CN106709480A - Partitioning human face recognition method based on weighted intensity PCNN model - Google Patents

Abstract

The invention relates to a human face recognition method based on a PCNN model, specifically a partitioning human face recognition method based on a weighted intensity PCNN model. The invention solves problems that a conventional human face recognition method based on the PCNN model is not fine in description of image features, carries out no-difference processing of the whole human face image through a group of parameters, and neglects the differences of all parts of a human face. On the basis of simplifying the PCNN model, the invention proposes the weighted intensity PCNN model, introduces the concepts of emission intensity of a spontaneous pulse, the emission intensity of a coupled pulse and the weighted intensity, and refines the output of the model. Meanwhile, the method employs the partitioning recognition during human face recognition. The method comprises the steps: enabling a human face image to be divided into blocks according to the difference of gray scale distribution of all parts of the human face image and the difference of local resolutions before recognition; adaptively setting the weight value of blocks according to the image blocks during recognition; finally enabling the recognition result of each block to be integrated in the recognition result of one human face image.

Description

Intersected human face recognition methods based on weighed intensities PCNN models

Technical field

Face identification method the present invention relates to be based on PCNN models, the specially piecemeal based on weighed intensities PCNN models Face identification method.

Background technology

Pulse Coupled Neural Network is the lock-out pulse of the brain visual cortexs of animal such as Eckhorn et al. researchs cat, monkey Provide what is found during phenomenon, and it is improved in 1993 by Johnson, it is proposed that the pulse coupled neural net of standard Network (Pulse Coupled Neural Network PCNN) model.1999, Lzhikevich was from the strict card of mathematical angle Understand that PCNN models are a kind of network models closest to biological neuron.PCNN models compared with conventional classical neural network, It is a kind of network of self-supervisory self study, it is not necessary to which training is capable of achieving image procossing, and PCNN models have excellent synthesis Space-time summation characteristic, dynamic synchronization pulse provide the characteristics such as characteristic, non-linear modulation and variable threshold, therefore, PCNN models with It is widely used in the aspects such as Image Edge-Detection, image segmentation, image co-registration, image recognition and image denoising.2005, PCNN models are introduced field of face identification by Taymoor M first.Chinese literature on PCNN model recognitions of face is 2011 Year just delivers first, but 2008, PCNN models are applied to recognition of face by domestic Southeast China University Zhang Yudong et al. In research.In recent years, the visual field of domestic scholars is progressed into PCNN models, its application study in terms of recognition of face Also gradually increase.

For being currently based in the face identification method of PCNN models, characteristic is typically from based on model n times iteration Some information of the bianry image of generation, such as duration of ignition sequence and information Entropy sequence, they are a kind of effects of plane Represent, expression is the overall feature of image, it is impossible to the change at edge or point in refinement ground reaction facial image.Therefore document [1] on the basis of simplified PCNN models, a kind of PCNN models that intensity is provided based on pulse be refer to.At the model Reason, each pixel in facial image can obtain its specific pulse and provide intensity, the average pulse that n times iteration is obtained Provide feature of the intensity as facial image, finally by the pulse for calculating facial image to be identified provide intensity matrix and each The average pulse of identity provides the COS distance between intensity matrix to judge its affiliated identity category, but the model does not have Intensity is provided to its pulse to be finely divided.

In actual applications, many uncertain parameters are included in PCNN models, the setting of these model parameters is determined simultaneously The quality of its performance is determined, has generally all been at present, through test of many times, to contrast gained processing result image to select suitable parameter Value.Based on this, existing many documents have been set for research to the automatic of PCNN model parameters.Document [2] is proposed with ash Degree-information content histogram characterizes characteristics of image, and the time in PCNN models of being estimated automatically using information content histogram is decayed ginseng Number, and its setting to PCNN other specifications still needs to set by rule of thumb.

In addition, the intensity distribution difference of face each several part is very big, as Figure 1-5, and traditional face based on PCNN Recognition methods is uniformly processed a width facial image as an entirety, ignores the difference between various pieces.And by testing The discrimination that may certify that the different zones of face be it is different, as shown in Figure 6.Therefore, it can in recognition of face to each portion Set up separately and put different weights.

