CN105069447A - Facial expression identification method - Google Patents

Abstract

The invention discloses a facial expression identification method and relates to a method for identifying images. The facial expression identification method is a facial expression identification method which extracts expression textural characteristics using a center symmetrical ternary patterns (CSTP) algorithm. The method comprises steps of preprocessing a facial expression image; extracting facial expression textural characteristics on each sub-block of the facial expression image; and determining final facial expression textural characteristics of the facial expression image. Therefore, the facial expression identification is completed. The method overcomes problems that texture description is not fine due to a complex identification background and the identification rate is not high in the prior art.

Description

A kind of recognition methods of human face expression

Technical field

Technical scheme of the present invention relates to the method for identifying figure, specifically a kind of recognition methods of human face expression.

Background technology

Expression recognition is computer vision field challenging research topic, significant in psychology, man-machine interaction research etc.In recent years, expression recognition technical development is rapid, and the method for main flow has Gabor filter method, active shape model ASM (ActiveShapeModels), principal component analysis (PCA) PCA (PrincipalComponentAnalysis) and textural characteristics method.Gabor wavelet can extract the textural characteristics of human face expression from different scale and different directions, but the high dimension that its computation process produces likely causes exhausting of calculator memory, and computation process is quite consuming time.ASM algorithm can react the change of expression intuitively, but the structure of model is quite complicated, have impact on the automaticity of algorithm to a certain extent.PCA algorithm can extract the overall textural characteristics of facial expression image, but it is difficult to avoid the too high problem of covariance matrix dimension that produces when processing data.Binary principal component analysis (PCA) 2DPCA is the expandable algorithm of PCA, to some extent solves the problems referred to above, but it is the feature extracting image in the horizontal direction, and have ignored the feature in vertical direction.Local binary patterns LBP operator is proposed by people such as Ojala the earliest, it can rendering image local grain information effectively, and LBP and extension feature thereof are applied to human face expression and detect, and have outstanding performance, its shortcoming is when illumination variation is violent, and the reliability of LBP can decline.For the limitation of LBP, centralization binary pattern (CBP) is applied to Expression Recognition by the people such as Fu Xiaofeng, CBP feature adopts " diagonal angle principle ", namely the gray-scale value of diagonal element is calculated, central point pixel is joined in traditional LBP operator and goes, and give central point pixel highest weight value.But it is only utilize neighboring pixel brightness change information that CBP and LBP exists equally, and does not consider the zonal local message of human face expression, cause the defect that the discrimination of expression is not high.

In a word, existing facial expression recognizing method exists because identifying the texture description that background complexity causes not meticulous, the defect that discrimination is not high.

Summary of the invention

Technical matters to be solved by this invention is: the recognition methods providing a kind of human face expression, a kind of use Central Symmetry three binarization mode (CenterSymmetricalTernaryPatterns, hereinafter referred to as CSTP) algorithm extracts the recognition methods of human face expression of expression textural characteristics, overcome the complicated texture description caused of prior art identification background not meticulous, the defect that discrimination is not high.

The present invention solves this technical problem adopted technical scheme: a kind of recognition methods of human face expression, and be a kind of recognition methods using CSTP algorithm to extract the human face expression of human face expression textural characteristics, concrete steps are as follows:

The first step, Facial Expression Image pre-service:

Adopt the Facial Expression Image that carried out in the face expression database of Face datection and geometrical normalization, and on former database, apply following formula (1) carry out gaussian filtering process, complete Facial Expression Image pre-service thus,

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{x^2+y^2}{2{\mathrm{\σ}}^2}}---\left(1\right),$

Wherein (x, y) is pixel coordinate, and σ is variance;

Second step, extracts the human face expression textural characteristics in the sub-block of each Facial Expression Image:

Carry out piecemeal to the Facial Expression Image that first step pre-service completes, be divided into 5 × 5 non-overlapped sub-blocks, utilize CSTP algorithm to extract human face expression textural characteristics in each Facial Expression Image sub-block, step is as follows;

(1) to calculate (P, R) neighborhood in about the pixel of center pixel symmetry differing from Δ g _i:

