CN103984919A - Facial expression recognition method based on rough set and mixed features - Google Patents

Abstract

The invention discloses a facial expression recognition method based on a rough set and mixed features. The method includes the following steps that (1) facial detection is carried out; (2) local geometric distortion features are extracted by means of a method combining an active appearance model and the rough set; (3) overall features of expressions are extracted with the combination of an improved weighting primary component analysis method and the rough set; (4) feature fusion is carried out on the extracted local geometric distortion features and the overall features by means of kernel canonical correlation analysis under a high-dimensional small sample to eliminate feature redundancy, and fused typical features are obtained; (5) the fused typical features serve as observation vectors of a discrete HMM for classification and recognition, and a classification result is obtained. As is presented by experiments, the improved method can shorten facial expression recognition time and improve the facial expression recognition rate.

Description

Based on the facial expression recognizing method of rough set and composite character

[technical field]

The invention belongs to technical field of computer information processing, particularly facial expression recognizing method.

[background technology]

Human face expression identification (Facial Expression Recognition, FER) be the process that computing machine carries out feature extraction and sorts out the expression information of face, it makes to calculate function knows people's expression information, and then deduction people's psychological condition, thereby it is mutual to realize high-grade intelligent between man-machine, is that of the multidisciplinary intersections such as pattern-recognition, physiology, psychology, computer vision is rich in challenging problem.

Although human face expression recognition technology is studied by a lot of people, but existing method does not disperse to emphasize expression shape change is had the information in the regions such as the people face eyes, eyebrow, face of significant contribution, and single feature extracting method can not be contained all effective informations, the multiplicity of not considering the importance degree of attributive character simultaneously, infosystem being contributed, cause thus the discrimination of human face expression identification not high, the long problem of recognition time of Expression Recognition.

[summary of the invention]

The object of the invention is to provide a kind of facial expression recognizing method, not high in order to solve the discrimination of human face expression of the prior art identification, the long problem of recognition time of Expression Recognition.

For achieving the above object, implementing facial expression recognizing method of the present invention comprises the steps:

Carry out face detection;

Adopt the method that active apparent model and rough set combine to extract the local geometric deformation characteristics of expressing one's feelings;

Adopt improved weighted principal component analyzing method and rough set to combine and extract the global feature of expression;

Adopt the kernel canonical correlation analysis under higher-dimension small sample to carry out Fusion Features elimination feature redundancy to the local geometric deformation characteristics of extracting and global feature, obtain the characteristic feature after merging;

Feature after utilization is merged, as the observed value vector of Discrete HMM, is carried out Classification and Identification, obtains classification results.

According to above-mentioned principal character, adopt the method that active apparent model and rough set combine to extract in the step of local geometric deformation characteristics of expression, the geometric description that these organ shape such as the eyebrow that adopts AAM algorithm to extract to describe in face, eyes, nose, face and structural relation change is as recognition feature, the position that uses AAM algorithm search to orient each key feature points of face, select totally 32 key points in eyebrow, eyes, nose, face, lower jaw region, then calculate different distance between different characteristic point as characteristic parameter.

According to above-mentioned principal character, adopt the method that active apparent model and rough set combine to extract in the step of local geometric deformation characteristics of expression, adopt improved rough set attribute reduction method to select the distance feature obtaining, after feature selecting, every width image is selected 14 main distance parameters.

According to above-mentioned principal character, adopt improved weighted principal component analyzing method and rough set to combine and extract in the step of the global feature of expressing one's feelings, for a testing image sequence, first adopt improved weighted principal component analyzing method, construct one using two eyes and face as the two-way Gaussian function in emergence central point San center as weighting function for identification, add level, the gradual change scale-up factor of vertical two directions, realize the two-way adjustable of weighting region, emphasize dispersedly expression shape change to have the eyes of significant contribution, eyebrow, the positional information of these three points of face, make facial expression feature more outstanding.

According to above-mentioned principal character, adopt improved weighted principal component analyzing method and rough set to combine and extract in the step of the global feature of expressing one's feelings, improved weighted principal component analyzing method obtains 52 × 10 eigenvectors matrix, for the higher problem of dimension, still adopt improved rough set attribute reduction method to carry out yojan to 52 dimensional feature vectors of the every width image obtaining, just 24 × 10 eigenvectors matrix of sequence thus obtains expressing one's feelings, its vector is turned to the column vector of 240 dimensions, form the global feature of an image sequence, the global feature vector of an image is made up of the vector of 240 dimensions.