Document [1]：Computer engineering and design, in January, 2017, the PCNN model parameters based on grid-search algorithms are adaptive Should, Li Han, Yang Xiaofeng etc..

Document [2]：Electronic letters, vol, in May, 2007, a kind of new PCNN model parameter evaluation methods, Zhao Zhijiang, Zhao Chun Sunshine, Zhang Zhihong.

The content of the invention

It is not thin enough that the present invention solves the description of characteristics of image present in the existing face identification method based on PCNN models Change and indistinguishably process view picture facial image using one group of parameter, ignore the different problem of face each several part, there is provided one Plant the intersected human face recognition methods based on weighed intensities PCNN models.

The present invention adopts the following technical scheme that realization：Intersected human face identification side based on weighed intensities PCNN models Method, is realized by following steps：

(1) training stage

Build or select existing containing A people, everyone B width facial images have the face database of A*B width facial images, Take everyone preceding N width in face database, N≤B/2, facial image as training image, by every width training image all according to eye Eyeball, nose and mouth carry out horizontal piecemeal, are divided into the block diagram of upper, middle and lower three for corresponding respectively to eyes, nose and mouth, will be every Three block images of every width facial image of people all weighted input intensity PCNN models, it is strong that each piecemeal can obtain a weighting Degree matrix, can so obtain A*N*3 weighed intensities matrix, and they correspond respectively to each of everyone each training image Piecemeal, then asks the average value of the N number of weighed intensities matrix for corresponding respectively to everyone upper, middle and lower piecemeal, i.e. average weighted intensity Matrix, and this value is recorded, used as the eigenmatrix of everyone each piecemeal, the training stage is always met together and obtains 3A feature square Battle array (average weighted intensity matrix), i.e.,：

(C_{On 1}, C_{On 2}... C_{On A})

(C_{In 1}, C_{In 2}... C_{In A})

(C_{Under 1}, C_{Under 2}... C_{Under A})

Wherein, C represents eigenmatrix, and which people first footmark represents, which piecemeal second footmark represents；

The Mathematical Modeling of described weighed intensities PCNN models is as follows：

F_ij(n)=I_ij (1)

L_ij(n)=∑ WY_kl(n-1) (2)

U_ij(n)=F_ij(n)(1+βL_ij(n)) (3)

θ_ij(n)=exp (- α_θ)θ_ij(n-1)+V_θY_ij(n-1) (4)

O_ij(n)=β F_ij(n)×L_ij(n) (6)

Wherein, to a certain neuron, the implication of parameters is as follows：N loop iteration number of times, [1~+∞)；I and j nerves First coordinate；K and l, coordinate is the coordinate of the peripheral neurons of the neuron of i and j；W inside connection matrix；F_ijFeed back input；I_ij External input signal；U_ijInternal activity；Y_ijPulse is exported, and 0 or 1；θ_ijDynamic threshold, [0~+∞)；Q_ijWeighed intensities； α_θThreshold value attenuation coefficient；V_θThreshold value amplification coefficient；β coefficient of connection；X and y strength formula coefficients；L_ijLink input.

As shown in Figure 8, weighed intensities PCNN models remain as a two-dimentional neutral net for individual layer, the Y1 in figure, Y2 ..., Yn refer to the pulse output of the neuron adjoining with the neuron.

Neuron in model can be received from outside outside input I_ij, while the chain from peripheral neurons can be received Meet input L_ijN (), they are different and change over time, while corresponding dynamic threshold θ_ijN () also changes over time, That is F_ij(n)/θ_ij(n) and O_ij(n)/θ_ijN the value of () is different, the two values are respectively just that intensity is provided in the spontaneous pulse of neuron Intensity, such as table 1 are provided with coupled pulse, is tested from the identification on YALE face databases, both obtain respectively as characteristic The discrimination that arrives is simultaneously different, and weighed intensities situation is when taking O in table_ijN () is used as characteristic, and O_ijN () takes (F_ij(n)+ O_ij(n))/θ_ijThe discrimination obtained when (n), it follows that weighed intensities can as face characteristic data, and can to from Hair pulse is provided intensity and provides the different weight of intensity imparting with coupled pulse, its two-value output Y compared to model_ijN () can be with The description characteristics of image for more refining.

