac{\underset{x,y\∈M_j}{\mathrm{\Σ}}T\{f(x,y)=i\}}{\underset{j=1}{\overset{k^2}{\mathrm{\Σ}}}\underset{x,y\∈M_j}{\mathrm{\Σ}}T\{f(x,y)=i\}}$

Then obtain subgraph j information entropy E by following formula _j,

$E_j=-\underset{j=1}{\overset{k^2}{\mathrm{\Σ}}}{P}_j^i\log P_j^i;$

Finally by entropy E _jdetermine weight w _j,

$w_j=\frac{E_j}{\underset{j=1}{\overset{k^2}{\mathrm{\Σ}}}{E}_j}.$

Hence one can see that, entropy E _jlarger, more image informations are contained in expression region, more easily to human face recognition, and therefore weight w _jlarger.

Determine weight by average pixel variance; Determine weight by average pixel variance, its basic thought is: the average pixel that first calculates corresponding sub-block region:

${\mathrm{\μ}}_j=\frac{1}{\mathrm{ab}}\underset{x=1}{\overset{a}{\mathrm{\Σ}}}\underset{y=1}{\overset{b}{\mathrm{\Σ}}}f(x,y)$

Then calculate its average pixel variances sigma _j,

${{\mathrm{\σ}}^2}_j=\frac{1}{\mathrm{ab}}\underset{x=1}{\overset{a}{\mathrm{\Σ}}}\underset{y=1}{\overset{b}{\mathrm{\Σ}}}{(f(x,y)-{\mathrm{\μ}}_j)}^2$

Finally by formula

determine weight.

So far, we know average pixel variances sigma _jlarger, corresponding subgraph similarity is less, required tax with weight w _jalso less.

232) design fuzzy controller

Arrive here, what consider is all that Dan Yin usually determines weights, in order to make result more objective, counts a fuzzy controller, as Fig. 4, selects information entropy E _jwith average pixel variances sigma _jbe used as the input of fuzzy controller, weight w _jas output.By entropy and average pixel variance be divided into little (S), in (M), large (B), structure fuzzy rule, known have 9 fuzzy rules, chooses suitable weight w from fuzzy rule base.

3, subgraph image is carried out to PCA dimensionality reduction

Principal component analysis (PCA) in face identification system (PCA) can reduce the dimension of original feature vector, can save like this computing time and improve recognition rate because noise contribution can be filtered, a feasible way is exactly in the different subareas of extracting histogram vectors through LBP by the method for PCA, in this situation, weight w _jstill can be applied on distance matrix.The method thinks that any secondary subgraph all can be decomposed into the linear combination of a series of vectors and coefficient, and this class coefficient is incoherent each other, just can well represent a sub-picture and computation complexity is reduced as long as choose some coefficients that energy is larger, be conducive to like this identification.PCA dimensionality reduction concrete steps are the proper vectors that calculate covariance matrix, estimate raw data by the linear combination of estimating eigenvalue of maximum.Taking recognition of face as example, calculating process can be expressed as follows:

Take out M facial image as training set, X _ibe the image vector of i training sample, calculating M width subgraph mean value is u: $u=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{X}_i;$

Utilize every width facial image data to deduct average face data I _i=X _i-u, covariance matrix is

$C=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}(X_i-u){(X_i-u)}^T=\frac{1}{M}{\mathrm{II}}^T$

Wherein I=[X ₁-u, X ₂-u ..., X _m-u];

Thereby calculate proper vector eigenvalue λ _k.Proper vector has determined the linear space of the eigenface of M width different images ψ formation: $x_i=\underset{k=1}{\overset{d}{\mathrm{\Σ}}}{x}_{\mathrm{ik}}{\mathrm{\ψ}}_k,$ i＝1，2，...，M

In these eigenface, K (K < M) individual eigenface and K maximum eigenwert match.Facial image can, by the subspace of the face of videoing with down conversion, project to eigenface: w _{n, k}=x _k(u _n-u), and wherein n=1,2 ..., M, k=1,2 ..., K.

Wherein projection coefficient is described the vectorial X of the facial image of input by each eigenface _n=[w _n1, w _n2..., w _nk] obtain, and a series of base vectors of facial image are exactly eigenface.

Because the single width intrinsic dimensionality of the 0RL image adopting reaches 1 × 10304 dimension, its training image storehouse is 200 × 10304 dimension matrixes, although carried out LBP feature extraction, but its dimension or excessive, thereby while carrying out pattern-recognition, data volume computing is large, selects PCA method to reduce in right amount the intrinsic dimensionality of image in experiment.

4, Classification and Identification

The classification of use sorting algorithm, sorting algorithm adopts City Block Measure, and formula is as follows:

$D(S,M)=\underset{j=1}{\overset{k^2}{\mathrm{\&Sigma;}}}{w}_j\left(\underset{i=1}{\overset{N_{\mathrm{PCA}}}{\mathrm{\&Sigma;}}}\right|S_{i,j}-M_{i,j}\left|\right)$

Wherein k ²to divide sub-block number, w _jrespective regions weight, N _pCAthat major component is selected number, S _{i, j}and M _{i, j}the statistics of region j pixel i, the number that pixel i occurs at region j.

