CN101710382A - Gabor human face recognizing method based on simplified intelligent single-particle optimizing algorithm - Google Patents

Abstract

The invention relates to a Gabor human face recognizing method based on a simplified intelligent single-particle optimizing algorithm, comprising a simplified intelligent single-particle optimizing algorithm and a link of selecting a Gabor filter group by the simplified intelligent single-particle optimizing algorithm, wherein the simplified intelligent single-particle optimizing algorithm comprises the following steps of: carrying out optimizing search on a solution space of a problem function by a particle and respectively carrying out intelligent updating on all dimensionality components of the single particle during iteration; and the link of selecting the Gabor filter group by the simplified intelligent single-particle optimizing algorithm comprises the following steps of: constructing a particle structure according to the quantity of Gabor filters, determining a particle search range and using a Fisher criterion as a fitness function; during the human face recognition, carrying out characteristic extraction on a human face image conforming to the authentication to form a characteristic database; after a new human face image is input, carrying out the characteristic extraction on the human face image by the selected Gabor filter; and comparing the obtained characteristic vector with the characteristics in the database by a minimal adjacent classifier to judge whether the input image is contained in the database or not and a corresponding identity. The invention improves the recognition precision and the characteristic extracting pertinence.

Description

Gabor face identification method based on simplified intelligent single-particle optimizing algorithm

Technical field

The present invention relates to facial image recognition technology field, more particularly, relate to a kind of Gabor face identification method based on simplified intelligent single-particle optimizing algorithm.

Background technology

Along with the social informatization degree improves constantly, the demand of personal identification work is increased day by day, original recognition methods can't have been satisfied the requirement of degree of accuracy and convenient degree.And then personal identification method based on facial image appearred.Current face identification method mainly is divided into two classes: (1) is discerned by the integrity attribute of research facial image pattern based on the recognition methods of one-piece pattern; (2) based on the recognition methods of signature analysis, relative ratios and other of people's face reference point are described the constitutive characteristic vectors such as parameter of face feature, discern by comparison.Practical face recognition algorithms need satisfy the requirement on accuracy of identification and the algorithm complex, promptly needs to promote discrimination on the one hand, reduces misclassification rate; Need computing enough simple on the other hand, can on limited resources such as embedded platform, realize.

Optimization problem is to seek that specific function is worth most and a class mathematical problem of correspondence position.Biological heuristic optimized Algorithm is a series of intelligent optimal value searching method that grows up by the simulating nature zone phenomenon, can effectively solve the optimization problem of using general mathematical tool to handle.Common biological heuritic approach such as ant group algorithm, genetic algorithm etc., or because search speed is slow, or, be difficult to use in the actual optimization problem that solves because the optimizing performance is not good.The particle swarm optimization algorithm that proposes by the simulation flock of birds group intelligent behavior of looking for food in recent years, simple and clear with its notion, performance efficient, be easy to advantage such as realizations, obtained paying attention to widely and application.

The face recognition algorithms relevant with present technique is for (the Gabor small echo of one dimension is proposed in nineteen forty-six by Gabor D based on the Gabor small echo, then Daugman J proposed two-dimentional Gabor small echo in 1985) and the method for PCA (Principal Component Analysis is called for short PCA) dimensionality reduction extraction feature.The method can obtain one of algorithm of better recognition effect in proposing in recent years for generally acknowledging.Its principle of work is as follows:

At first, facial image and Gabor wavelet filter group are distinguished convolution to extract feature.As shown in Figure 1, the Gabor bank of filters is made up of 40 measure-alike two-dimensional Gabor small echo matrixes, includes 8 directions and 5 yardsticks.Thereby can effectively extract the whole characteristic information of face image on different directions and yardstick.

Facial image and this 40 Gabor small echos are carried out the two-dimensional convolution computing respectively, obtain 40 measure-alike two dimensional character matrixes.For reducing calculated amount, generally facial image and Gabor bank of filters can be transformed into frequency domain by the Two-dimensional FFT conversion respectively, thereby convolutional calculation is converted into matrix multiplication in actual applications, again matrix of consequence be changed back temporal signatures by the IFFT conversion.