The experimental result contrast of intensity is provided in more than 1 kinds of pulse of table

(2) test phase

A) training image is removed in face database, remaining is test image, everyone has (B-N) width test image, A people is total to There is A* (B-N) width test image, take any one width test image, it is obtained according to the method as the step (1) training stage The weighed intensities matrix of each piecemeal, that is, obtain three weighed intensities matrixes：

QD_{On m}

QD_{In m}

QD_{Under m}

Wherein, QD represents weighed intensities matrix, and first footmark m represents any one width in A* (B-N) width test image, Which piecemeal second footmark represents；

B) each of the width test image (with the test image described in step a)) is calculated according to piecemeal weight calculation method The weight of piecemeal, that is, obtain three weights：

E_{On m}

E_{In m}

E_{Under m}

Wherein, E is the weights of piecemeal；First footmark m represents any one width in A* (B-N) width test image；Second Which piecemeal individual footmark represents；

The specific calculating process of described piecemeal weight calculation method is as follows：

LH (i, j)=H (F (i, j)_r) (8)

F(i,j)_r=I (a, b) | a ∈ [i-r, i+r], b ∈ [j-r, j+r] } (9)

H () is image entropy function, F (i, j)_rIt is the subgraph slided in variable window centered on (i, j), I (a, b) Represent F (i, j)_rInterior pixel, r is the radius for sliding variable window, and (i, j) represents the position of certain pixel, and p and q is piecemeal The horizontal and vertical length of image, cd represents any block in the piecemeal of upper, middle and lower three, and CM is local special type contribution degree, and is adopted With CM as piecemeal weights E；Because by Figure 10-11 as can be seen that local special type contribution degree CM and correspondence portion in formula (10) Divide discrimination Changing Pattern identical, understand that both degrees of correlation are higher according to statistic analysis result, therefore using CM as piecemeal Weights E.

C) calculate consistent in the width test image (with step a) and b)) each piecemeal weighed intensities matrix and step Suddenly the COS distance of the eigenmatrix of everyone the corresponding piecemeal in (), that is, seek both cosine similarities, obtains 3A Cosine similarity：

(S_{On m1}, S_{On m2}... S_{On mA})

(S_{In m1}, S_{In m2}... S_{In mA})

(S_{Under m1}, S_{Under m2}... S_{Under mA})

Wherein, S represents cosine similarity, and first footmark m still represents any one width in A* (B-N) width test image, Second footmark represents the someone in 1 to A, and which piecemeal the 3rd footmark represents；

D) by the cosine similarity obtained by the corresponding section technique of each piecemeal of the width test image and different people with should The corresponding weights of piecemeal are multiplied, i.e.

(S_{On m1}*E_{On m}, S_{On m2}*E_{On m}... S_{On mA}*E_{On m})

(S_{In m1}*E_{In m}, S_{In m2}*E_{In m}... S_{In mA}*E_{In m})

(S_{Under m1}*E_{Under m}, S_{Under m2}*E_{Under m}... S_{Under mA}*E_{Under m})

Following correspondence addition is carried out again, i.e.

S_m1=S_{On m1}*E_{On m}+S_{In m1}*E_{In m}+S_{Under m1}*E_{Under m}

S_m2=S_{On m2}*E_{On m}+S_{In m2}*E_{In m}+S_{Under m2}*E_{Under m}

S_mA=S_{On mA}*E_{On m}+S_{In mA}*E_{In m}+S_{Under mA}*E_{Under m}

E) (S is taken_m1, S_m2... S_mA) in the corresponding people of maximum as the width test image recognition result.