Δg _i＝g _i-g _i+P/2,i＝0,1,…,P/2-1(2)，

Wherein, g is pixel, and P is surrounding neighbors number, and R is radius, R=2;

(2) statistics is somebody's turn to do the pixel of (P, R) neighborhood to the average of difference

$\stackrel{\‾}{{\mathrm{\Δg}}_i}=\left(\underset{i=0}{\overset{P/2-1}{\Σ}}{\mathrm{\Δg}}_i\right)/(P/2)---\left(3\right);$

(3) capping value U and lower limit L:

$U=\stackrel{\‾}{{\mathrm{\Δg}}_i}*(1+t)---\left(4\right),$

$L=\stackrel{\‾}{{\mathrm{\Δg}}_i}*(1-t)---\left(5\right),$

Wherein, t is threshold parameter, is empirical value, t ∈ (0,1), and t adjusts with the change of illumination;

(4) the CSTP human face expression textural characteristics of each pixel is extracted:

${\mathrm{CSTP}}_{P,R}=\underset{i=0}{\overset{P/2-1}{\Σ}}{2}^i\·S({\mathrm{\Δg}}_i,U,L)---\left(6\right),$

Wherein, $S({\mathrm{\&Delta;g}}_i,U,L)=\left\{\begin{array}{c}1,{\mathrm{\&Delta;g}}_i>U\\ 0,L\&le;{\mathrm{\&Delta;g}}_i\&le;U\\ -1,{\mathrm{\&Delta;g}}_i<L\end{array}\right.---\left(7\right);$

The cataloged procedure of CSTP is, from the upper left corner clockwise, as the gray-scale value g of surrounding pixel point _iwith the g taking center pixel as symmetry _i+P/2difference DELTA g _ibe greater than U, then this position is encoded to 1, as Δ g _ibe less than L, then this position is encoded to-1; Be clearly " 0 " all "-1 " in coding, be called holotype, coding in " 1 " be clearly " 0 ", then "-1 " is set to " 1 ", be called negative mode, after the coding completing positive and negative two patterns, align negative mode carry out corresponding inclusive-OR operation, each pixel in each sub-block carries out the feature extraction of CSTP, namely extract the CSTP human face expression textural characteristics of each pixel, the CSTP histogram H adding up each sub-block is exactly the human face expression textural characteristics in each Facial Expression Image sub-block of extracting;

3rd step, determine the human face expression textural characteristics that Facial Expression Image is final:

(1) the information entropy E of each the Facial Expression Image sub-block in second step is calculated _j:

$E_j=-\underset{i=0}{\overset{m-1}{\&Sigma;}}{P}_j^i{\mathrm{logP}}_j^i---\left(8\right),$

Wherein, represent the probability that a jth sub-block i-th grade of pixel occurs, m represents progression, gets 256;

(2) the weights W of each Facial Expression Image sub-block is calculated _j:

$W_j=E_j/\underset{j=0}{\overset{n-1}{\&Sigma;}}{E}_j---\left(9\right),$

Wherein, n is each Facial Expression Image sub-block number;

(3) histogram vectors after each sub-block weighting of each Facial Expression Image is calculated

Wherein H _jit is the histogram of each the Facial Expression Image jth sub-block utilizing second step to extract;

To connect the histogram vectors after each sub-block weighting of each Facial Expression Image as final human face expression textural characteristics;

4th step, completes the identification of human face expression:

Adopt SVMKernelMethods tool box training SVM classifier, idiographic flow is as follows:

(1) the human face expression textural characteristics that training sample and test sample book are extracted is inputted, this human face expression textural characteristics is through gained after second step above and the 3rd step, go out the training and testing classification sample matrix of training sample human face expression textural characteristics matrix and test sample book human face expression textural characteristics matrix difference correspondence according to these human face expression textural characteristics sample architecture, the value in classification sample matrix is the class categories of sample;

(2) for local facial express one's feelings textural characteristics adopt gaussian kernel function, check figure is set to 8, Lagrange factor c=100, regulating parameter λ=10 of Quadratic Optimum method ^-7, first training sample human face expression textural characteristics matrix and training classification sample matrix are sent into svmmulticlassoneagainstall function and obtain support vector, weight is with biased, again by training sample textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is trained, finally by test sample book textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is predicted, complete the identification of human face expression thus, in JAFFE storehouse, experiment obtains anger, detest, fear, glad, neutral, sad and surprised 7 kinds of expressions, and experiment obtains anger in CMU-AMP storehouse, glad and surprised 3 kinds of expressions, complete the identification of human face expression thus.