Then, the eigenmatrix group is extracted for discerning personnel selection face eigenvector.As shown in Figure 2, extract the eigenmatrix that obtains through the Gabor wave filter, reset in regular turn and classify the one-dimensional characteristic vector as for each.

Thereby the full feature matrix group will be converted into 40 one-dimensional characteristic vectors.These vectors are joined end to end, form single long eigenvector, and adopt PCA mapping carrying out dimension-reduction treatment.The important component in the feature can be effectively obtained in the PCA computing, gives up non-important component, thereby reduces vector length.Pass through down-sampling at last, obtain the lower identification personnel selection face eigenvector of length.

At last, eigenvector is used for recognition of face.The facial image that meets authentication is carried out Gabor respectively extract, obtain identification eigenvector composition characteristic database.When discerning, facial image to input carries out feature extraction, each vector divides other mahalanobis distance in the eigenvector of calculating acquisition and the database then, carry out identification decision by minimum neighbour's sorter: if the minimum mahalanobis distance of itself and each eigenvector is not more than predefined threshold values, then input picture is identified as the corresponding identity of the feature that obtains this minor increment, is not included in the database otherwise be judged to be face characteristic.

Although it is higher that use Gabor small echo and PCA dimensionality reduction extraction feature are carried out the method precision of recognition of face, two major defects are arranged:

One, algorithm complex are too high, have a strong impact on practical value.Use the Gabor bank of filters facial image to be carried out in the computing of feature extraction, need carry out 40 Two-dimensional FFT and IFFT conversion, and the phase multiplication of 40 complex matrixs.Embedded platform is difficult to satisfy its calculation requirement.Even use PC as implementation platform, arithmetic speed is still slow on the one hand, and it is also too high to discern cost on the other hand.

Its two, it is not high to use the Gabor small echo that whole face image is carried out the feature extraction specific aim.40 Gabor small echos be chosen for experiment experience design, do not carry out special optimization, its feature extraction inefficiency at recognition of face.The method that the pattern of the whole face image of dependence is discerned can't be distinguished the characteristic information content and the significance level to discerning thereof of human facial zones of different, thus the easier influence that is interfered.

Summary of the invention

The technical problem to be solved in the present invention is, proposes a kind of Gabor face identification method based on simplified intelligent single-particle optimizing algorithm, uses Gabor small echo and PCA dimensionality reduction extraction feature to carry out the existing defective of recognition of face in the prior art to solve.

The technical solution adopted for the present invention to solve the technical problems is: construct a kind of Gabor face identification method based on simplified intelligent single-particle optimizing algorithm, comprising:

A kind of Gabor face identification method based on simplified intelligent single-particle optimizing algorithm is characterized in that, comprising:

Simplified intelligent single-particle optimizing algorithm comprises and adopts a particle that the solution space of problem function is carried out the optimizing search, and each dimension component of single-particle is carried out intelligent updating respectively during iteration; And

Use simplified intelligent single-particle optimizing algorithm to select the Gabor bank of filters, comprise according to the Gabor number of filter being configured to particle structure, determine the scope of particle search and use the Fisher criterion as fitness function;

When carrying out recognition of face, the facial image that meets authentication is carried out feature extraction form property data base, and behind the new facial image of input, use the described Gabor bank of filters of selecting that this facial image is carried out eigenvalue extracting with the composition characteristic vector, use in minimum neighbour's sorter and the database feature to compare to the eigenvector that obtains, whether be contained in the database and the identity of correspondence to judge input picture;

Wherein, each wave filter forms an eigenwert, and the eigenwert that the whole filter group produces is arranged and formed the one-dimensional characteristic vector.