Compared with prior art, the present invention has following beneficial effect: the method that the present invention adopts modified LBP and PCA to merge is extracted characteristics of image.LBP method have anti-selective power by force, the advantage such as different impact of illumination while not gathered image, and follow-on LBP method in the present invention, by different subgraphs being arranged to different weights, has improved image recognition rate, and obtaining of weights adopted fuzzy rule.Come after texture feature extraction by LBP, its dimension is also larger, then carrys out dimensionality reduction by PCA.Experimental result shows, the feature extraction that mixes LBP and PCA can improve discrimination and the recognition speed of face.

Experiment content

The data acquisition of this experiment test ORL face database, this is current most widely used recognition of face database, its result comparison is more intense.Be characterized in without detecting the processing such as face and size normalization.ORL database has everyone 10 width different images, chooses that in ORL database, everyone 5 width images of 50 people are as training sample, and all the other 5 width figure of this 50 people are as test sample book.LBP operator is chosen the LBP that 8 field radiuses are 2 _8,2.For the advantage of outstanding this method, carry out series of experiments, comprise the experiment contrasts such as LBP feature extracting method, single characteristic weighing LBP feature extraction, FUZZY WEIGHTED LBP feature extraction, FUZZY WEIGHTED LBP+PCA feature extraction, experimental result shows to adopt the method for FUZZY WEIGHTED LBP feature extraction to be obviously better than the method for weighting that adopts separately information entropy or average pixel variance to obtain, and is obviously better than not having the method for PCA dimensionality reduction by the method for PCA dimensionality reduction in recognition speed.Use this method can obtain better discrimination and recognition speed.

Claims (7)

1. the face identification method based on fuzzy rule, is characterized in that comprising the steps:

1) gather facial image and carry out pre-service;

2) pretreated image is carried out to image division, image is divided into k ²individual subgraph;

3) the LBP histogram vectors of extraction subgraph, LBP histogram vectors is as the textural characteristics of subgraph;

4) calculate respectively the information entropy E of subgraph _j, subgraph pixel variances sigma _j;

5) fuzzy controller that adopts dual input list to export; The wherein information entropy E of subgraph _j, subgraph pixel variances sigma _jas input, weight w _jas output;

6) the textural characteristics subarea vector of subgraph is carried out to PCA dimensionality reduction, obtain PCA dimensionality reduction vector;

7) carry out recognition of face according to the textural characteristics of each subgraph and the weights of fuzzy controller output.

2. the face identification method based on fuzzy rule according to claim 1, is characterized in that step 3) extract the LBP histogram vectors of subgraph, the statistics H of the pixel i of subgraph j _{i, j}, H _{i, j}employing following formula calculates,

$H_{i,j}=\underset{x,y}{\mathrm{\&Sigma;}}g\{f(x,y)=i\},i=0,...,n-1$

In formula, j represents subgraph, and n, i represent gray level,

3. the face identification method based on fuzzy rule according to claim 1, is characterized in that described step 4) calculate the information entropy E of subgraph _jfirst adopt formula calculate the statistical probability of subgraph j pixel i

wherein k ²for subgraph number, f (x, y) is image pixel; Adopt again formula

calculate the information entropy E of subgraph j _j;

Calculate subgraph pixel variances sigma _jfirst adopt

calculate the mean pixel μ of subgraph j _j, wherein a represents that subgraph is long, and b represents that subgraph is wide, and f (x, y) is image pixel; Adopt again formula ${{\mathrm{\&sigma;}}^2}_j=\frac{1}{\mathrm{ab}}\underset{x=1}{\overset{a}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{b}{\mathrm{\&Sigma;}}}{(f(x,y)-{\mathrm{\&mu;}}_j)}^2$ Calculate subgraph j pixel variances sigma _j.

4. the face identification method based on fuzzy rule according to claim 1, is characterized in that described step 5) the fuzzy rule of fuzzy controller adopt triangular membership functions, wherein the information entropy E of subgraph _j, subgraph pixel variances sigma _jas input, weight w _jas output; To the information entropy E of subgraph _j, subgraph pixel variances sigma _jbe divided into little (S), in (M), large (B), obtain nine fuzzy rules, the i rule of fuzzy system is expressed as follows:

Rule?i：If?σ _j?is?A _1，iand?E _j?is?A _2，i?Then?w _j?is?B _i。

Wherein A _{1, i}and A _{2, i}represent respectively σ _jand E _jthe degree of membership value of getting, B _irepresent the weighted value obtaining, A _{1, i}, A _{2, i}and B _ispan is large, medium and small.

5. the face identification method based on fuzzy rule according to claim 1, is characterized in that described step 7) adopt sorting algorithm, by formula

calculate; Wherein k ²to divide subgraph number, w _jrespective regions weight, N _pCAthat major component is selected number, S _{i, j}and M _{i, j}the statistics of region j pixel i, the number that pixel i occurs at region j, wherein S _{i, j}represent sample, M _{i, j}represent template, D (S, M) represents histogram intersection distance.

6. the face identification method based on fuzzy rule according to claim 1, is characterized in that described step 3) with step 4) order replaces.

7. the face identification method based on fuzzy rule according to claim 1, is characterized in that described step 5) and step 6) order replaces.

Images