CN106503665A - A kind of face identification method based on synergetic neural network - Google Patents

Abstract

The invention discloses a kind of face identification method based on synergetic neural network, it is used as sample by gathering substantial amounts of facial image, and then the coupling sub-network of synergetic neural network is built carrying out recognition of face, feedback regulation is carried out to the coupling sub-network of synergetic neural network by sample again and makes it have learning ability, the characteristics of improving discrimination to facial image, and have simple.

Description

A kind of face identification method based on synergetic neural network

Technical field

The invention belongs to Intelligent Recognition field, more specifically, is a kind of recognition of face based on synergetic neural network Method.

Background technology

Phase late 1980s, synergetics founder professor Haken propose and for synergy principle to apply to pattern knowledge Other new ideas Synergetic Pattern Recognition, main research soil domestic at present are concentrated on Synergetic Algorithm for Pattern Recognition, including The research of the research of Cooperative Study algorithm, S order parameter meaning and its effectively reconstruct, identification parameter optimization and uneven attention parameters Research, and the application study of Synergetic Pattern Recognition etc..

At present in existing face identification method, mainly there are eigenfaces and neural net method, belong to based on face The recognition methods of global characteristics.Collaboration face identification method is the knowledge from facial image entirety, based on face global characteristics Not, but, collaboration face identification method to different angles, refusing for low resolution face is sincere higher, now multiplex clustering algorithm, The methods such as genetic algorithm, PCA feature extractions are improved, but effect is still unsatisfactory.

Content of the invention

It is an object of the invention to overcoming the deficiencies in the prior art, there is provided a kind of recognition of face based on synergetic neural network Method, by deep learning construction coupling subnet, further to improve the effect of recognition of face.

For achieving the above object, of the invention for a kind of face identification method based on synergetic neural network, its feature It is, comprises the following steps：

(1), IMAQ

Multiple facial images that different experiments person is gathered using camera, and as sample image, while by same experiment Multiple facial images of person are classified as a classification；

(2), Image semantic classification

Sample image is cut out, and makes the resolution ratio of every width sample image identical, then the sample image conversion after cutting out For gray level image, part gray level image is finally selected as master sample, remainder is used as training sample；

(3), synergetic neural network is built

(3.1), initial prototype pattern and kinetics equation are built

Just master sample is column vector according to the grayvalue transition of pixel, column vector is averaged and carries out zero-mean And normalized, obtain initial prototype vector v_k；

By training sample according to pixel grayvalue transition be column vector, column vector is carried out at zero-mean and normalization Vectorial q is obtained after reason；

According to Synergy, by the use of vectorial q as the kinetics equation of mode construction image recognition to be identified：

Wherein, λ_kFor attention parameters, only when it is timing, pattern could be identified；v_kFor initial prototype vector； For v_kOrthogonal adjoint vector；F (t) is fluctuating force；q⁺It is expressed as the adjoint vector of q；Represent single order inverses of the q with regard to the time； K represents that facial image classification, k=k' represent the facial image for specifically taking a certain classification；B, C are respectively constant coefficient；

(3.2), synergetic neural network competition sub-network is built

Introduce S order parameter ξ_k, S order parameter ξ_kRepresent vector q under least square meaning in v_kOn projection, i.e.,：

And then kinetics equation discretization, obtain competing sub-network model：

ξ_k(n+1)-ξ_k(n)=γ (λ_k-D+Bξ_k ²(n))ξ_k(n) (2)

Wherein, n represents that iterations, γ represent iteration step length；

(3.3), the coupling sub-network of synergetic neural network is built

(3.3.1), by initial prototype vector v_kThe formula (1) being updated to vectorial q in step (3.2), is calculated Initial S order parameter ξ_k(0)；

(3.3.2), by initial S order parameter ξ_k(0) formula (2) being updated in (0) step (3.2) carries out dynamic evolution；

To ξ_k(0) progressively carry out n iteration, when certain S order parameter component be equal to 1, remaining S order parameter component be equal to 0 when, repeatedly In generation, stops；

(3.3.3), the S order parameter ξ after n iteration_kN () is judging whether training sample to be identified recognizes correctly, such as Fruit is correct, then train next training sample；If incorrect, the corresponding initial prototype vector vs of the vectorial q_kAsk for Average, and zero averaging normalized is carried out, obtain new prototype vector V_k, repeat step (3.3.1)-(3.3.3)；

(3.3.4) the prototype matrix of new prototype vector composition, after the training of all training samples terminates, is obtained：

V=(V₁,V₂,...V_k)^T

The pseudo inverse matrix of the matrix is sought, the coupling sub-network of synergetic neural network is obtained：

V⁺=(V₁ ⁺,V₂ ⁺,...V_k ⁺)^T

(4), facial image to be measured is recognized using synergetic neural network

(4.1), according to step (1)-(2) methods described collection, process facial image, and as sample to be detected, then will Sample to be detected is processed into column vector q' to be measured according to step (3) methods described；

(4.2), using coupling sub-network V of synergetic neural network⁺Initial sequence ginseng to be measured is calculated with column vector q' to be measured Amount ξ_k(0) ', ξ_k(0) '=V⁺q'；

(4.3), by initial S order parameter ξ to be measured_k(0) ' the formula (2) that is updated in step (3.2) carries out dynamic evolution, When certain S order parameter component is equal to 1, and remaining S order parameter component is equal to 0, end of identification, it is corresponding that its S order parameter component is equal to 1 Prototype vector is recognition result.

