CN104408468A - Face recognition method based on rough set and integrated learning - Google Patents

Abstract

The invention discloses a face recognition method based on rough set and integrated learning. The face recognition method includes steps of 1, extracting face images to preprocess and generate decision tables; 2, subjecting the decision tables to attribute reduction to obtain the optimal attribute reduction space; 3, selecting an optimal attribute reduction combination random from the optimal attribute reduction space and storing; 4, calculating dependency degree of each condition attribute in the optimal attribute reduction combination; 5, reducing reduction space according to the dependency degree; 6, selecting the optimal attribute reduction combination from the reduced reduction space and storing; 7, judging the number of the stored optimal attribute reduction combination is up to the preset value, if yes, returning to the step 4, otherwise, training a basic classifier by the stored optimal attribute reduction combination and recognizing the face images. Through attribute reduction of the face images, recognition rate and stability of the face images are improved. The face recognition method can be used for anti-terrorist safety, case investigation and evidence collection and confidentiality of business and individual privacy.

Description

Based on the face identification method of rough set and integrated study

Technical field

Technical field of image processing of the present invention, particularly a kind of face identification method, can be used for anti-terrorism safety, case investigate and collect evidence and business, individual privacy maintain secrecy.

Background technology

Closely for decades, face recognition technology becomes the study hotspot in the field such as image procossing, pattern-recognition gradually.Face recognition technology is a kind of biometrics identification technology carrying out identity authentication according to the facial characteristics of people, because face characteristic is that mankind itself has, and be difficult to forge and steal, the advantage not needing direct contact, be not easily therefore easily perceived by humans when adding identification, therefore it is in anti-terrorism safety, case investigation and evidence collection and business and individual privacy area of security all extensive application.Academia and industry member more and more pay attention to the research of face recognition technology, and it is the technology that in current living things feature recognition, in the professional technique index such as feasibility, reliability and accuracy, numerical value is higher.

Face identification method is vulnerable to ambient light, and human body face is blocked by hair, jewelry, and people's age increases the old and feeble impact waiting many factors of face.Traditional face identification method based on single support vector machines, attempts to make sorter on empiric risk and generalization ability, reach a kind of and compromises, and then improve the performance of sorter.Single SVM classifier, when there being a large amount of training sample, can obtain good recognition performance, but in actual applications, obtains a large amount of training samples unpractical often.

Integrated study decides final classification, to obtain performance more better than single sorter by the classification results of multiple single sorter being carried out certain combination.Typical integrated learning approach comprises Bagging, Adaboost and Randomsubspace, and they adopt the different training set of strategy generation for training basic classification device.The training set of training basic classification device to need in Bagging method is made up of the sample randomly drawing some in original training set; Adaboost method is according to the classification results of each basic classification device to sample each in training set, and last time general classification accuracy rate, the weights of each sample are adjusted, reduce the weights of correct sample of being classified by basic classification device, improve by the weights of the sample of basic classification device classification error, produce new training set with this by the weights changing each sample; The Randomsubspace method attribute that Stochastic choice is different from original training set forms new training set.

Above-mentioned integrated learning approach, owing to all not considering the relation in face sample between attribute preferably, does not also effectively extract face characteristic, therefore can not obtain good recognition of face effect.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, propose a kind of face identification method based on rough set and integrated study, improve face identification rate.

Technical thought of the present invention is: from facial image, extracting the many groups of features being conducive to classifying, and guaranteeing when not having a large amount of training sample, by several basic classification devices integrated, improve the discrimination to face.

According to above-mentioned thinking, technical scheme of the present invention comprises the steps:

(1) facial image pre-service:

(1a) from standard faces storehouse, obtain T width facial image, and be divided into N width training sample image and M width test sample image, wherein T >=2, N >=1, M >=1, N+M=T;

(1b) for training sample image sets up decision table S, it is N × L matrix, every a line represents a width facial image, each row represents an attributive character of facial image, wherein L>2, the conditional attribute of front L-1 list diagram picture, the decision attribute of L list diagram picture;

(2) carry out Property element to decision table S, produce many group best attributes and about subtracts combination, what about subtract combination formation best attributes by these best attributes about subtracts space K={A ₁.., A _i.., A _z, wherein A _ibe that i-th group of best attributes about subtracts combination, i=1 ..., z, z>=2, A _i={ c ₁.., c _j.., c _n, c _jfor a jth conditional attribute of image, j=1 ..., n, 1<n<L;

(3) about subtract space from best attributes, Stochastic choice one group of best attributes about subtracts combination, and preserves.