According to above-mentioned principal character, adopt the kernel canonical correlation analysis under higher-dimension small sample to carry out Fusion Features elimination feature redundancy to the local geometric deformation characteristics of extracting and global feature, obtain in the step of the characteristic feature after merging, adopt Canonical Correlation Analysis to carry out Fusion Features to extract two kinds of features above, before fusion, need data set to do respectively standardization, select again suitable kernel function type, the local geometric features and the global feature that extract above merge, the dimension of proper vector z' after Fusion Features is 40 dimensions, observed value vector by normalized as Discrete HMM.

According to above-mentioned principal character, feature after utilization is merged is as the observed value vector of Discrete HMM, carrying out that Classification and Identification obtains in the step of classification results is the observed value vector as HMM after the proper vector normalization of 40 dimensions that obtain after merging, according to the observed value of input, each model training is asked p (o/ λ _i), 1≤i≤6.If i ^*=argmax _1≤i≤6[p (o/ λ _i)], i ^*be the affiliated expression of this segment table feelings sequence classification.

Compared with prior art, the present invention can improve the discrimination of human face expression identification by said method, and the recognition time of Expression Recognition is shorter, thereby is easier to implement.

[brief description of the drawings]

Fig. 1 is for implementing process flow diagram of the present invention.

Fig. 2 is the positioning result figure of AAM.

Fig. 3 is the face characteristic point diagram of demarcating.

Fig. 4 is distance feature description list.

Fig. 5 is HMM structural representation.

Fig. 6 is that kernel canonical correlation analysis obtains the table with test results that fusion feature is identified.

Fig. 7 is the result comparison diagram of four kinds of expression recognition methods.

[embodiment]

The present invention is extensively reading after the domestic and international existing document about human face expression feature extraction and Expression Recognition technology, relatively and use for reference existing successful facial expression recognizing method, gordian technique to corresponding human face expression feature extraction and identification is carried out Improvement and perfection, the algorithm of oneself is proposed simultaneously, improve the discrimination of human face expression identification, and shortened the recognition time of Expression Recognition.

Refer to shown in Fig. 1, implement facial expression recognizing method of the present invention and comprise the steps:

Step 1: carry out face detection;

Step 2: adopt the method that active apparent model and rough set combine to extract the local geometric deformation characteristics of expressing one's feelings;

Step 3: adopt improved weighted principal component analyzing method and rough set to combine and extract the global feature of expression;

Step 4: adopt the kernel canonical correlation analysis under higher-dimension small sample to carry out Fusion Features elimination feature redundancy to the local geometric deformation characteristics of extracting and global feature, obtain the characteristic feature after merging;

Step 5: the feature after utilization is merged, as the observed value vector of Discrete HMM, is carried out Classification and Identification, obtains classification results.

Below respectively step 2 to step 5 is described in detail.

1) local geometric deformation characteristics is extracted

Adopt the method (AAM-RS) that active apparent model and rough set combine to extract the local geometric deformation characteristics of expressing one's feelings.Eyebrow in face, eyes, nose, face etc. have been constructed the colourful expression shape change of face, adopt AAM algorithm to extract the geometric description of describing these organ shape and structural relation variation etc. as recognition feature.The position that use AAM algorithm search is oriented each key feature points of face is as Fig. 2, select totally 32 key points in eyebrow, eyes, nose, face, lower jaw region, as shown in Figure 3, then calculate different distance between different characteristic point as characteristic parameter.The distance (distance in Fig. 3 between unique point 6 and unique point 13) that represents two inner eye corners with d is used as feature normalized factor.

After normalization, putting i is defined as to the Euclidean distance Dis (i, j) of a j:

$\mathrm{Dis}(i,j)=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}/d$

After normalization, putting i is defined as to the vertical range Hei (i, j) of a j:

Hei(i,j)＝|y _i-y _j|/d (1-2)

After normalization, putting i is defined as to the horizontal distance W id (i, j) of a j:

Wid(i,j)＝|x _i-x _j|/d (1-3)

Wherein x _iand x _jrepresent respectively the horizontal ordinate of some i and some j; y _iand y _jrepresent respectively the ordinate of some i and j.The distance feature for Expression Recognition being derived by unique point is described as shown in Figure 4.