The goal of the invention of the present invention is realized in：

A kind of face identification method based on synergetic neural network of the present invention, is used as sample by gathering substantial amounts of facial image This, and then build the coupling sub-network of synergetic neural network to carry out recognition of face, then by sample to synergetic neural network Coupling sub-network carries out feedback regulation and makes it have learning ability, improves the discrimination to facial image, and has simple The characteristics of.

Meanwhile, a kind of face identification method based on synergetic neural network of the present invention also has the advantages that：

(1), by using the coupling sub-network identification facial image of synergetic neural network, it is achieved that Cooperative Mode theory exists Application breakthrough in terms of neutral net, provides another powerful method for neural metwork training.

(2), collaboration face identification method be from facial image entirety, based on the identification of face global characteristics, can Target face is fast and effectively recognized.

(3), the study of synergetic neural network mainly includes choosing prototype pattern and calculates adjoint vector, due to adjoint vector Calculated according to prototype vector, so the selection of prototype vector is most important, adopt a kind of side of feedback regulation here Formula constructs prototype vector so as to more representative, enhances the effect of identification.

Description of the drawings

Fig. 1 is face identification method flow chart of the present invention based on synergetic neural network；

Fig. 2 synergetic neural networks mate subnet model；

Fig. 3 is 3 layers of synergetic neural network structure chart；

Fig. 4 test face specimen discerning processes and recognition result.

Specific embodiment

Below in conjunction with the accompanying drawings the specific embodiment of the present invention is described, so as to those skilled in the art preferably Understand the present invention.Requiring particular attention is that, in the following description, when known function and design detailed description perhaps When can desalinate main contents of the invention, these descriptions will be ignored here.

Embodiment

Fig. 1 is face identification method flow chart of the present invention based on synergetic neural network.

In the present embodiment, as shown in figure 1, a kind of face identification method based on synergetic neural network of the present invention, including Following steps：

(1), IMAQ

Multiple facial images that different experiments person is gathered using camera, and as sample image, while by same experiment Multiple facial images of person are classified as a classification；

In the present embodiment, the facial image in orl storehouses is taken, each experimenter totally 10 facial images, totally 40 experiments 400 facial images of person, and using this 400 facial images all as sample image.

(2), Image semantic classification

Sample image is cut out, and makes the resolution ratio of every width sample image identical, then the sample image conversion after cutting out For gray level image, part gray level image is finally selected as master sample, remainder is used as training sample；

In the present embodiment, 400 sample images are cut out as 68*58 pixels, then 400 colored samples after cutting out This image is converted to gray-scale map, because gray level image still shows the entirety of whole image and Local textural feature, bright Degree, contrast etc., but greatly reducing amount of calculation；Select at random 40 people everyone 2 totally 80 as master sample, per People 6 totally 240 as training sample, remaining everyone 2 totally 80 as test sample.

(3), synergetic neural network is built

(3.1), initial prototype pattern and kinetics equation are built

Just master sample is column vector according to the grayvalue transition of pixel, column vector is averaged and carries out zero-mean And normalized, obtain initial prototype vector v_k；

By training sample according to pixel grayvalue transition be column vector, column vector is carried out at zero-mean and normalization Vectorial q is obtained after reason；

According to Synergy, by the use of vectorial q as the kinetics equation of mode construction image recognition to be identified：

Wherein, λ_kFor attention parameters, only when it is timing, pattern could be identified；v_kFor initial prototype vector； For v_kOrthogonal adjoint vector；F (t) is fluctuating force；q⁺It is expressed as the adjoint vector of q；Represent single order inverses of the q with regard to the time； K represents that facial image classification, k=k' represent the facial image for specifically taking a certain classification；B, C are respectively constant coefficient；

In the present embodiment, facial image classification k=40, k' are the number in 1～40；

(3.2), synergetic neural network competition sub-network is built

Introduce S order parameter ξ_k, S order parameter ξ_kRepresent vector q under least square meaning in v_kOn projection, i.e.,：

And then kinetics equation discretization, obtain competing sub-network model：

ξ_k(n+1)-ξ_k(n)=γ (λ_k-D+Bξ_k ²(n))ξ_k(n) (2)