(4) best attributes calculating up-to-date preservation about subtracts the dependence angle value formula of each conditional attribute in combination: wherein r _p(c _j) be conditional attribute c _jto the dependence angle value of P, P deletes conditional attribute c for about subtracting combination from best attributes _jafter property set, POS _p(c _j) be conditional attribute c _jthe positive region of relative priority collection P, || be the gesture of " ";

(5) the maximum conditional attribute c corresponding to dependence angle value is found out _m, wherein m ∈ [1, n], comprises this conditional attribute c from about subtracting the K of space to delete _mbest attributes about subtract combination, with remaining best attributes about subtract combination formation one reduce about subtract space;

(6) about subtract space from what reduce, utilize accurately-various valuation functions to select one group of best attributes about to subtract combination, and preserve.

(7) judge whether the quantity that the best attributes of preserving about subtracts combination reaches the value preset: if quantity does not reach the value preset, then return step (4), if quantity reaches the value preset, then perform step (8);

(8) utilize each best attributes of preserving about to subtract combination structure training sample set, obtain the many groups training sample set rebuild; Train a basic classification device with each training sample set, utilize different basic classification devices to classify to test sample book collection, obtain different classification results; Adopt the majority voting method in integrated study to merge these different classification results, obtain the final recognition result of test sample book collection.

Tool of the present invention has the following advantages:

1. the inventive method carries out Property element owing to utilizing fuzzy coarse central algorithm to facial image, effectively compensate for the defect that traditional Property element method reduces attribute resolution characteristic.

2. the inventive method is owing to adopting the sorting technique of integrated study, improves the problem that traditional single sorter classification accuracy rate is lower and robustness is poor.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the facial image schematic diagram in ORL and CMU PIE database;

Fig. 3 is the influence curve figure of number for performance of the present invention of basic classification device.

Embodiment

With reference to Fig. 1, specific embodiment of the invention step is as follows:

Step 1, facial image pre-service.

(1a) facial image is chosen.

This example uses ORL and CMU PIE two standard sets face database, chooses the 400 width images be made up of 40 people, choose the 3329 width images be made up of 68 people from CMU PIE from ORL; Utilize the imresize function in MATLAB software that the size of image is adjusted to 32 × 32, as shown in Figure 2, wherein Fig. 2 (a) is the facial image schematic diagram of first man in ORL database, and Fig. 2 (b) is the facial image schematic diagram of first man in CMU PIE database.

From often organizing facial image, random selecting half image is as training sample image, and second half image is as test sample image.

(1b) for training sample image sets up decision table S: utilize the reshape function in MATLAB software to convert each width facial image to a row vector respectively, each row vector comprises 1024 elements in the present embodiment, and N width facial image forms the matrix A of N × 1024; The column vector B of N × 1 is formed with the decision attribute of N width facial image; The matrix A obtained and column vector B are merged, obtain the decision table S of N width training sample image, it is N × 1025 matrix.

Step 2, Property element is carried out to decision table S.

What the present embodiment utilized Richard Jensen to write the fuzzy coarse central algorithm be embedded in weka software carries out Property element to decision table S, produce many group best attributes and about subtract combination, about subtract combination formation best attributes by these best attributes and about subtract space K={A ₁.., A _i.., A _z, wherein A _ibe that i-th group of best attributes about subtracts combination, i=1 ..., z, z>=2, A _i={ c ₁.., c _j.., c _n, c _jfor a jth conditional attribute of image, j=1 ..., n, 1<n<L;

The method of decision table S being carried out to Property element is not limited to fuzzy coarse central algorithm, except fuzzy coarse central algorithm can also use WADF algorithm, specifically see " Multiknowledge for decision making " Knowledge andinformation systems, 2005,7 (2): 246-266.

Step 3, about subtract space from best attributes, Stochastic choice one group of best attributes about subtracts combination, and preserves.

Step 4, the best attributes calculating up-to-date preservation about subtract the dependence angle value of each conditional attribute in combination.