For solving redundancy feature and the too high problem of dimension, remove those redundancy features in primitive character, so both ensure that the feature after selecting has stronger recognition capability equally, reduce again the working time of whole system.Adopt improved rough set attribute reduction method to select the distance feature obtaining herein, rough set theory related to the present invention is defined as follows:

Define 1 one knowledge-representation system S and can be expressed as S=<U, C, D, V, f>, wherein, U is the set of object, and C ∪ D=R is community set, and subset C and D are called conditional attribute and decision attribute, v=∪ a ∈ RVa is the set of property value, and Va has represented the scope of property value a ∈ R, and f is an information function, the property value that it specifies each object x in U.

Define the cluster relation of equivalence S on 2 given domain U and U, if , and ∩ P (common factor of all relation of equivalence in P) remains a relation of equivalence on domain U, is called can not differentiate relation on P, is designated as IND (P), and also normal brief note is P.And

$\mathrm{IND}\left(p\right)=\{(x,y)\&Element;U\&times;U:f(x,a)=f(y,a)\&ForAll;a\&Element;p\},$ (if x, y) ∈ IND (p), x, y is called with respect to P and can not differentiates.

It is an equivalence relation family that R is established in definition 3, and r ∈ R, if IND (R)=IND (R-{r}) claims that r is the knowledge that can be divided out in R; If P=R-{r} is independently, P is a yojan in R.If r can not be divided out in R, equivalence relation family R is independently; Otherwise R is correlated with, it is dependence.If if Q is independently, and IND (Q)=IND (P), claim that Q is a yojan of P.In P the set of the relation that is necessary composition be called the core of P, be denoted as CORE (P).Core and yojan have following relation: CORE (P)=∩ RED (P).Wherein, RED (P) represents all yojan of P.Endorse and be interpreted as in the time of Reduction of Knowledge, it be can not cancellation knowledge characteristic set.

If domain U is divided into m equivalence class by definition 4 certain community set, each equivalence class has element to be respectively n1, n2 ... nm, the knowledge quantity that this community set has so $W(n_1,n_2,...,n_m)=W\left(\mathrm{1,1}\right)\&times;\underset{1<i<j<m}{\mathrm{\&Sigma;}}{n}_i\&times;n_j.$ It meets:

①W(n)＝0；

②W(n ₁,...,n _i,...,n _j,...,n _m)＝W(n ₁,...,n _j,...,n _i,...,n _m)；

③W(n ₁,n ₂,...,n _m)＝W(n ₁,n ₂,...,n _m)+W(n ₂,n ₃,...,n _m)；

④W(n ₁,n _i+n _j)＝W(n ₁+n _i)+W(n ₁+n _j)；

It is U that domain is established in definition 5, and P is the set of some attribute in information table, and Q is another community set in information table, and community set Q is designated as W with respect to the Relative knowledge quantity of community set P _{u, Q/P}=W _{u, P ∪ Q}-W _u,P.

Definition 6 is provided with infosystem S, and a (x) is the value of x on attribute a, c _ijin representing matrix, i is capable, and the element of j row, can be defined as by identification matrix: wherein i, j=1,2 ..., n, here n=|U|.

It is a decision table R=C ∪ D that S=(U, R, V, f) is established in definition 7, for conditional attribute collection, D is decision kind set, if

U/C＝{X1,X2,…,Xn}

U/D＝{Y1,Y2,…,Yn}

D is defined as follows about the support of C:

$K_C\left(D\right)=\frac{1}{\left|U\right|}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|\underset{\&OverBar;}{C}{Y}_i\right|=\frac{1}{\left|U\right|}\underset{i=1}{\overset{1}{\mathrm{\&Sigma;}}}\left|{\mathrm{POS}}_C\left(Y_i\right)\right|Y_i\&Element;U/D$

Wherein, || represent the element number of graph-inclusion, decision attribute support is measuring of decision table entirety classification capacity, therefore also referred to as classification quality.

Introduced feature selection algorithm removes those redundancy features in primitive character, has so both ensured that the feature after selecting has the strong recognition capability of religion, the working time of having reduced again whole system equally.Use improved heuristic attribute reduction method herein, algorithm detailed process is described below:

Input: decision system S=(U, C ∩ D), wherein, U is domain, is object set, and C is conditional attribute set, and D is decision attribute set (only having a decision attribute herein).