The existing face identification method based on PCNN models is mostly using some based on the intrinsic two-value output of model Used as face characteristic data, characteristic is not enough refined and indistinguishably processes view picture facial image using one group of parameter information, Ignore the difference of face each several part.The present invention proposes the intersected human face identification side based on weighed intensities PCNN models based on this Method.On the basis of PCNN models are simplified, it is proposed that weighed intensities PCNN models, introduce spontaneous pulse and provide intensity, coupling The concept of intensity and weighed intensities, the output of refined model are provided in pulse.Carry out being recognized using piecemeal during recognition of face simultaneously.Know The difference of the not preceding different and local discrimination according to facial image each several part intensity profile, piecemeal has been carried out by facial image. When carrying out recognition of face, the weights of piecemeal can adaptively set according to block image；The parameter of model can be according to block image Content setting (can be according to improved mesh parameter optimizing, document [1]), being then based on above-mentioned parameter each piecemeal can be carried out respectively Identification.The recognition result of the comprehensive each piecemeal of recognition result meeting of final width facial image.

Effect experiment：

1st, intersected human face identification experiment

After employing piecemeal recognition strategy, the discrimination of each face database, less using the recognition of face side of partition strategy Method, is improved to some extent, and is shown in Table 2,3.

The identification experiment of the YALE face databases intersected human face of table 2

The identification experiment of the ORL face databases intersected human face of table 3

2nd, the face recognition experiment based on weighed intensities PCNN models

After the face identification method based on weighed intensities PCNN models is employed, the discrimination of each face database is relatively based on The face identification method of unweighted intensity PCNN models, is improved to some extent, and sees table 4 below, 5.

The YALE face database weighed intensities PCNN model face recognition experiments of table 4

The ORL face database weighed intensities PCNN model face recognition experiments of table 5

3rd, the intersected human face identification experiment based on weighed intensities PCNN models

Experimental result on YALE face databases is shown in Figure 12, it can be seen that discrimination of this paper algorithms when N takes 3,4 and 5 is equal Higher than the face identification method based on ICA, PCA, the amplitude lifted when N takes 3 is maximum.

The experimental result of ORL face databases is shown in Figure 13 (N takes 5), it can be seen that the discrimination on ORL face databases also has accordingly Lifting.

Brief description of the drawings

Fig. 1 is facial image；

Fig. 2 is the overall gray level distribution of face in Fig. 1；

For the grey level histogram of eyes, (wherein transverse axis represents that the intensity profile of eye portion in Fig. 1 is interval to Fig. 3, longitudinal axis table Show the number of pixels of same grayscale value)；

For the grey level histogram of nose, (wherein transverse axis represents that the intensity profile of nasal portion in Fig. 1 is interval to Fig. 4, longitudinal axis table Show the number of pixels of same grayscale value)；

For the grey level histogram of mouth, (wherein transverse axis represents that the intensity profile of face part in Fig. 1 is interval to Fig. 5, and the longitudinal axis is represented The number of pixels of same grayscale value)；

Fig. 6 is the intensity profile and corresponding discrimination of face different zones；

Fig. 7 is the process schematic of the average weighted intensity matrix that any one personal upper piecemeal is sought in step (1)；

Fig. 8 is weighed intensities PCNN model schematics；

Fig. 9 is to seek the process schematic of any one each divided group intensity matrix of width test image in step (2) a)；

Figure 10 is the variation relation figure of local feature contribution degree and local discrimination；

Figure 11 is the matching line chart of local feature contribution degree and local discrimination；

Figure 12 is the experimental result on YALE face databases；

Figure 13 is the experimental result of ORL face databases.

Specific embodiment

Based on the intersected human face recognition methods of weighed intensities PCNN models, realized by following steps：

(1) training stage

Build or select existing containing A people, everyone B width facial images have the face database of A*B width facial images； Take everyone preceding N width in face database, N≤B/2, facial image as training image, by every width training image all according to eye Eyeball, nose and mouth carry out horizontal piecemeal, are divided into upper, lower three block diagrams of neutralization for corresponding respectively to eyes, nose and mouth, will be every Three block images of every width facial image of people all weighted input intensity PCNN models, it is strong that each piecemeal can obtain a weighting Degree matrix, can so obtain A*N*3 weighed intensities matrix, and they correspond respectively to each of everyone each training image Piecemeal, then asks the average value of the N number of weighed intensities matrix for corresponding respectively to everyone upper, middle and lower piecemeal, i.e. average weighted intensity Matrix, and this value is recorded, used as the eigenmatrix of everyone each piecemeal, the training stage is always met together and obtains 3A feature square Battle array (average weighted intensity matrix), i.e.,：