(4a) training sample set U={x ₁.., x _k.., x _n, wherein x _kfor a kth sample image, for property set P, calculate Indiscernible relation I (P):

$I\left(P\right)=\{(x_k,x_l)\&Element;U\&times;U|c\&Element;P,g(x_k,c)=g(x_l,c)\}$

Wherein, P deletes conditional attribute c for about subtracting combination from best attributes _jafter property set, c is the conditional attribute in property set P, x _lbe l sample image, g (x _k, c) represent the value corresponding to c attribute of a kth sample image, g (x _l, c) represent the value corresponding to c attribute of l sample image;

(4b) according to the Indiscernible relation that step (4a) obtains, the equivalence class [x] produced based on property set P is calculated _p: [x] _p={ x| (x _k, x _l) ∈ I (P), x ∈ U};

(4c) according to the equivalence class obtained in step (4b), computation attribute collection P is for the lower aprons of sample set X with upper approximate

$\left\{\begin{array}{c}\underset{\&OverBar;}{P}\left(X\right)=\{x\&Element;U|{\left[x\right]}_P\&SubsetEqual;X\}\\ \stackrel{\&OverBar;}{P}\left(X\right)=\{x\&Element;U|{\left[x\right]}_P\&cap;X\&NotElement;\mathrm{\&phi;}\}\end{array}\right.,$

Wherein, X is based on property set c _ithe equivalence class produced expression comprises, and ∩ represents crossing, and φ represents empty set;

(4d) according to the lower aprons obtained in step (4c) design conditions attribute c _jthe positive region of relative priority collection P:

${\mathrm{POS}}_P\left(c_j\right)=\&cup;\underset{\&OverBar;}{P}\left(X\right);$

(4e) according to the conditional attribute c obtained in step (4d) _jthe positive region POS of relative priority collection P _p(c _j), calculate the dependence angle value that best attributes about subtracts each conditional attribute in combination:

$r_P\left(c_j\right)=\frac{\left|{\mathrm{POS}}_P\left(c_j\right)\right|}{N},$

Wherein, || represent the gesture of " ".

What step 5, structure reduced about subtracts space.

Find out the maximum conditional attribute c corresponding to dependence angle value _m, wherein m ∈ [1, n], comprises this conditional attribute c from about subtracting the K of space to delete _mbest attributes about subtract combination, with remaining best attributes about subtract combination formation one reduce about subtract space.

Step 6, about subtract space from what reduce, utilize accurately-various valuation functions to select one group of best attributes about to subtract combination, and preserve.

(6a) the empirical loss A of basic classification device is calculated _emp(F _t, N):

$A_{\mathrm{emp}}(F_t,N)=\frac{1}{a_{t}N}\underset{b=1}{\overset{a_t}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{[f_b\left(x_k\right)-y_k]}^2,$

Wherein, basic classification device group F _tbe made up of multiple fixed sorter and a basic classification device, these fixed sorters about subtract combined training based on the best attributes of preserving to obtain, and a basic classification device about subtracts combined training based on the best attributes about subtracted in space reduced to obtain; a _trepresent basic classification device group F _tthe quantity of middle basic classification device, f _b(x _k) represent basic classification device group F _tin b sorter to the result of a kth face Images Classification, y _krepresent the decision attribute of a kth facial image, wherein for about subtracting the sum of the sorter of combined training based on the best attributes about subtracted in space reduced, b ∈ [1, a _t], k ∈ [1, N];

(6b) the diversity numerical value D between basic classification device is calculated _div(F _t, N):

The diversity numerical value calculated between basic classification device has multiple method to realize, comprise CFD, ENT and DF method, specifically see " Measures of diversity in classifier ensembles and their relationship with the ensembleaccuracy " Machine learning, 2003,51 (2): 181-207, the present embodiment utilizes the diversity numerical value between DF calculating basic classification device, and its formula is as follows:

$\left\{\begin{array}{c}{D}_{\mathrm{div}}(F_t,N)=\frac{2}{a_t(a_t-1)}\underset{r=1}{\overset{a_t-1}{\mathrm{\&Sigma;}}}\underset{q=r+1}{\overset{a_t}{\mathrm{\&Sigma;}}}{\mathrm{df}}_{\mathrm{rq}}\\ {\mathrm{df}}_{\mathrm{rq}}=\frac{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(1-O_{\mathrm{rk}})\&CenterDot;(1-O_{\mathrm{qk}})}{N}\end{array}\right.,$

Wherein, O _rkbe defined as: kth width facial image, then O if r sorter can correctly be classified _rk=1, otherwise O _rk=0, O _qkbe defined as, kth width facial image, then O if q sorter can correctly be classified _qk=1, otherwise O _qk=0, r ∈ [1, a _t-1], q ∈ [r+1, a _t];

(6c) the empirical loss A of the basic classification device obtained according to step (6a) _emp(F _t, N) and the basic classification device that obtains of step (6b) between diversity numerical value D _div(F _t, N), obtain accurately-various valuation functions:

AD(F _t,N)＝1-A _emp(F _t,N)+ω×D _div(F _t,N)，

Wherein, ω is balance empirical loss and multifarious parameter, in the present embodiment ω=1, but the value of ω is not limited to 1, and it rule of thumb obtains;

(6d) according to accurate-various valuation functions that step (6c) obtains, for each the group best attributes about subtracted in space reduced about subtract combination calculating one accurately-various valuation functions value AD (F _t, N), select the maximum best attributes corresponding to accurate-various valuation functions value about to subtract combination and preserved.