Above-mentioned facial expression recognizing method, described SVM classifier is known.

The invention has the beneficial effects as follows: compared with prior art, outstanding substantive distinguishing features of the present invention and marked improvement as follows:

(1) the inventive method has broken away from the dependence to center pixel, overcome the complicated texture description caused of prior art identification background not meticulous, the defect that recognition performance is low, meticulousr to local texture description, inclusive-OR operation enhances the important information of expression, makes the sign of whole human face expression textural feature space sample and classification performance be further enhanced and improve;

(2) the inventive method by calculate the information entropy of each sub-block construct each sub-block adaptive weighted after proper vector, each block feature vector after adaptive weighted is concatenated as the final human face expression textural characteristics of Facial Expression Image, merge local feature and global characteristics, improve the performance of Expression Recognition.

Accompanying drawing explanation

Below in conjunction with drawings and Examples, the present invention is further described.

Fig. 1 is the schematic process flow diagram of the inventive method.

Fig. 2 (a) is the Facial Expression Image after human face expression pre-service in JAFFE storehouse.

Fig. 2 (b) is the Facial Expression Image after human face expression pre-service in CMU-AMP storehouse.

Fig. 3 (a) is the three-valued schematic diagram of CSTP.

Fig. 3 (b) is CSTP inclusive-OR operation schematic diagram.

Fig. 4 (a) is that in the inventive method, in JAFFE storehouse, CSTP gets the discrimination schematic diagram under different t value.

Fig. 4 (b) is that in the inventive method, in CMU-AMP storehouse, CSTP gets the discrimination schematic diagram under different t value.

Embodiment

Embodiment illustrated in fig. 1ly show, the flow process of the recognition methods of human face expression of the present invention is: the identification of the final human face expression textural characteristics of Facial Expression Image pre-service → the extract human face expression textural characteristics in the sub-block of each Facial Expression Image → determine Facial Expression Image → complete human face expression.

In more detail, the textural characteristics in the sub-block of each Facial Expression Image of said extracted comprises: to calculate in (P, R) neighborhood about the pixel of center pixel symmetry difference Δ g _i→ statistics is somebody's turn to do the pixel of (P, R) neighborhood to the average of difference the CSTP textural characteristics of each pixel of → capping value U and lower limit L → extract; Above-mentionedly determine that the final textural characteristics of Facial Expression Image comprises: the information entropy E calculating each the Facial Expression Image sub-block in previous step _j→ calculate the weights W of each Facial Expression Image sub-block _j→ calculate the histogram vectors after each sub-block weighting of each Facial Expression Image

Fig. 2 is the pretreated schematic diagram of Facial Expression Image in the inventive method, wherein: Fig. 2 (a) illustrated embodiment shows the Facial Expression Image after human face expression pre-service in the methods of the invention in JAFFE storehouse, be respectively angry, detest, fear, glad, neutral, sad and surprised 7 kinds of expressions.Fig. 2 (b) illustrated embodiment shows the Facial Expression Image in the pretreated CMU-AMP storehouse of human face expression in the methods of the invention, is surprised, angry and glad 3 kinds of expressions respectively.

Explanation embodiment illustrated in fig. 3 in the methods of the invention in sub-block each pixel carry out CSTP texture feature extraction and be defined as follows:

${\mathrm{CSTP}}_{P,R}=\underset{i=0}{\overset{(P/2)-1}{\&Sigma;}}{2}^i\&CenterDot;S({\mathrm{\&Delta;g}}_i,U,L)---\left(11\right),$

Wherein, $S({\mathrm{\&Delta;g}}_i,U,L)=\left\{\begin{array}{c}1,{\mathrm{\&Delta;g}}_i>U\\ 0,L\&le;{\mathrm{\&Delta;g}}_i\&le;U\\ -1,{\mathrm{\&Delta;g}}_i<L\end{array}\right.---\left(12\right),$