Wherein, n represents that iterations, γ represent iteration step length；

(3.3), the coupling sub-network of synergetic neural network is built

(3.3.1), by initial prototype vector v_kThe formula (1) being updated to vectorial q in step (3.2), is calculated Initial S order parameter ξ_k(0)；

(3.3.2), by initial S order parameter ξ_k(0) formula (2) being updated in step (3.2) carries out dynamic evolution；

To ξ_k(0) progressively carry out n iteration, when certain S order parameter component be equal to 1, remaining S order parameter component be equal to 0 when, repeatedly In generation, stops；

(3.3.3), the S order parameter ξ after n iteration_kN () whether this just recognizes to judge training sample (n) to be identified Really, if correctly, the next training sample of training；If incorrect, the corresponding initial prototype vector vs of the vectorial q_k Average is asked for, and carries out zero averaging normalized, obtain new prototype vector V_k, repeat step (3.3.1)- (3.3.3)；

Wherein, using S order parameter ξ_kN () whether this recognizes that correct determination methods are to judge training sample (n) to be identified：

After n iteration, by ξ_kN the classification corresponding to the S order parameter component of ()=1 is carried out with the class belonging to sample to be identified Relatively, if same category, then recognize correctly, on the contrary identification mistake.

(3.3.4) the prototype matrix of new prototype vector composition, after the training of all training samples terminates, is obtained：

V=(V₁,V₂,...V_k)^T

The pseudo inverse matrix of the matrix is sought, the coupling sub-network of synergetic neural network as shown in Figure 2 is obtained：

V⁺=(V₁ ⁺,V₂ ⁺,...V_k ⁺)^T

In the present embodiment, neutral net as shown in Figure 3 can be constructed；Wherein, in input layer, the unit of input layer is received The i-th class sample q of all kinds of pattern vector q to be identified_i(0), q is the gray scale that the first step trains face pattern in storehouse in the present invention The column vector of the digital picture that figure is obtained after matrixing.

Intermediate layer represents all kinds of prototype vector neuron v_k, during the network operation, according to v_kAdjoint vector v_k ⁺It is obtained just Beginning S order parameter ξ_k(0), network according to the kinetics equation operation in competition subnet and develops, and takes iterations n=30 here, repeatedly Ride instead of walk long γ=1, ξ_k(0) end-state ξ is reached with the development of time_k(n).Q is obtained using the S order parameter_i(n).

In output layer, the pattern of output layer can be expressed as q_i(n)=ξ_k(n)V^T, q_iN () is the i-th class sample q_i(0) evolution n End-state after secondary.Relatively q_i(n) and q_i(0) if same class then goes successively to input layer and reads sample q, if not same Class then returns to intermediate layer, q_i(0) corresponding v_kAverage as new prototype vector, until sample q reads finish, i.e., Complete the structure of synergetic neural network.

(4), facial image to be measured is recognized using synergetic neural network

(4.1), according to step (1)-(2) methods described collection, process facial image, and as sample to be detected, then will Sample to be detected is processed into column vector q' to be measured according to step (3) methods described；

(4.2), using coupling sub-network V of synergetic neural network⁺Initial sequence ginseng to be measured is calculated with column vector q' to be measured Amount ξ_k(0) ', ξ_k(0) '=V⁺q'；

(4.3), by initial S order parameter ξ to be measured_k(0) ' the formula (2) that is updated in step (3.2) carries out dynamic evolution, When certain S order parameter component is equal to 1, and remaining S order parameter component is equal to 0, end of identification, it is corresponding that its S order parameter component is equal to 1 Prototype vector is recognition result.

In the present embodiment, use the piece image in test sample storehouse as test sample, such as Fig. 4 (a), using collaboration Coupling sub-network V of neutral net⁺Obtain initial S order parameter ξ_k(0) ', enter action mechanical equation and developed；S order parameter is discrete Change curve such as Fig. 4 (b), represent the S order parameter ξ of target image_k(0) '=0.4, after 10 step of iteration, its value gradually levels off to 1；Identification Process such as Fig. 4 (c), target image are gradually approached to prototype；Recognition result such as Fig. 4 (d).

Although being described to illustrative specific embodiment of the invention above, in order to the technology of the art Personnel understand the present invention, it should be apparent that the invention is not restricted to the scope of specific embodiment, the common skill to the art For art personnel, as long as various change is in appended claim restriction and the spirit and scope of the present invention for determining, these Change is it will be apparent that all utilize the innovation and creation of present inventive concept in the row of protection.