Among the present invention, the particle of described simplified intelligent single-particle optimizing algorithm more new formula is:

$v_i^{n+1}=(a\&times;{\mathrm{<figure-callout\; id="1"\; label="r"\; filenames="H2009101886678E00012.png"\; state="\{\{state\}\}">r</figure-callout>}}_1/n^{p\&times;r_2})+2{\&times;L}_i^n---\left(1\right)$

$x_i^{n+1}=\left\{\begin{array}{cc}{x}_i^n+v_i^{n+1}& \mathrm{if\; f}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ x_i^n& \mathrm{if\; f}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(2\right)$

$L_i^{n+1}=\left\{\begin{array}{cc}{v}_i^{n+1}& \mathrm{if\; f}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ L_i^n& \mathrm{if\; f}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(3\right)$

Wherein a is the diversity factor, and p is a descending factors, and v is a velocity, x is a position vector, and L is study factor vector, and n is the iterations of current dimension in the formula, i is the dimension of the current renewal of vector, and f () is a fitness function, and r1 and r2 are with each two random values that produce of iterations n.

Among the present invention,

More in the new formula (1), next iteration particle's velocity vector is determined by diversity part and study part: the study part is formed for learning information factor L adds inertial coefficient 2, the diversity part is then by diversity factor a and descending factors p decision, represented the search of particle to attempt behavior, diversity partly is to be the power decreasing function of variable with iterations n, and its variation makes particle change Local Search gradually into by the global search at iteration initial stage;

More in the new formula (3), result's decision that the variation of the study factor is attempted by this iteration: if the position after this iteration is upgraded is better than the position of last iteration, the speed knowledge of then drawing this is as the study factor, otherwise the maintenance last iteration uses the study factor that obtains better result.

Among the present invention,

The key parameter of decision search performance is diversity factor a and descending factors p;

Wherein a determines the step-length that particle search is attempted, the rate of regression of p decision particle search step-length;

When particle tended to global search, needs increased a value and reduce the p value with the bigger trial scope of acquisition, otherwise when tending to Local Search, then needed to reduce a value and increase the p value to carry out precise search and to avoid crossing the optimal value zone.

Among the present invention, set diversity factor a and these two key parameter factors of descending factors p adaptive updates that carries out with iteration, more new formula is as follows for it:

$a^{k+1}=\left\{\begin{array}{cc}{a}^k-\frac{(a_{\max }-a_{\min })}{K}& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ a^k+\frac{(a_{\max }-a_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(4\right)$

$p^{k+1}=\left\{\begin{array}{cc}{p}^k+\frac{(p_{\max }-p_{\min })}{K}& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ p^k-\frac{(p_{\max }-p_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(5\right)$

$t^{k+1}=\left\{\begin{array}{cc}0& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ t^k+1& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(6\right)$

Wherein, k is current whole number of iterations in the formula, and K is the largest global iterations, and max is a maximal value, and min is a minimum value, and t is the disturbance factor.

Among the present invention,

Be [xmin, xmax] if establish the hunting zone of particle on each dimension, then a variation range is [amin=xmin * 0.01, amax=xmax * 100] in the formula, and the p variation range is [pmin=3, pmax=30], and t is set has r (t for the disturbance factor ^k)=uniform (0,2 * t ^k) be absorbed in stagnation with the variation that prevents these two key parameters of a, P.

Among the present invention, the step of described simplified intelligent single-particle optimizing algorithm comprises:

S1, initialization particle position x also calculates its initial adaptive value, initiation parameter a=a _Max, p=p _Min, t=0;

S2, the whole iterations k=1 of initialization;

S3, initialization dimension label i=1, and the study factor is set $L_i^0=0;$

S4, initialization dimension iterations n=1;

S5, according to formula (1), particle rapidity, position and the study factor are upgraded in (2), (3);

S6 is if n less than maximum dimension iterations N, then forwards S4 to;

S7 is if i is less than maximum dimension label i _Max, then forward S3 to;

S8, according to formula (4), (5), (6) undated parameter a, p and t;

S9 is if k less than largest global iterations K, then forwards S2 to.

Among the present invention, described Gabor wave filter is used to carry out feature extraction, and each described Gabor wave filter contains 7 parameters, comprise: 5 parameters of the form of two-dimensional Gabor small echo, wavelength L, angle T, phase place P, aspect ratio G and bandwidth B, and the coordinate X and the Y that act on people's face position.

Among the present invention, the hunting zone of each described Gabor wave filter is as follows:

The hunting zone of X is [0, PicSizeX], and the hunting zone of Y is [0, PicSizeY], the hunting zone of L be [2, MIN (PicSizeX, PicSizeY)], the hunting zone of T is [0, π], the hunting zone of P is [π, π], and the hunting zone of G is [0.2,1], the plain scope of searching of B is [0.4,2.5], and wherein PicSizeX and PicSizeY are the size of input facial image.