Fig. 3 (a) is the three-valued schematic diagram of CSTP, wherein from left to right, have nine blockages in this figure first image subblock, the numerical value in each blockage represents the gray-scale value of each pixel, and two pixels in first image subblock in same pattern circle represent a pixel pair, totally 4 pixels pair, be respectively (251,178), (135,90), (53,23) and (246,198).To calculate in this neighborhood about the pixel of center pixel symmetry difference Δ g from upper left corner clockwise direction _i, its numerical value, as shown in four blockages containing numerical value in this figure second image subblock, is 73,45,30,48 respectively viewed from the clockwise direction of the upper left corner; Then the pixel of four blockages containing numerical value in second image subblock is added up to the average of difference utilize obtain higher limit U and lower limit L with threshold value t, finally judge 3 kinds of patterns (-1,0,1) according to formula (12), its result such as 4 blockages containing numerical value in this figure the 3rd image subblock are depicted as 1,0 ,-1,0.

Fig. 3 (b) is CSTP or computing schematic diagram, intuitively by containing 4 blockages of numerical value in Fig. 3 (a) the 3rd image subblock by becoming a column vector form clockwise, splitting into positive and negative two patterns and being weighted respectively;-1 part above in a line column vector coding replaces by 0, and be defined as holotype, binary coding is 1000, and it is 8 that weighting is transformed into the decimal system; 1 part below in a line column vector coding replaces by 0, and-1 takes absolute value, and is defined as negative mode, and binary coding is 0010, and it is 2 that weighting is transformed into the decimal system; Finally carry out inclusive-OR operation, binary coding is 1010, and it is 10 that weighting is transformed into the decimal system; Inclusive-OR operation strengthens expression information, prevents the loss of its some important informations in characteristic extraction procedure.

The relation of threshold value t and human face expression average recognition rate in CSTP in Fig. 4 (a) illustrated embodiment display JAFFE database, the figure illustrates the corresponding relation of the average recognition rate of threshold value t value and human face expression in CSTP different in JAFFE database, when in the CSTP in JAFFE database, threshold value t gets 0.5 and 0.7, effect is best, and the discrimination of human face expression can reach 95.71%.

Fig. 4 (b) illustrated embodiment shows the relation of threshold value t and human face expression average recognition rate in CSTP in CMU-AMP database, the figure illustrates the corresponding relation of the average recognition rate of threshold value t value and human face expression in CSTP different in CMU-AMP database, when in the CSTP in CMU-AMP database, threshold value t gets 0.9, the average recognition rate of human face expression is the highest.

The experiment basis of Fig. 4 (a) and Fig. 4 (b) illustrated embodiment is that training sample and test sample book are divided into 5 × 5 sub-blocks, and P=8, R=2, SVM check figure is set to 8.

Embodiment

The recognition methods of a kind of human face expression of the present embodiment, be a kind of recognition methods using CSTP algorithm to extract the human face expression of human face expression textural characteristics, concrete steps are as follows:

The first step, Facial Expression Image pre-service:

Adopt the Facial Expression Image that carried out in the face expression database of Face datection and geometrical normalization, and on former database, apply following formula (1) carry out gaussian filtering process, complete Facial Expression Image pre-service thus,

$G(x,y)=\frac{1}{2{\mathrm{\&pi;\&sigma;}}^2}{e}^{-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2}}---\left(1\right),$

Wherein (x, y) is pixel coordinate, and σ is variance;

Second step, extracts the human face expression textural characteristics in the sub-block of each Facial Expression Image:

Carry out piecemeal to the Facial Expression Image that first step pre-service completes, be divided into 5 × 5 non-overlapped sub-blocks, utilize CSTP algorithm to extract human face expression textural characteristics in each Facial Expression Image sub-block, step is as follows;

(1) to calculate (P, R) neighborhood in about the pixel of center pixel symmetry differing from Δ g _i:

Δg _i＝g _i-g _i+P/2,i＝0,1,…,P/2-1(2)，

Wherein, g is pixel, and P is surrounding neighbors number, and R is radius, R=2;

(2) statistics is somebody's turn to do the pixel of (P, R) neighborhood to the average of difference