(C_{On 1}, C_{On 2}... C_{On A})

(C_{In 1}, C_{In 2}... C_{In A})

(C_{Under 1}, C_{Under 2}... C_{Under A})

Wherein, C represents eigenmatrix, and which people first footmark represents, which piecemeal second footmark represents；

The Mathematical Modeling of described weighed intensities PCNN models is as follows：

F_ij(n)=I_ij (1)

L_ij(n)=∑ WY_kl(n-1) (2)

U_ij(n)=F_ij(n)(1+βL_ij(n)) (3)

θ_ij(n)=exp (- α_θ)θ_ij(n-1)+V_θY_ij(n-1) (4)

O_ij(n)=β F_ij(n)×L_ij(n) (6)

Wherein, to a certain neuron, the implication of parameters is as follows：N loop iteration number of times, [1~+∞)；I and j nerves First coordinate；K and l, coordinate is the coordinate of the peripheral neurons of the neuron of i and j；W inside connection matrix；F_ijFeed back input；I_ij External input signal；U_ijInternal activity；Y_ijPulse is exported, and 0 or 1；θ_ijDynamic threshold, [0~+∞)；Q_ijWeighed intensities； α_θThreshold value attenuation coefficient；V_θThreshold value amplification coefficient；β coefficient of connection；X and y strength formula coefficients；L_ijLink input；Wherein, it is internal Connection matrix W, threshold value attenuation coefficient α_θ, threshold value amplification coefficient V_θ, coefficient of connection β, the determination of strength formula coefficient x and y, to this It is known for art personnel, can be at least set by the grid optimizing method in document [1], specially：(1) W god Represented through the inverse of distance between unit and neuron, i.e. W=[0.7 1 0.7；1 0 1；0.7 1 0.7]；Threshold value amplifies system Number V_θLarger initial value need to be assigned, while in order to not waste calculating, appropriate takes an integer value, makes it more than facial image Maximum gradation value.(2) assume that two parts weights in formula (7) are that x and y are 1, now need the parameter of setting to only have β and α_θ, at this moment just using the grid search optimizing algorithm in document [1], (application result chooses application and is based on weighed intensities PCNN The recognition accuracy that the intersected human face recognition methods of model is obtained on correspondence test set), obtain β and α_θValue.(3) will obtain β and α_θModel is substituted into, using the grid search optimizing algorithm in document [1] to two parts weights (x and y) in formula (7) Optimizing, obtains x and y.

(2) test phase

A) training image is removed in face database, remaining is test image, everyone has (B-N) width test image, A people is total to There is A* (B-N) width test image, take any one width test image, it is obtained according to the method as the step (1) training stage The weighed intensities matrix of each piecemeal, that is, obtain three weighed intensities matrixes：

QD_{On m}

QD_{In m}

QD_{Under m}

Wherein, QD represents weighed intensities matrix, and first footmark m represents any one width in A* (B-N) width test image, Which piecemeal second footmark represents；

B) the width test image is calculated (with any one width test chart described in step a) according to piecemeal weight calculation method Picture) each piecemeal weight, that is, obtain three weights：

E_{On m}

E_{In m}

E_{Under m}

Wherein, E is the weights of piecemeal；First footmark m represents any one width in A* (B-N) width test image；Second Which piecemeal individual footmark represents；

The specific calculating process of described piecemeal weight calculation method is as follows：

LH (i, j)=H (F (i, j)_r) (8)

F(i,j)_r=I (a, b) | a ∈ [i-r, i+r], b ∈ [j-r, j+r] } (9)

H () is image entropy function, F (i, j)_rIt is the subgraph slided in variable window centered on (i, j), I (a, b) Represent F (i, j)_rInterior pixel, r is the radius for sliding variable window, and (i, j) represents the position of certain pixel, and p and q is piecemeal The horizontal and vertical length of image, cd represents any block in the piecemeal of upper, middle and lower three, and CM is local special type contribution degree, and is adopted With CM as piecemeal weights E；