Among the present invention, when facial image is carried out feature extraction, be the center with people's specified point on the face, the matrix and the Gabor wave filter dot product of the image identical with Gabor wave filter size, each wave filter will obtain an eigenwert.

The invention has the beneficial effects as follows, the present invention is applied to the heuristic optimized Algorithm of biology among the face identification method first, proposed to solve the SISPO algorithm of actual optimization problem on the one hand, on the other hand by in the selection and optimization of the SISPO algorithm being introduced the Gabor wave filter, the wave filter of calculation of complex is selected work and general recognition of face work is separated, and used recognition methods based on signature analysis, thereby accuracy of identification and feature extraction specific aim have been promoted, reduced computation complexity significantly, can on embedded platform, realize at an easy rate, effectively reduce the identification cost.

Description of drawings

The invention will be further described below in conjunction with drawings and Examples, in the accompanying drawing:

Fig. 1 is that prior art uses Gabor wavelet filter group to extract the synoptic diagram of face characteristic;

Fig. 2 is the synoptic diagram that prior art is arranged as eigenmatrix one n dimensional vector n;

Fig. 3 is the synoptic diagram of Gabor bank of filters structure particle of the present invention;

Fig. 4 is that the present invention uses the Gabor bank of filters to extract the synoptic diagram of privileged site feature.

Embodiment

Further understand and understanding for making architectural feature of the present invention and the effect reached had, cooperate detailed explanation, be described as follows in order to preferred embodiment and accompanying drawing:

Computation complexity when problem to be solved by this invention uses the Gabor bank of filters to extract face characteristic for reducing promotes recognition efficiency, improves the robustness of Verification System, enables effectively to realize on lower-cost embedded platform.By improved particle swarm optimization algorithm being applied to the Gabor bank of filters and being used in choosing of people's face particular location, improve the specific aim of feature extraction.Simultaneously Feature Selection work and identifying are separated, the optimization that computational complexity is high is only chosen needs operation once, and computation complexity significantly reduces when using identification.Thereby form face recognition technology with Practical significance.

The technical scheme of Gabor face identification method of the present invention mainly comprises two aspects: improved particle swarm optimization algorithm, and use to improve the back algorithm and the method for using the Gabor bank of filters to carry out recognition of face is made amendment and optimize.

(1) particle swarm optimization algorithm is improved.

Original particle swarm optimization algorithm is absorbed in the local optimum point easily when solving challenge, and precocious convergence phenomenon occurs.Particularly when the dimension of problem function increases, the optimizing performance of algorithm will obviously descend.This makes original particle cluster algorithm be difficult to effectively be applied to recognition of face to optimize in this higher-dimension complicated problems.

For addressing this problem, the present invention has used improved simplified intelligent single-particle optimizing algorithm (Simplified Intelligence Single Particle Optimization is called for short SISPO).Be compared to original particle swarm optimization algorithm, the SISPO algorithm has following characteristics:

Only use single particle to carry out the optimizing iteration, rather than whole population is searched for.

During iteration each dimension component of single-particle is carried out intelligent updating respectively, rather than all dimension is upgraded identical value simultaneously.

Introduce study factor L, comprise diversity part and study part, can effectively carry out optimizing.

Optimize performance and be better than general particle cluster algorithm, particularly effect is more obvious on the high problem of complexity.

Key parameter diversity factor a and descending factors p need not preestablish, and it is worth self-adaptation adjustment in iteration.

In the SISPO algorithm, only use a particle that the solution space of problem function is carried out the optimizing search.This is because when using population to search for, the search of each particle trend depends on three parts: regional learning information was searched for by learning information and whole particle colony that the inertia of trend, particle self search for the zone.In optimizing,, can think that its position is global optimum by mistake, thereby whole colony is sent out wrong learning information, influence the search performance of algorithm on the contrary if there is some particle to sink into the part when being worth most.This is worth on a little numerous optimization problems particularly evident most in the part; the information regular meeting of putting immediate particle apart from global optimum in the colony is sunk into the information that other particle on the local value point sends and is obscured; make colony lose direction, cause optimizing iteration result to be tending towards stagnating.And when using the single particle search, owing to there is not the interference of other particle, this problem can not appear fully.