Step 7, judge whether the quantity that the best attributes of preserving about subtracts combination reaches the value preset: if quantity does not reach the value preset, then return step 4, if quantity reaches the value preset, then perform step 8;

Step 8, training basic classification device, test sample image carries out recognition of face, obtains the final result identified.

(8a) utilize each best attributes of preserving about to subtract combination structure training sample set, obtain the many groups training sample set rebuild:

(8a1) provide a best attributes and about subtract combination A _i={ c ₁.., c _j.., c _n, wherein c _jfor a jth conditional attribute of training sample set;

(8a2) from training sample set U, delete best attributes about subtract combination A _iin do not have occur attribute, obtain the training sample set rebuild wherein for a kth sample image.

(8b) a basic classification device is trained with each training sample set, different basic classification devices is utilized to classify to test sample book collection, obtain different classification results, use SVM as basic classification device in the present embodiment, but the selection of basic classification device is not limited to SVM, also can trade-off decision set as basic classification device;

(8c) adopt the majority voting method in integrated study to merge these different classification results, obtain the final recognition result of test sample image.

Effect of the present invention can be illustrated by emulation experiment:

1. experiment condition

Testing microcomputer CPU used is Intel Core (TM) 2Duo 3GHz internal memory 2GB, and programming platform is MatlabR2012a.

These two standard faces storehouses of ORL and CMU PIE are selected in experiment.Wherein, ORL standard faces storehouse is by 40 people, and everyone forms by 10 width images, and image comprises the dimensional variation within different expressions, small attitude and 20%; For CMU PIE standard faces storehouse, we choose and are tested by the facial image of 68 people totally 3329 different light.The size of these two groups of facial images used in experiment all adjusts to 32 × 32, as shown in Figure 2, wherein Fig. 2 (a) is the facial image of first man in ORL database, and Fig. 2 (b) is the facial image of first man in CMU PIE database.

In experiment, random selecting half image is as training sample, and second half image is as test sample book.

2. experiment content and result

Experiment one: the fuzzy coarse central algorithm be embedded in weka software utilizing Richard Jensen to write, produces best attributes and about subtracts combination, and the number that the best attributes for the generation of different face databases about subtracts combination is as shown in table 1:

Table 1 best attributes about subtracts the number of combination

Data set	Best attributes about subtracts the number of combination
		ORL	133
CMU PIE	135

As can be seen from Table 1, ORL database uses this algorithm can produce 133 groups of best attributes and about subtract combination, CMU PIE database uses this algorithm can produce 135 groups of best attributes and about subtract combination.

Experiment two: use SVM as basic classification device, emulate the accuracy of the present invention when the quantity of basic classification device is increased to 30 from 5, often organize emulation and repeat 30 times, calculate the average correct classification rate of 30 emulation, as final classification accuracy rate, result as shown in Figure 3, the horizontal ordinate of Fig. 3 is the number of sorter, ordinate is classification accuracy rate, wherein Fig. 3 (a) is the accuracy curve map on ORL database, and Fig. 3 (b) is the accuracy curve map on CMU PIE database;

As can be seen from Figure 3, when the quantity of basic classification device reaches 25, the present invention can obtain comparatively high-class accuracy.

Experiment three: the present invention and additive method are compared under same Setup Experiments prerequisite, in the present embodiment, the quantity of basic classification device is decided to be 25.These methods comprise: Single represents that the single SVM classifier of use carries out the method for classifying on facial image, Bagging represents the classical integrated approach of existing one, AdaBoost represents the classical integrated approach of existing another kind, and RS represents existing stochastic subspace integrated approach.