$\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}=\left(\underset{i=0}{\overset{P/2-1}{\&Sigma;}}{\mathrm{\&Delta;g}}_i\right)/(P/2)---\left(3\right);$

(3) capping value U and lower limit L:

$U=\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}*(1+t)---\left(4\right),$

$L=\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}*(1-t)---\left(5\right),$

Wherein, t is threshold parameter, is empirical value, t ∈ (0,1), and t adjusts with the change of illumination;

(4) the CSTP human face expression textural characteristics of each pixel is extracted:

${\mathrm{CSTP}}_{P,R}=\underset{i=0}{\overset{P/2-1}{\&Sigma;}}{2}^i\&CenterDot;S({\mathrm{\&Delta;g}}_i,U,L)---\left(6\right),$

Wherein, $S({\mathrm{\&Delta;g}}_i,U,L)=\left\{\begin{array}{c}1,{\mathrm{\&Delta;g}}_i>U\\ 0,L\&le;{\mathrm{\&Delta;g}}_i\&le;U\\ -1,{\mathrm{\&Delta;g}}_i<L\end{array}\right.---\left(7\right);$

The cataloged procedure of CSTP is, from the upper left corner clockwise, as the gray-scale value g of surrounding pixel point _iwith the g taking center pixel as symmetry _i+P/2difference DELTA g _ibe greater than U, then this position is encoded to 1, as Δ g _ibe less than L, then this position is encoded to-1; Be clearly " 0 " all "-1 " in coding, be called holotype, coding in " 1 " be clearly " 0 ", then "-1 " is set to " 1 ", be called negative mode, after the coding completing positive and negative two patterns, align negative mode carry out corresponding inclusive-OR operation, each pixel in each sub-block carries out the feature extraction of CSTP, namely extract the CSTP human face expression textural characteristics of each pixel, the CSTP histogram H adding up each sub-block is exactly the human face expression textural characteristics in each Facial Expression Image sub-block of extracting;

3rd step, determine the human face expression textural characteristics that Facial Expression Image is final:

(1) the information entropy E of each the Facial Expression Image sub-block in second step is calculated _j:

$E_j=-\underset{i=0}{\overset{m-1}{\&Sigma;}}{P}_j^i{\mathrm{logP}}_j^i---\left(8\right),$

Wherein, represent the probability that a jth sub-block i-th grade of pixel occurs, m represents progression, gets 256;

(2) the weights W of each Facial Expression Image sub-block is calculated _j:

$W_j=E_j/\underset{j=0}{\overset{n-1}{\&Sigma;}}{E}_j---\left(9\right),$

Wherein, n is each Facial Expression Image sub-block number;

(3) histogram vectors after each sub-block weighting of each Facial Expression Image is calculated

Wherein H _jit is the histogram of each the Facial Expression Image jth sub-block utilizing second step to extract;

To connect the histogram vectors after each sub-block weighting of each Facial Expression Image as final human face expression textural characteristics;

4th step, completes the identification of human face expression:

Adopt SVMKernelMethods tool box training SVM classifier, idiographic flow is as follows:

(1) the human face expression textural characteristics that training sample and test sample book are extracted is inputted, this human face expression textural characteristics is through gained after second step above and the 3rd step, go out the training and testing classification sample matrix of training sample human face expression textural characteristics matrix and test sample book human face expression textural characteristics matrix difference correspondence according to these human face expression textural characteristics sample architecture, the value in classification sample matrix is the class categories of sample;

(2) for local facial express one's feelings textural characteristics adopt gaussian kernel function, check figure is set to 8, Lagrange factor c=100, regulating parameter λ=10 of Quadratic Optimum method ^-7, first training sample human face expression textural characteristics matrix and training classification sample matrix are sent into svmmulticlassoneagainstall function and obtain support vector, weight is with biased, again by training sample textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is trained, finally by test sample book textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is predicted, complete the identification of human face expression thus, last experiment in JAFFE storehouse obtains anger, detest, fear, glad, neutral, sad and surprised 7 kinds of expressions, and experiment obtains surprised in CMU-AMP storehouse, angry and glad 3 kinds of expressions, complete the identification of human face expression thus.