Output: the yojan RED of conditional attribute set C.

Step 1 initialization: , W (ai)=0;

Step 2 is calculated can identification matrix M, and by the core CORE that can identification matrix generates property set, { ai}, joins the attribute that only comprises individual element in matrix entries in CORE CORE=CORE ∪;

Step 3RED=CORE, and in puncture table M, comprise all of core, the knowledge quantity W (ai) of computation attribute, and calculate | M|;

Step 4while (| M| ≠ 0)

{AR＝C-RED；

Knowledge quantity W (ai) in AR is worth in each from matrix M of minimum attribute and is deleted; The attribute that only comprises individual element in the item of matrix M is joined in RED_TEMP, deletes all that in M, comprise RED_TEMP, calculate WRED (ai), and calculate | M|, RED=RED+RED_TEMP; // delete the little attribute of those knowledge values, then find out the unsuppressible attribute (containing the item of single property element) of deleting after this attribute and join in yojan set and go.}

Step 5Return RED.

These 36 features are carried out to yojan with the improved attribute reduction method in above-mentioned, obtain 14 to system have significant contribution feature (be D1, D3, D4, D6, D8, D9, D17, D18, D23, D24, D26, D29, D31, D32), the variation of face component structure while describing different expression.

2) global feature extracts

For a testing image sequence, first adopt improved weighted principal component analyzing (UWPCA) method, one of improved weighted principal component analyzing (UWPCA) method construct using two eyes and face as the two-way Gaussian function in emergence central point San center as weighting function for identification, add the gradual change scale-up factor of level, vertical two directions, realize the two-way adjustable of weighting region, obtain the adjustable weighting function of the two-way yardstick in following San center:

$\begin{array}{c}\mathrm{\&omega;}(i,j)=\exp \{-\{[{(i-x_1)}^2/a_1^2+{(j-y_1)}^2/b_1^2][{(i-x_2)}^2/a_2^2+{(j-y_2)}^2/b_2^2]\\ [{(i-x_3)}^2/a_3^2+(j-y_3)/b_3^2]^(1/3)\left\}\right\}\end{array}---(2-1)$

Wherein (i, j) represents the position of this dimensional feature (pixel) in image, (x ₁, y ₁), (x ₂, y ₂), (x ₃, y ₃) be respectively the position at three weighting centers.The feature of this function is to have 3 emergence centers, compared with original emergence function, it is not the single information of emphasizing some points or two points, but the positional information of emphasizing 3 points that can disperse, and adds the gradual change scale-up factor of level, vertical two directions.Emphasizing dispersedly has eyes, the eyebrow of significant contribution, the positional information of these three points of face to expression shape change, makes facial expression feature more outstanding.Found through experiments, in the time that three emergence center bright spot shapes of emergence function are respectively ellipse, recognition effect is better than it and is shaped as bowlder, works as a ₁=a ₂=15, b ₁=b ₂=20, a ₃=15, b ₃=25 o'clock, recognition effect was best.

Obtain 52 × 10 eigenvectors matrix through said method, for the higher problem of dimension, still adopt improved rough set attribute reduction method to carry out yojan to 52 dimensional feature vectors of the every width image obtaining, just 24 × 10 eigenvectors matrix of sequence thus obtains expressing one's feelings, its vector is turned to the column vector of 240 dimensions, the global feature that forms an image sequence, the global feature vector of an image is made up of the vector of 240 dimensions.

3) local and overall expressive features merges

Adopt Canonical Correlation Analysis (CCA) to carry out Fusion Features to extract two kinds of features above, in fusion process, owing to participating in, may there are the different of dimension selection from y in two data set x of fusion or each component differs greatly, be unfavorable for the extraction of correlated characteristic, before fusion, need data set to do respectively standardization

$x^{*}=\frac{x-{\mathrm{\&mu;}}_x}{{\mathrm{\&sigma;}}_x},y^{*}=\frac{y-{\mathrm{\&mu;}}_y}{{\mathrm{\&sigma;}}^y}---(3-1)$

Wherein μ _x=E (x), μ _y=E (y) is respectively the mean vector of sample; σ _x, σ _ythe average of the standard deviation vector that is respectively sample on each component.Kernel function K ' after definition mapping _ij=< φ ' (x _i), φ ' (x _j) >, have:

$K_{\mathrm{ij}}^{\&prime;}\left[\right[\mathrm{\&phi;}\left(x_i\right)-\frac{1}{n}\underset{p=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{\&phi;}\left(x_p\right)],[\mathrm{\&phi;}\left(x_j\right)-\frac{1}{n}\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{\&phi;}\left(x_q\right)\left]\right]---(3-2)$

From theory of reproducing kernel space:

$M=\frac{1}{N}{K}_x^T{\mathrm{JK}}_y,P=\frac{1}{N}{K}_x^T{\mathrm{JK}}_x+{\mathrm{\&eta;}}_1{K}_x,Q=\frac{1}{N}{K}_y^T{\mathrm{JK}}_y+{\mathrm{\&eta;}}_2{K}_y,J=I-\frac{1}{N}{A}^T---(3-3)$

A is that all elements is 1 N dimensional vector, η 1=η 2=η.

Solve vector α and β, the generalized eigenvalue characteristic of correspondence vector of a demand solution P and M, or decompose and try to achieve by Cholesky.Try to achieve after α and β, extract the non-linear canonical correlation feature between x and y.

Non-linear correlation feature between combination part and the entirety extracted is as the linked character of the two:

$z^{\&prime;}={\left[\begin{array}{c}\mathrm{\&beta;}\\ \mathrm{\&beta;}\end{array}\right]}^T\left[\begin{array}{c}{K}_x\\ K_y\end{array}\right]---(3-4)$

Z ' is the feature after tried to achieve fusion.

It is as follows that expression merges detailed process:

1) the feature input vector that two kinds of feature extracting methods obtain is respectively X, and Y, utilizes the standardization pre-service before formula (3-1) merges, and obtains respectively X', Y'.

2) select again suitable kernel function type, try to achieve kernel function k _x, k _y, and utilize formula (3-2) to carry out centralization pre-service, obtain respectively k ' _x, k ' _y.

3) set up matrix M according to formula (3-3), P, Q.

4) by solving the generalized eigenvalue characteristic of correspondence vector of P and M, or decomposed to try to achieve by Cholesky and obtain α, β, then try to achieve fusion feature collection z' according to formula (3-4).

According to above algorithm steps 1)～4), the local geometric features of extracting above and global feature are merged, and the dimension of the proper vector z' after Fusion Features is 40 dimensions, the observed value vector by normalized as Discrete HMM.

4) Classification and Identification based on HMM

After the proper vector normalization of 40 dimensions that obtain after merging, as the observed value vector of HMM, according to the observed value of input, each model training is asked p (o/ λ _i), 1≤i≤6.If i ^*=argmax _1≤i≤6[p (o/ λ _i)], i ^*be the affiliated expression of this segment table feelings sequence classification.HMM Expression Recognition system architecture schematic diagram as shown in Figure 5.The algorithm proposing is herein tested in the expression storehouse of Cohn-Kanade (Cohn-Kanade AU-coded Facial Expression Database), choosing at random the facial expression image of 40 people in Cohn-Kanade expression storehouse tests, front 10 people's facial expression image sequence is as training sample, 30 people's of residue facial expression image sequence is as test sample book, everyone comprises 10 width images by the order growing from weak to strong at every kind of expression, totally 2400 width images.

Merge two category features through kernel canonical correlation analysis, eliminate feature redundancy, obtain the characteristic feature after merging, train six hidden Markov models.After having trained, we test every kind of random 30 sequences of selection of expression, and test result as shown in Figure 6.Frightened, glad as can be seen from the table, the discrimination of surprised three classes expressions is higher, and detest and sad discrimination relatively low.Average recognition rate is 91.12%.Glad, surprised discrimination is relatively high is that two kinds of expressions of this collecting, mostly all in very big state, are expressed one's feelings fairly proper because in our image library used.And detest relatively, other several Expression Recognition rates are very low, and reason is in image library, and the most amplitude of this expression is less, and move lack of standardizationly, are easily mistaken for sadness.

For the contrast validity of Fusion Features herein, at training and testing image sequence all under identical condition, compare that the present invention and canonical correlation analysis merge and two kinds of single features extract and the recognition result of HMM, as shown in Figure 7, horizontal ordinate represents six kinds of basic facial expressions, and ordinate represents discrimination.As can be seen from Figure 7, average recognition rate of the present invention is higher than other three kinds.