The Comparative result of table 2 distinct methods on ORL and CMU PIE database (average % ± variance %)

Database	Single	Bagging	AdaBoost	RS	The present invention
						ORL	52.75±4.5	69.38±4.7	71.33±4.3	68.91±4.9	78.75±3.9
CMU PIE	62.26±1.4	87.32±1.2	88.64±1.7	86.99±2.3	91.01±2.8

As can be seen from Table 2, the present invention has higher accuracy rate compared with the existing methods.

To sum up, the face identification method based on rough set and integrated study that the present invention proposes, carries out Property element to facial image, effectively compensate for the defect that traditional Property element method reduces attribute resolution characteristic; Adopt the sorting technique of integrated study, improve the problem that traditional single sorter classification accuracy rate is lower and robustness is poor.

Claims (7)

1., based on a face identification method for rough set and integrated study, comprise the steps:

(1) facial image pre-service:

(1a) from standard faces storehouse, obtain T width facial image, and be divided into N width training sample image and M width test sample image, wherein T >=2, N >=1, M >=1, N+M=T;

(1b) for training sample image sets up decision table S, it is N × L matrix, every a line represents a width facial image, each row represents an attributive character of facial image, wherein L>2, the conditional attribute of front L-1 list diagram picture, the decision attribute of L list diagram picture;

(2) carry out Property element to decision table S, produce many group best attributes and about subtracts combination, what about subtract combination formation best attributes by these best attributes about subtracts space K={A ₁.., A _i.., A _z, wherein A _ibe that i-th group of best attributes about subtracts combination, i=1 ..., z, z>=2, A _i={ c ₁.., c _j.., c _n, c _jfor a jth conditional attribute of image, j=1 ..., n, 1<n<L;

(3) about subtract space from best attributes, Stochastic choice one group of best attributes about subtracts combination, and preserves;

(4) best attributes calculating up-to-date preservation about subtracts the dependence angle value formula of each conditional attribute in combination: wherein r _p(c _j) be conditional attribute c _jto the dependence angle value of P, P deletes conditional attribute c for about subtracting combination from best attributes _jafter property set, POS _p(c _j) be conditional attribute c _jthe positive region of relative priority collection P, || be the gesture of " ";

(5) the maximum conditional attribute c corresponding to dependence angle value is found out _m, wherein m ∈ [1, n], comprises this conditional attribute c from about subtracting the K of space to delete _mbest attributes about subtract combination, with remaining best attributes about subtract combination formation one reduce about subtract space;

(6) about subtract space from what reduce, utilize accurately-various valuation functions to select one group of best attributes about to subtract combination, and preserve;

(7) judge whether the quantity that the best attributes of preserving about subtracts combination reaches the value preset: if quantity does not reach the value preset, then return step (4), if quantity reaches the value preset, then perform step (8);

(8) utilize each best attributes of preserving about to subtract combination structure training sample set, obtain the many groups training sample set rebuild; Train a basic classification device with each training sample set, utilize different basic classification devices to classify to test sample book collection, obtain different classification results; Adopt the majority voting method in integrated study to merge these different classification results, obtain the final recognition result of test sample book collection.

2. the face identification method based on rough set and integrated study according to claim 1, decision table S is set up for training sample image: be utilize the reshape function in MATLAB software to convert each width facial image to a row vector respectively wherein described in step (1b), each row vector comprises L-1 conditional attribute, obtains the matrix A of a N × (L-1) of N width facial image; The column vector B of N × 1 is formed with the decision attribute of N width facial image; The matrix A obtained and column vector B are merged, obtain the decision table of N width facial image, it is N × L matrix, and wherein the concrete value of L is determined by the size of facial image.

3. the face identification method based on rough set and integrated study according to claim 1, in wherein said step (2), best attributes about subtracts combination, refer to the minimum attribute set of training sample image, and this subset has the resolution characteristic of primitive attribute.

4. the face identification method based on rough set and integrated study according to claim 1, wherein step (2) is described carries out Property element to decision table S, producing many group best attributes and about subtract combination, is that the fuzzy coarse central algorithm be embedded in weka software utilizing Richard Jensen to write realizes.