C) calculate consistent in the width test image (with step a) and b)) each piecemeal weighed intensities matrix and step Suddenly the COS distance of the eigenmatrix of everyone the corresponding piecemeal in (), that is, seek both cosine similarities, obtains 3A Cosine similarity：

(S_{On m1}, S_{On m2}... S_{On mA})

(S_{In m1}, S_{In m2}... S_{In mA})

(S_{Under m1}, S_{Under m2}... S_{Under mA})

Wherein, S represents cosine similarity, and first footmark m still represents any one width in A* (B-N) width test image, Second footmark represents the someone in 1 to A, and the 3rd footmark represents which piecemeal, e.g., S_{On m1}Represent the width test image The cosine similarity of the upper piecemeal of upper piecemeal and first man, S_{On m1}=sim (QD_{On m}, C_{On 1})；

D) by the cosine similarity obtained by the corresponding section technique of each piecemeal of the width test image and different people with should The corresponding weights of piecemeal are multiplied, i.e.

(S_{On m1}*E_{On m}, S_{On m2}*E_{On m}... S_{On mA}*E_{On m})

(S_{In m1}*E_{In m}, S_{In m2}*E_{In m}... S_{In mA}*E_{In m})

(S_{Under m1}*E_{Under m}, S_{Under m2}*E_{Under m}... S_{Under mA}*E_{Under m})

Following correspondence addition is carried out again, i.e.

S_m1=S_{On m1}*E_{On m}+S_{In m1}*E_{In m}+S_{Under m1}*E_{Under m}

S_m2=S_{On m2}*E_{On m}+S_{In m2}*E_{In m}+S_{Under m2}*E_{Under m}

S_mA=S_{On mA}*E_{On m}+S_{In mA}*E_{In m}+S_{Under mA}*E_{Under m}

E) (S is taken_m1, S_m2... S_mA) in the corresponding people of maximum as the width test image recognition result.

Claims (1)

1. a kind of intersected human face recognition methods based on weighed intensities PCNN models, it is characterised in that realized by following steps：

(1) training stage

Build or select existing containing A people, everyone B width facial images have the face database of A*B width facial images, in people Take everyone preceding N width in face storehouse, N≤B/2, facial image as training image, by every width training image all according to eyes, nose Son and mouth carry out horizontal piecemeal, are divided into the block diagram of upper, middle and lower three for corresponding respectively to eyes, nose and mouth, by everyone every Three block images of width facial image all weighted input intensity PCNN models, each piecemeal can obtain a weighed intensities square Battle array, can so obtain A*N*3 weighed intensities matrix, and they correspond respectively to each point of everyone each training image Block, then asks the average value of the N number of weighed intensities matrix for corresponding respectively to everyone upper, middle and lower piecemeal, i.e. average weighted intensity square Battle array, and this value is recorded, used as the eigenmatrix of everyone each piecemeal, the training stage is always obtained 3A eigenmatrix, I.e.：

(C_{On 1}, C_{On 2}... C_{On A})

(C_{In 1}, C_{In 2}... C_{In A})

(C_{Under 1}, C_{Under 2}... C_{Under A})

Wherein, C represents eigenmatrix, and which people first footmark represents, which piecemeal second footmark represents；

The Mathematical Modeling of described weighed intensities PCNN models is as follows：

F_ij(n)=I_ij (1)

L_ij(n)=∑ WY_kl(n-1) (2)

U_ij(n)=F_ij(n)(1+βL_ij(n)) (3)

θ_ij(n)=exp (- α_θ)θ_ij(n-1)+V_θY_ij(n-1) (4)

$Y_{ij}\left(n\right)=\left\{\begin{array}{cc}1& U_{ij}\left(n\right)\&GreaterEqual;{\mathrm{\&theta;}}_{ij}\left(n\right)\\ 0& U_{ij}\left(n\right)<{\mathrm{\&theta;}}_{ij}\left(n\right)\end{array}\right.---\left(5\right)$

O_ij(n)=β F_ij(n)×L_ij(n) (6)

$Q_{ij}\left(n\right)=\underset{1}{\overset{n}{\mathrm{\&Sigma;}}}\&lsqb;x*F_{ij}\left(n\right)/{\mathrm{\&theta;}}_{ij}\left(n\right)+y*O_{ij}\left(n\right)/{\mathrm{\&theta;}}_{ij}\left(n\right)\&rsqb;---\left(7\right)$