But because single-particle has lacked colony's learning information, the SISPO algorithm needs new mechanism and guarantees its search performance.One of improvement is except that the whole iteration of particle, and the value on each dimension of particle position vector is all carried out iterative search respectively, promptly all carries out optimizing on each dimension respectively and attempts.Another improvement is to have constructed brand-new particle more new formula is as follows:

$v_i^{n+1}=(a\&times;{\mathrm{<figure-callout\; id="1"\; label="r"\; filenames="H2009101886678E00012.png"\; state="\{\{state\}\}">r</figure-callout>}}_1/n^{p\&times;r_2})+2{\&times;L}_i^n---\left(1\right)$

$x_i^{n+1}=\left\{\begin{array}{cc}{x}_i^n+v_i^{n+1}& \mathrm{if\; f}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ x_i^n& \mathrm{if\; f}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(2\right)$

$L_i^{n+1}=\left\{\begin{array}{cc}{v}_i^{n+1}& \mathrm{if\; f}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ L_i^n& \mathrm{if\; f}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(3\right)$

Wherein v is a velocity, and x is a position vector, and L is study factor vector.N is the iterations of current dimension in the formula, and i is the dimension of the current renewal of vector, and f () is a fitness function, and r1 and r2 are with each two random values that produce of iterations n.By new formula (1) more as can be seen, next iteration particle's velocity vector is by diversity part and the decision of study part.The study part is that learning information factor L adds inertial coefficient 2 compositions, the knowledge of having represented particle to obtain from past experience.According to current update mode, particle's velocity when the study partial dynamic is adjusted next iteration promotes particle trend problem optimum point.On behalf of the search of particle, the diversity part attempt behavior then by diversity factor a and descending factors p decision.Diversity partly is to be the power decreasing function of variable with iterations n, and its variation makes particle change Local Search gradually into by the global search at iteration initial stage.Make particle have certain irregular behavior by adding the random value disturbance, thereby avoid it to be absorbed in the local optimum point.By new formula (3) more as can be seen, result's decision that the variation of the study factor is attempted by this iteration: if the position after this iteration is upgraded is better than the position of last iteration, the speed knowledge of then drawing this is as the study factor, otherwise the maintenance last iteration uses the study factor that obtains better result.

By new formula more as can be seen, the key parameter of decision search performance is diversity factor a and descending factors p.Wherein a has determined the step-length that particle search is attempted, and p has determined the rate of regression of particle search step-length.When particle tends to global search, should increase a value and reduce the p value to obtain bigger trial scope.Otherwise when tending to Local Search, then need to reduce a value and increase the p value to carry out precise search and to avoid crossing the optimal value zone.In SISPO, set the carry out adaptive updates of these two key parameter factors, thereby do not need artificial special the setting with iteration.More new formula is as follows for it:

$a^{k+1}=\left\{\begin{array}{cc}{a}^k-\frac{(a_{\max }-a_{\min })}{K}& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ a^k+\frac{(a_{\max }-a_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(4\right)$

$p^{k+1}=\left\{\begin{array}{cc}{p}^k+\frac{(p_{\max }-p_{\min })}{K}& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ p^k-\frac{(p_{\max }-p_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(5\right)$

$t^{k+1}=\left\{\begin{array}{cc}0& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ t^k+1& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(6\right)$

Be [xmin, xmax] if establish the hunting zone of particle on each dimension, then a variation range is [amin=xmin * 0.01, amax=xmax * 100] in the formula, and the p variation range is [pmin=3, pmax=30].K is current whole number of iterations in the formula, and K is the largest global iterations.T is set r (t is arranged for the disturbance factor ^k)=uniform (0,2 * t ^k), effect is that the variation that prevents two key parameters is absorbed in stagnation.The adaptive updates formula shows, when this whole iteration can't obtain better as a result the time, the diversity part will more be tending towards enlarging the hunting zone seeking other optimum points, otherwise will more be tending towards fine search to seek the optimal location in the one's respective area.