5. the face identification method based on rough set and integrated study according to claim 1, wherein said step (4) conditional attribute c _jto the dependence angle value r of P _p(c _j), obtain as follows:

(4a) training sample set U={x ₁.., x _k.., x _n, wherein x _kfor a kth sample image, for property set P, calculate Indiscernible relation I (P):

$I\left(P\right)=\{(x_k,x_l)\&Element;U\&times;U|\&ForAll;c\&Element;P,g(x_k,c)=g(x_l,c)\},$

Wherein, c is the conditional attribute in property set P, x _lbe l sample image, g (x _k, c) represent the value corresponding to c attribute of a kth sample image, g (x _l, c) represent the value corresponding to c attribute of l sample image;

(4b) according to the Indiscernible relation that step (4a) obtains, the equivalence class [x] produced based on property set P is calculated _p: [x] _p={ x| (x _k, x _l) ∈ I (P), x ∈ U};

(4c) according to the equivalence class obtained in step (4b), computation attribute collection P is for the lower aprons of sample set X pand upper approximate (X)

$\left\{\begin{array}{c}\underset{\&OverBar;}{P}\left(X\right)=\{x\&Element;U|{\left[x\right]}_P\&SubsetEqual;X\}\\ \stackrel{\&OverBar;}{P}\left(X\right)=\{x\&Element;U|{\left[x\right]}_P\&cap;X\&NotEqual;\mathrm{\&phi;}\}\end{array}\right.,$

Wherein, X is based on property set c _ithe equivalence class produced φ represents empty set;

(4d) according to the lower aprons obtained in step (4c) p(X), design conditions attribute c _jthe positive region of relative priority collection P:

POS _P(c _j)＝∪ P(X)。

6. the face identification method based on rough set and integrated study according to claim 1, about subtract space from what reduce wherein described in step (6), accurately-various valuation functions is utilized to select one group of best attributes about to subtract the method for combination, carry out as follows:

(6a) the empirical loss A of basic classification device is calculated _emp(F _t, N):

$A_{\mathrm{emp}}(F_t,N)=\frac{1}{a_{t}N}\underset{b=1}{\overset{a_t}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{[f_b\left(x_k\right)-y_k]}^2,$

Wherein, basic classification device group F _tbe made up of multiple fixed sorter and a basic classification device, these fixed sorters about subtract combined training based on the best attributes of preserving to obtain, and a basic classification device about subtracts combined training based on the best attributes about subtracted in space reduced to obtain; a _trepresent basic classification device group F _tthe quantity of middle basic classification device, f _b(x _k) represent basic classification device group F _tin b sorter to the result of a kth face Images Classification, y _krepresent the decision attribute of a kth facial image, wherein for about subtracting the sum of the sorter of combined training based on the best attributes about subtracted in space reduced, b ∈ [1, a _t], k ∈ [1, N];

(6b) the diversity D between basic classification device is calculated _div(F _t, N):

$\left\{\begin{array}{c}{D}_{\mathrm{div}}(F_t,N)=\frac{2}{a_t(a_t-1)}\underset{r=1}{\overset{a_t-1}{\mathrm{\&Sigma;}}}\underset{q=r+1}{\overset{a_t}{\mathrm{\&Sigma;}}}d{f}_{\mathrm{rq}}\\ d{f}_{\mathrm{rq}}=\frac{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(1-O_{\mathrm{rk}})\&CenterDot;(1-O_{\mathrm{qk}})}{N}\end{array}\right.,$

Wherein, O _rkbe defined as: kth width facial image, then O if r sorter can correctly be classified _rk=1, otherwise O _rk=0, O _qkbe defined as, kth width facial image, then O if q sorter can correctly be classified _qk=1, otherwise O _qk=0, r ∈ [1, a _t-1], q ∈ [r+1, a _t];

(6c) the empirical loss A of the basic classification device obtained according to step (6a) _emp(F _t, N) and the basic classification device that obtains of step (6b) between diversity D _div(F _t, N), what calculating was reduced about subtracts accurate-various valuation functions value that each best attributes in space about subtracts combination:

AD(F _t,N)＝1-A _emp(F _t,N)+ω×D _div(F _t,N)，

Wherein, ω is balance empirical loss and multifarious parameter;

(6d) the maximum best attributes corresponding to accurate-various valuation functions value is selected about to subtract combination.

7. the face identification method based on rough set and integrated study according to claim 1, utilizes best attributes about to subtract combination to rebuild training sample set, carry out as follows wherein described in step (8):

(7a) provide a best attributes and about subtract combination A _i={ c ₁.., c _j.., c _n, wherein c _jfor a jth conditional attribute of training sample set;

(7b) from training sample set U, delete best attributes about subtract combination A _iin do not have occur attribute, obtain the training sample set rebuild $\stackrel{\&OverBar;}{U}=\{\stackrel{\&OverBar;}{x_1},..,\stackrel{\&OverBar;}{x_k},..,\stackrel{\&OverBar;}{x_N}\},$ Wherein for a kth sample image.