The present embodiment uses the Gaussian function of SVM classifier as classification function, carries out multicategory classification.

The present embodiment is tested at JAFFE and CMU-AMP Facial expression database.Wherein JAFFE expresses one's feelings in storehouse and has 10 people, and everyone has 7 kinds of expressions, namely angry, detest, fear, glad, neutral, sad and surprised, often kind of express one's feelings 3 width images, totally 210 width images.In experiment, everyone often plants expression and chooses 2 as training sample, totally 140, remaining 70 as test sample book, the size of image is normalized to 128 × 128.CMU-AMP expression has 13 people in storehouse, and everyone is containing surprised, angry, happiness these 3 kinds expressions, and often kind of expression chooses 16 width images, totally 624 width images.In experiment, everyone often plants expression random selecting 8 as training sample, and totally 312 training samples, choose everyone 8 as test sample book in residual image, and totally 312 as test sample book, and the size of image is normalized to 64 × 64.The present embodiment is that the MATLABR2013b platform under Windows8 environment has run.Fig. 2 is 7 kinds of expressive parts samples in the JAFFE face expression database of the present embodiment and 3 kinds of expressive parts samples in CMU-AMP face expression database.The experiment basis of embodiment is that training sample in two expression storehouses and test sample book are divided into 5 × 5 sub-blocks, and P=8, R=2, SVM check figure is set to 8.

Table 1 lists the discrimination (%) of the human face expression on the JAFFE database of the present embodiment.Table 2 lists the Expression Recognition rate (%) on the CMU-AMP database of the present embodiment.

The discrimination (%) of the human face expression on table 1.JAFFE database

Expression Recognition rate (%) on table 2.CMU-AMP database

The data of consolidated statement 1 and table 2 are known, the average recognition rate of the inventive method obviously will exceed these four kinds of classic algorithm of 2DPCA, Gabor+PCA, LBP and CBP, in JAFFE database, fear that this Expression Recognition rate is lower than Gabor+PCA and CBP, for fear with detest and surprised expression similarity high, CSTP feature similarity, be easy to fear be classified as and detest and surprised, this is also the following problem needing solution; The recognition methods of extracting the human face expression of expression textural characteristics in CMU-AMP database with CSTP algorithm is better than other several algorithms in anger and these two kinds of human face expressions of happiness and average recognition rate, surprised this human face expression is lower than 2DPCA, reason is that in CMU-AMP database, human face expression classification is very few, the reason that training sample is too much, illustrates that the recognition methods of the human face expression of the present invention CSTP algorithm extraction expression textural characteristics has superiority on process small sample problem simultaneously.

Claims (1)

1. a recognition methods for human face expression, is characterized in that: be a kind of recognition methods using CSTP algorithm to extract the human face expression of human face expression textural characteristics, concrete steps are as follows:

The first step, Facial Expression Image pre-service:

Adopt the Facial Expression Image that carried out in the face expression database of Face datection and geometrical normalization, and on former database, apply following formula (1) carry out gaussian filtering process, complete Facial Expression Image pre-service thus,

$G(x,y)=\frac{1}{2{\mathrm{\&pi;\&sigma;}}^2}{e}^{-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2}}---\left(1\right),$

Wherein (x, y) is pixel coordinate, and σ is variance;

Second step, extracts the human face expression textural characteristics in the sub-block of each Facial Expression Image:

Carry out piecemeal to the Facial Expression Image that first step pre-service completes, be divided into 5 × 5 non-overlapped sub-blocks, utilize CSTP algorithm to extract human face expression textural characteristics in each Facial Expression Image sub-block, step is as follows;

(1) to calculate (P, R) neighborhood in about the pixel of center pixel symmetry differing from Δ g _i:

Δg _i＝g _i-g _i+P/2,i＝0,1,…,P/2-1(2)，

Wherein, g is pixel, and P is surrounding neighbors number, and R is radius, R=2;

(2) statistics is somebody's turn to do the pixel of (P, R) neighborhood to the average of difference

$\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}=\left(\underset{i=0}{\overset{P/2-1}{\&Sigma;}}{\mathrm{\&Delta;g}}_i\right)/(P/2)---\left(3\right);$