The SISPO algorithm flow is as shown in table 1:

Table 1.SISPO algorithm flow

(2) to improvement based on the face identification method of Gabor bank of filters.

Of the present invention being improved to selected specific Gabor wave filter the ad-hoc location of facial image carried out dot-product operation to extract the characteristic information at this position, by the Gabor wave filter that uses one group of process to select each significant points of people's face is carried out feature extraction, reduce computation complexity, the method for improving recognition efficiency.

At first need to use the SISPO algorithm to select and optimize the Gabor bank of filters and act on the position of people's face respectively.The form of two-dimensional Gabor small echo is determined by 5 parameters: wavelength Lambda (L), angle Theta (T), phase place Phi (P), aspect ratio Gamma (G) and bandwidth B andwidth (B).Add that it acts on the coordinate X and the Y of people's face position, each Gabor wave filter that is applied to feature extraction contains 7 parameters.When using the SISPO algorithm to select the Gabor bank of filters, at first definite number of filter that needs, then construct particle structure as shown in Figure 3:

The hunting zone of each filter parameter is as shown in table 2:

Table 2. hunting zone is provided with

Wherein PicSizeX and PicSizeY are the size of input facial image.

After obtaining to be added into the facial image of database, use the fitness function f () of Fisher criterion: establish if contain m people in the database as optimization, everyone comprises n and opens face-image, after using the Gabor bank of filters to extract eigenvector, then distance should be as far as possible little in the class of the eigenvector of same people's image acquisition, and the between class distance of the eigenvector that the Different Individual image obtains should be big as far as possible, distance should be the smaller the better divided by the value of between class distance in the class, is a minimum value optimization problem.Use the SISPO algorithm this to be optimized search, finally obtain one group of Gabor wavelet filter and extract the position of feature respectively at facial image.

The Fisher criterion is the typical method in the traditional mode recognition methods, it emphasizes to regard the product of normal vector in the linear equation and sample the projection of sample vector on unit normal vector as, if can accomplish that the projection of inhomogeneous sample on normal vector presents in the class and assemble, class is to the effect of separating, and is then favourable to reducing misclassification.The best normal vector calculating formula of gained is S _w ^-1(m ₁-m ₂).The Bayesian decision result of this result and normal distribution covariance matrix etc. is similar, if these explanation two classes distribute around average is close really separately, the Fisher criterion can make error rate less.

As shown in Figure 4, when facial image is carried out feature extraction, be the center with people's specified point on the face, the matrix and the Gabor wave filter dot product of the image identical with Gabor wave filter size, then each wave filter will obtain an eigenwert.These eigenwerts are arranged as eigenvector promptly can be used for identification, the dimension of eigenvector is the number of Gabor wave filter.Owing to be that the high critical positions of face characteristic information amount is extracted, rather than whole face image is carried out computing remake the dimensionality reduction operation, therefore the eigenvector that obtains has more specific aim.In addition since computing slowly wave filter choose work and only need carry out once, be separated with the identification work in later stage.And only need to carry out dot-product operation rather than original convolution algorithm when carrying out recognition of face, also removed the calculated amount that the PCA dimensionality reduction produces from.Thereby computational complexity of the present invention is very low, can conveniently realize on embedded platform, has reduced the identification cost.

When using the present invention to carry out recognition of face, at first the facial image that meets authentication is carried out feature extraction and form property data base.After when reality is discerned, importing new facial image, the Gabor bank of filters that use is selected is carried out eigenvalue extracting with the composition characteristic vector to this person's face image, use in minimum neighbour's sorter and the database feature to compare to the eigenvector that obtains then, to judge whether input picture is contained in the database and corresponding identity, wherein, each wave filter forms an eigenwert, and the eigenwert that the whole filter group produces is arranged and formed the one-dimensional characteristic vector.

Although comprise when fewer in number,, its member then needs to choose again the Gabor bank of filters to guarantee accuracy of identification if changing to some extent at property data base.But can predict, only the facial image that need obtain enough scales carries out the optimization of choosing of wave filter, the Gabor bank of filters that can obtain being of universal significance, can effectively extract the facial key character of specific crowd, carry out feature extraction thereby when property data base need change, need not again selecting filter.