Compared with prior art, the present invention can improve the discrimination of human face expression identification by said method, and the recognition time of Expression Recognition is shorter, thereby is easier to implement.

Claims (7)

1. the facial expression recognizing method based on rough set and composite character, comprises the steps:

Carry out face detection;

Adopt the method that active apparent model and rough set combine to extract the local geometric deformation characteristics of expressing one's feelings;

Adopt improved weighted principal component analyzing method and rough set to combine and extract the global feature of expression;

Adopt the kernel canonical correlation analysis under higher-dimension small sample to carry out Fusion Features elimination feature redundancy to the local geometric deformation characteristics of extracting and global feature, obtain the characteristic feature after merging;

Feature after utilization is merged, as the observed value vector of Discrete HMM, is carried out Classification and Identification, obtains classification results.

2. the facial expression recognizing method based on rough set and composite character as claimed in claim 1, it is characterized in that: adopt the method that active apparent model and rough set combine to extract in the step of local geometric deformation characteristics of expression, adopt AAM algorithm to extract the eyebrow of describing in face, eyes, nose, the geometric description that these organ shape such as face and structural relation change is as recognition feature, the position that uses AAM algorithm search to orient each key feature points of face, select eyebrow, eyes, nose, face, totally 32 key points in lower jaw region, then calculate different distance between different characteristic point as characteristic parameter.

3. the facial expression recognizing method based on rough set and composite character as claimed in claim 2, it is characterized in that: adopt the method that active apparent model and rough set combine to extract in the step of local geometric deformation characteristics of expression, adopt improved rough set attribute reduction method to select the distance feature obtaining, after feature selecting, every width image is selected 14 main distance parameters.

4. the facial expression recognizing method based on rough set and composite character as claimed in claim 3, it is characterized in that: adopt improved weighted principal component analyzing method and rough set to combine and extract in the step of the global feature of expressing one's feelings, for a testing image sequence, first adopt improved weighted principal component analyzing method, construct one using two eyes and face as the two-way Gaussian function in emergence central point San center as weighting function for identification, add level, the gradual change scale-up factor of vertical two directions, realize the two-way adjustable of weighting region, emphasize dispersedly expression shape change to have the eyes of significant contribution, eyebrow, the positional information of these three points of face, make facial expression feature more outstanding.

5. the facial expression recognizing method based on rough set and composite character as claimed in claim 4, it is characterized in that: adopt improved weighted principal component analyzing method and rough set to combine and extract in the step of the global feature of expressing one's feelings, improved weighted principal component analyzing method obtains 52 × 10 eigenvectors matrix, for the higher problem of dimension, still adopt improved rough set attribute reduction method to carry out yojan to 52 dimensional feature vectors of the every width image obtaining, just 24 × 10 eigenvectors matrix of sequence thus obtains expressing one's feelings, its vector is turned to the column vector of 240 dimensions, form the global feature of an image sequence, the global feature vector of an image is made up of the vector of 240 dimensions.

6. the facial expression recognizing method based on rough set and composite character as claimed in claim 5, it is characterized in that: adopt the kernel canonical correlation analysis under higher-dimension small sample to carry out Fusion Features elimination feature redundancy to the local geometric deformation characteristics of extracting and global feature, obtain in the step of the characteristic feature after merging, adopt Canonical Correlation Analysis to carry out Fusion Features to extract two kinds of features above, before fusion, need data set to do respectively standardization, select again suitable kernel function type, the local geometric features and the global feature that extract above merge, the dimension of proper vector z' after Fusion Features is 40 dimensions, observed value vector by normalized as Discrete HMM.

7. the facial expression recognizing method based on rough set and composite character as claimed in claim 6, it is characterized in that: the feature after utilization is merged is as the observed value vector of Discrete HMM, carrying out that Classification and Identification obtains in the step of classification results is the observed value vector as HMM after the proper vector normalization of 40 dimensions that obtain after merging, according to the observed value of input, each model training is asked p (o/ λ _i), 1≤i≤6.If i ^*=argmax _1≤i≤6[p (o/ λ _i)], i ^*be the affiliated expression of this segment table feelings sequence classification.