Wherein, to a certain neuron, the implication of parameters is as follows：N loop iteration number of times, [1~+∞)；I and j neurons are sat Mark；K and l, coordinate is the coordinate of the peripheral neurons of the neuron of i and j；W inside connection matrix；F_ijFeed back input；I_ijIt is outside Input signal；U_ijInternal activity；Y_ijPulse is exported, and 0 or 1；θ_ijDynamic threshold, [0~+∞)；Q_ijWeighed intensities；α_θThreshold Value attenuation coefficient；V_θThreshold value amplification coefficient；β coefficient of connection；X and y strength formula coefficients；L_ijLink input.

(2) test phase

A) training image is removed in face database, remaining is test image, everyone has (B-N) width test image, the A total A* of people (B-N) width test image, takes any one width test image, obtains its each according to the method as the step (1) training stage The weighed intensities matrix of piecemeal, that is, obtain three weighed intensities matrixes：

QD_{On m}

QD_{In m}

QD_{Under m}

Wherein, QD represents weighed intensities matrix, and first footmark m represents any one width in A* (B-N) width test image, second Which piecemeal individual footmark represents；

B) weight of each piecemeal of the width test image is calculated according to piecemeal weight calculation method, that is, obtains three weights：

E_{On m}

E_{In m}

E_{Under m}

Wherein, E is the weights of piecemeal；First footmark m represents any one width in A* (B-N) width test image；Second angle Which piecemeal mark represents；

The specific calculating process of described piecemeal weight calculation method is as follows：

LH (i, j)=H (F (i, j)_r) (8)

F(i,j)_r=I (a, b) | a ∈ [i-r, i+r], b ∈ [j-r, j+r] } (9)

${\mathrm{CM}}_{cd}=\frac{1}{p\&times;q}\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{q}{\mathrm{\&Sigma;}}}LH(i,j)---\left(10\right)$

H () is image entropy function, F (i, j)_rIt is the subgraph slided in variable window centered on (i, j), I (a, b) is represented F(i,j)_rInterior pixel, r is the radius for sliding variable window, and (i, j) represents the position of certain pixel, and p and q is block image Horizontal and vertical length, cd represents any block in the piecemeal of upper, middle and lower three, and CM is for local special type contribution degree and uses CM As the weights E of piecemeal.

C) the weighed intensities matrix and everyone the corresponding piecemeal in step (1) of each piecemeal of the width test image are calculated Eigenmatrix COS distance, that is, seek both cosine similarities, obtain 3A cosine similarity：

(S_{On m1}, S_{On m2}... S_{On mA})

(S_{In m1}, S_{In m2}... S_{In mA})

(S_{Under m1}, S_{Under m2}... S_{Under mA})

Wherein, S represents cosine similarity, and first footmark m still represents any one width in A* (B-N) width test image, second Individual footmark represents the someone in 1 to A, and which piecemeal the 3rd footmark represents；

D) by the cosine similarity obtained by the corresponding section technique of each piecemeal of the width test image and different people and the piecemeal Corresponding weights are multiplied, i.e.

(S_{On m1}*E_{On m}, S_{On m2}*E_{On m}... S_{On mA}*E_{On m})

(S_{In m1}*E_{In m}, S_{In m2}*E_{In m}... S_{In mA}*E_{In m})

(S_{Under m1}*E_{Under m}, S_{Under m2}*E_{Under m}... S_{Under mA}*E_{Under m})

Following correspondence addition is carried out again, i.e.

S_m1=S_{On m1}*E_{On m}+S_{In m1}*E_{In m}+S_{Under m1}*E_{Under m}

S_m2=S_{On m2}*E_{On m}+S_{In m2}*E_{In m}+S_{Under m2}*E_{Under m}

S_mA=S_{On mA}*E_{On m}+S_{In mA}*E_{In m}+S_{Under mA}*E_{Under m}

E) (S is taken_m1, S_m2... S_mA) in the corresponding people of maximum as the width test image recognition result.