The present invention is applied to the heuristic optimized Algorithm of biology among the face identification method first.Proposed to solve the SISPO algorithm of actual optimization problem on the one hand.On the other hand by in the selection and optimization of the SISPO algorithm being introduced the Gabor wave filter, the wave filter of calculation of complex is selected work and general recognition of face work is separated, and used recognition methods based on signature analysis, thereby accuracy of identification and feature extraction specific aim have been promoted, reduced computation complexity significantly, can on embedded platform, realize at an easy rate, effectively reduce the identification cost.

In the present invention, the way of using the SISPO algorithm to select the Gabor bank of filters can use other similar optimized Algorithm to finish, for example ant group algorithm, genetic algorithm, neural network etc.Or directly one group of Gabor wave filter of arteface is discerned.But the recognition effect that these ways obtain generally is difficult to reach the result who uses the SISPO algorithm to select acquisition.

It should be noted that at last, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is had been described in detail with reference to preferred embodiment, those of ordinary skill in the art is to be understood that, can make amendment or be equal to replacement technical scheme of the present invention, and not breaking away from the spirit and scope of technical solution of the present invention, it all should be encompassed in the middle of the claim scope of the present invention.

Claims (10)

1. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm is characterized in that, comprising:

Simplified intelligent single-particle optimizing algorithm comprises and adopts a particle that the solution space of problem function is carried out the optimizing search, and each dimension component of single-particle is carried out intelligent updating respectively during iteration; And

Use simplified intelligent single-particle optimizing algorithm to select the Gabor bank of filters, comprise according to the Gabor number of filter being configured to particle structure, determine the scope of particle search and use the Fisher criterion as fitness function;

When carrying out recognition of face, the facial image that meets authentication is carried out feature extraction form property data base, and behind the new facial image of input, use the described Gabor bank of filters of selecting that this facial image is carried out eigenvalue extracting with the composition characteristic vector, use in minimum neighbour's sorter and the database feature to compare to the eigenvector that obtains, whether be contained in the database and the identity of correspondence to judge input picture;

Wherein, each wave filter forms an eigenwert, and the eigenwert that the whole filter group produces is arranged and formed the one-dimensional characteristic vector.

2. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 1 is characterized in that, the particle of described simplified intelligent single-particle optimizing algorithm more new formula is:

$V_i^{n+1}=(a\&times;r_1/n^{p\&times;r_2})+2\&times;L_i^n---\left(1\right)$

$x_i^{n+1}=\left\{\begin{array}{cc}{x}_i^n+v_i^{n+1}& \mathrm{iff}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ x_i^n& \mathrm{iff}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(2\right)$

$L_i^{n+1}=\left\{\begin{array}{cc}{v}_i^{n+1}& \mathrm{iff}\left(x_i^n\right)>f\left(x_i^{n+1}\right)\\ L_i^n& \mathrm{iff}\left(x_i^n\right)\&le;f\left(x_i^{n+1}\right)\end{array}\right.---\left(3\right)$

Wherein a is the diversity factor, and p is a descending factors, and v is a velocity, x is a position vector, and L is study factor vector, and n is the iterations of current dimension in the formula, i is the dimension of the current renewal of vector, and f () is a fitness function, and r1 and r2 are with each two random values that produce of iterations n.

3. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 2 is characterized in that,

More in the new formula (1), next iteration particle's velocity vector is determined by diversity part and study part: the study part is formed for learning information factor L adds inertial coefficient 2, the diversity part is then by diversity factor a and descending factors p decision, represented the search of particle to attempt behavior, diversity partly is to be the power decreasing function of variable with iterations n, and its variation makes particle change Local Search gradually into by the global search at iteration initial stage;

More in the new formula (3), result's decision that the variation of the study factor is attempted by this iteration: if the position after this iteration is upgraded is better than the position of last iteration, the speed knowledge of then drawing this is as the study factor, otherwise the maintenance last iteration uses the study factor that obtains better result.

4. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 2 is characterized in that,

The key parameter of decision search performance is diversity factor a and descending factors p;

Wherein a determines the step-length that particle search is attempted, the rate of regression of p decision particle search step-length;

When particle tended to global search, needs increased a value and reduce the p value with the bigger trial scope of acquisition, otherwise when tending to Local Search, then needed to reduce a value and increase the p value to carry out precise search and to avoid crossing the optimal value zone.

5. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 4, it is characterized in that, set diversity factor a and these two key parameter factors of descending factors p adaptive updates that carries out with iteration, more new formula is as follows for it:

$a^{k+1}=\left\{\begin{array}{cc}{a}^k-\frac{(a_{\max }-a_{\min })}{K}& \mathrm{if\; f}\left(x^k\right)>f\left(x^{k+1}\right)\\ a^k+\frac{(a_{\max }-a_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{if\; f}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}---\left(4\right)\right.$

$p^{k+1}=\left\{\begin{array}{cc}{p}^k+\frac{(p_{\max }-p_{\min })}{K}& \mathrm{iff}\left(x^k\right)>f\left(x^{k+1}\right)\\ p^k-\frac{(p_{\max }-p_{\min })}{K}\&times;r\left(t^k\right)& \mathrm{iff}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}---\left(5\right)\right.$

$t^{k+1}=\left\{\begin{array}{cc}0& \mathrm{iff}\left(x^k\right)>f\left(x^{k+1}\right)\\ t^k+1& \mathrm{iff}\left(x^k\right)\&le;f\left(x^{k+1}\right)\end{array}\right.---\left(6\right)$

Wherein, k is current whole number of iterations in the formula, and K is the largest global iterations, and max is a maximal value, and min is a minimum value, and t is the disturbance factor.

6. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 5 is characterized in that,

Be [xmin, xmax] if establish the hunting zone of particle on each dimension, then a variation range is [amin=xmin * 0.01, amax=xmax * 100] in the formula, and the p variation range is [pmin=3, pmax=30], and t is set has r (t for the disturbance factor ^k)=uniform (0,2 * t ^k) be absorbed in stagnation with the variation that prevents these two key parameters of a, P.

7. according to claim 5 or 6 described Gabor face identification methods, it is characterized in that the step of described simplified intelligent single-particle optimizing algorithm comprises based on simplified intelligent single-particle optimizing algorithm:

S1, initialization particle position x also calculates its initial adaptive value, initiation parameter a=a _Max, p=p _Min, t=0;

S2, the whole iterations k=1 of initialization;

S3, initialization dimension label i=1, and the study factor is set $L_i^0=0;$

S4, initialization dimension iterations n=1;

S5, according to formula (1), particle rapidity, position and the study factor are upgraded in (2), (3);

S6 is if n less than maximum dimension iterations N, then forwards S4 to;

S7 is if i is less than maximum dimension label i _Max, then forward S3 to;

S8, according to formula (4), (5), (6) undated parameter a, p and t;

S9 is if k less than largest global iterations K, then forwards S2 to.

8. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 1, it is characterized in that, described Gabor wave filter is used to carry out feature extraction, each described Gabor wave filter contains 7 parameters, comprising: 5 parameters of the form of two-dimensional Gabor small echo, wavelength L, angle T, phase place P, aspect ratio G and bandwidth B, and the coordinate X and the Y that act on people's face position.

9. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 8 is characterized in that the hunting zone of each described Gabor wave filter is as follows:

The hunting zone of X is [0, PicSizeX], and the hunting zone of Y is [0, PicSizeY], the hunting zone of L be [2, MIN (PicSizeX, PicSizeY)], the hunting zone of T is [0, π], the hunting zone of P is [π, π], and the hunting zone of G is [0.2,1], the plain scope of searching of B is [0.4,2.5], and wherein PicSizeX and PicSizeY are the size of input facial image.

10. the Gabor face identification method based on simplified intelligent single-particle optimizing algorithm according to claim 1, it is characterized in that, when facial image is carried out feature extraction, with people's specified point on the face is the center, the matrix and the Gabor wave filter dot product of the image identical with Gabor wave filter size, each wave filter will obtain an eigenwert.

Images