Claims (2)

1. a kind of face identification method based on synergetic neural network, it is characterised in that comprise the following steps：

(1), IMAQ

Multiple facial images that different experiments person is gathered using camera, and as sample image, while by same experimenter's Multiple facial images are classified as a classification；

(2), Image semantic classification

Sample image is cut out, and makes the resolution ratio of every width sample image identical, then the sample image after cutting out is converted into ash Degree image, finally selects part gray level image as master sample, and remainder is used as training sample；

(3), synergetic neural network is built

(3.1), initial prototype pattern and kinetics equation are built

Just master sample is column vector according to the grayvalue transition of pixel, averages and carries out zero-mean to column vector and return One change is processed, and obtains initial prototype vector v_k；

By training sample according to pixel grayvalue transition be column vector, column vector is carried out after zero-mean and normalized Obtain vectorial q；

According to Synergy, by the use of vectorial q as the kinetics equation of mode construction image recognition to be identified：

$\stackrel{\·}{q}=\underset{k}{\Σ}{\mathrm{\λ}}_k{v}_k\left(v_k^{+}q\right)-B\underset{k=k^{\′}}{\Σ}{\left(v_{k^{\′}}^{+}q\right)}^2\left(v_k^{+}q\right)v_k-C\left(q^{+}q\right)q+F\left(t\right)$

Wherein, λ_kFor attention parameters, only when it is timing, pattern could be identified；v_kFor initial prototype vector；For v_k Orthogonal adjoint vector；F (t) is fluctuating force；q⁺It is expressed as the adjoint vector of q；Represent single order inverses of the q with regard to the time；K tables Face image of leting others have a look at classification, k=k' represent the facial image for specifically taking a certain classification；B, C are respectively constant coefficient；

(3.2), synergetic neural network competition sub-network is built

Introduce S order parameter ξ_k, S order parameter ξ_kRepresent vector q under least square meaning in v_kOn projection, i.e.,：

${\mathrm{\ξ}}_k=v_k^{+}q;---\left(1\right)$

And then kinetics equation discretization, obtain competing sub-network mould：

${\mathrm{\ξ}}_k(n+1)-{\mathrm{\ξ}}_k\left(n\right)=\mathrm{\γ}({\mathrm{\λ}}_k-D+{{\mathrm{B\ξ}}_k}^2(n\left)\right){\mathrm{\ξ}}_k\left(n\right)---\left(2\right)$

$D=(B+C)\underset{k^{\′}}{\Σ}{{\mathrm{\ξ}}_{k^{\′}}}^2\left(n\right)$

Wherein, n represents that iterations, γ represent iteration step length；

(3.3), the coupling sub-network of synergetic neural network is built

(3.3.1) initial prototype vector v and vector q are updated to the formula (1) in step (3.2), initial sequence is calculated Parameter ξ_k(0)；

(3.3.2), by initial S order parameter ξ_k(0) formula (2) being updated in step (3.2) carries out dynamic evolution；

To ξ_k(0) progressively carry out n iteration, when certain S order parameter component be equal to 1, remaining S order parameter component be equal to 0 when, iteration is stopped Only；

(3.3.3), the S order parameter ξ after n iteration_kN () is judging whether training sample to be identified is recognized correctly, if just Really, then next training sample is trained；If incorrect, the corresponding initial prototype vector vs of the vectorial q_kAsk for average, And zero averaging normalized is carried out, obtain new prototype vector V_k, repeat step (3.3.1)-(3.3.3)；

(3.3.4) the prototype matrix of new prototype vector composition, after the training of all training samples terminates, is obtained：

V=(V₁,V₂,...V_k)^T

The pseudo inverse matrix of the matrix is sought, the coupling sub-network of synergetic neural network is obtained：

V⁺=(V₁ ⁺,V₂ ⁺,…V_k ⁺)^T

(4), facial image to be measured is recognized using synergetic neural network

(4.1), according to step (1)-(2) methods described collection, process facial image, and as sample to be detected, then will be to be checked Test sample sheet is processed into column vector q' to be measured according to step (3) methods described；

(4.2), coupling sub-network V+ and column vector q' to be measured using synergetic neural network calculates initial S order parameter ξ to be measured_k (0) ', ξ_k(0) '=V+q'；

(4.3), by initial S order parameter ξ to be measured_k(0) ' the formula (2) that is updated in step (3.2) carries out dynamic evolution, when certain Individual S order parameter component is equal to 1, when remaining S order parameter component is equal to 0, and end of identification, its S order parameter component are equal to 1 corresponding prototype Vector is recognition result.

2. a kind of face identification method based on synergetic neural network according to claim 1, it is characterised in that the step Suddenly in (3.3.3), using S order parameter ξ_kN () is judging whether training sample to be identified recognizes that correct determination methods are：

After n iteration, by ξ_kN the classification corresponding to the S order parameter component of ()=1 is compared with the class belonging to sample to be identified, If same category, then recognize correctly, on the contrary identification mistake.