(3) capping value U and lower limit L:

$U=\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}*(1+t)---\left(4\right),$

$L=\stackrel{\&OverBar;}{{\mathrm{\&Delta;g}}_i}*(1-t)---\left(5\right),$

Wherein, t is threshold parameter, is empirical value, t ∈ (0,1), and t adjusts with the change of illumination;

(4) the CSTP human face expression textural characteristics of each pixel is extracted:

${\mathrm{CSTP}}_{P,R}=\underset{i=0}{\overset{P/2-1}{\&Sigma;}}{2}^i\&CenterDot;S({\mathrm{\&Delta;g}}_i,U,L)---\left(6\right),$

Wherein, $S({\mathrm{\&Delta;g}}_i,U,L)=\left\{\begin{array}{c}1,{\mathrm{\&Delta;g}}_i>U\\ 0,L\&le;{\mathrm{\&Delta;g}}_i\&le;U\\ -1,{\mathrm{\&Delta;g}}_i<L\end{array}\right.---\left(7\right);$

The cataloged procedure of CSTP is, from the upper left corner clockwise, as the gray-scale value g of surrounding pixel point _iwith the g taking center pixel as symmetry _i+P/2difference DELTA g _ibe greater than U, then this position is encoded to 1, as Δ g _ibe less than L, then this position is encoded to-1; Be clearly " 0 " all "-1 " in coding, be called holotype, coding in " 1 " be clearly " 0 ", then "-1 " is set to " 1 ", be called negative mode, after the coding completing positive and negative two patterns, align negative mode carry out corresponding inclusive-OR operation, each pixel in each sub-block carries out the feature extraction of CSTP, namely extract the CSTP human face expression textural characteristics of each pixel, the CSTP histogram H adding up each sub-block is exactly the human face expression textural characteristics in each Facial Expression Image sub-block of extracting;

3rd step, determine the human face expression textural characteristics that Facial Expression Image is final:

(1) the information entropy E of each the Facial Expression Image sub-block in second step is calculated _j:

$E_j=-\underset{i=0}{\overset{m-1}{\&Sigma;}}{P}_j^i{\mathrm{logP}}_j^i---\left(8\right),$

Wherein, represent the probability that a jth sub-block i-th grade of pixel occurs, m represents progression, gets 256;

(2) the weights W of each Facial Expression Image sub-block is calculated _j:

$W_j=E_j/\underset{j=0}{\overset{n-1}{\&Sigma;}}{E}_j---\left(9\right),$

Wherein, n is each Facial Expression Image sub-block number;

(3) histogram vectors after each sub-block weighting of each Facial Expression Image is calculated

Wherein H _jit is the histogram of each the Facial Expression Image jth sub-block utilizing second step to extract;

To connect the histogram vectors after each sub-block weighting of each Facial Expression Image as final human face expression textural characteristics;

4th step, completes the identification of human face expression:

Adopt SVMKernelMethods tool box training SVM classifier, idiographic flow is as follows:

(1) the human face expression textural characteristics that training sample and test sample book are extracted is inputted, this human face expression textural characteristics is through gained after second step above and the 3rd step, go out the training and testing classification sample matrix of training sample human face expression textural characteristics matrix and test sample book human face expression textural characteristics matrix difference correspondence according to these human face expression textural characteristics sample architecture, the value in classification sample matrix is the class categories of sample;

(2) for local facial express one's feelings textural characteristics adopt gaussian kernel function, check figure is set to 8, Lagrange factor c=100, regulating parameter λ=10 of Quadratic Optimum method ^-7, first training sample human face expression textural characteristics matrix and training classification sample matrix are sent into svmmulticlassoneagainstall function and obtain support vector, weight is with biased, again by training sample textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is trained, finally by test sample book textural characteristics matrix, support vector, weight, biased and Gaussian function is sent in svmmultival function and is predicted, complete the identification of human face expression thus, in JAFFE storehouse, experiment obtains anger, detest, fear, glad, neutral, sad and surprised 7 kinds of expressions, and experiment obtains anger in CMU-AMP storehouse, glad and surprised 3 kinds of expressions, complete the identification of human face expression thus.