CN101615248B - Age estimation method, equipment and face recognition system - Google Patents

Abstract

The embodiment of the invention provides an age estimation method and equipment. The method comprises the following steps: processing face images of users to extract initial characteristic data; selecting subsets of the initial characteristic data as low-dimension characteristic data with dimensionality less than that of the initial characteristic data; carrying out relevant component analysis on the low-dimension characteristic data to obtain training data; according to the obtained training data, training regression analysis parameters used for age estimation; adopting the regression analysis parameters to estimate the ages of the users. The embodiment of the invention can reduce the computation of the age estimation in the face recognition system.

Description

Age estimation method, equipment and face identification system

Technical field

The present invention relates to face identification system, more specifically, relate to age estimation method and equipment in the face identification system.

Background technology

Realize the face identification system of a robust, inevitably will solve the influence of factors such as for example illumination, expression, attitude and age.Estimation of Age is exactly the method according to pattern classification, utilizes the characteristic of facial image, and this individual age is estimated.Man-machine interaction based on the age has huge potential application in daily life.Yet the age recognition technology is but related to seldom under study for action automatically.A chief reason is that the characteristic of some uniquenesses of the aging effect of people face makes estimation of Age need use some off-gauge sorting techniques.This also makes estimation of Age become a challenging problem.

If can solve the age identification problem, then be expected to realize following target:

The man-machine interaction of given age: if computing machine can be judged user's age exactly, computing environment and type of interaction can be made adjustment according to user's age.Be different from the man-machine interaction of standard, such system can combine to guarantee that the teenager is away from unhealthy information with the control of Internet access security.

Facial image index based on the age: age identification automatically can be used for the facial image in the database is sorted; Also can it be used in the electron album, the user can come their photo is sorted, manages through setting the range of age easily like this.

Realization from moving face ageing system: age identification automatically depends on because the perception and the classification of the variation of the facial characteristics that change of age produced, and same method also is needed from the aging simulation system of moving face.

A kind of age recognition methods based on EHMM (built-in type hidden Markov model) is arranged in the prior art; At first; Through analysis, confirm the nonlinear relationship between the variation of age and people's face key feature, and set up nonlinear model based on this relation to great amount of samples; Utilize PARI (partial ageingratio image, local ageing ratio chart picture) to estimate the texture of aging people's face then; At last, utilize the characteristic of from the image of rebuilding, extracting to train EHMM to be used for recognition of face.

In realizing process of the present invention, the inventor finds to exist in the above-mentioned prior art following problem:

The realization cost of existing method is bigger, extract a large amount of profile informations and texture information respectively to the sample set of each age bracket, and operand is very big.

Summary of the invention

The embodiment of the invention provides age estimation method, equipment and the face identification system in a kind of face identification system, to reduce the operand of estimation of Age.

The embodiment of the invention provides a kind of age estimation method, comprising: processing face images of users is to extract initial characteristic data; The subclass of selecting said initial characteristic data is as low dimensional feature data, and the dimension of wherein said low dimensional feature data is less than the dimension of said initial characteristic data; To the constituent analysis of being correlated with of said low dimensional feature data, to obtain training data; According to the said training data that is obtained, training is used for the regretional analysis parameter of estimation of Age; And utilize said regretional analysis parameter to estimate said user's age.

The embodiment of the invention also provides a kind of estimation of Age equipment, comprising: the primitive character extraction module, and processing face images of users is to extract initial characteristic data; Feature selection module, the subclass of selecting said initial characteristic data is as low dimensional feature data, and the dimension of wherein said low dimensional feature data is less than the dimension of said initial characteristic data; Relevant component analysis module is to the constituent analysis of be correlated with of said low dimensional feature data, with the acquisition training data; Training module according to by the said training data that said relevant component analysis module obtained, is trained the regretional analysis parameter that is used for estimation of Age; And estimation module, utilize said regretional analysis parameter to estimate said user's age.

The embodiment of the invention also provides a kind of face identification system, comprises above-mentioned estimation of Age equipment.

The embodiment of the invention adopts the method for relevant constituent analysis to carry out dimensionality reduction for facial image, adopts the mode that returns to carry out estimation of Age, can reduce the operand in the age estimation method.

Description of drawings

In order to be illustrated more clearly in the technical scheme of the embodiment of the invention; To do to introduce simply to the accompanying drawing of required use in embodiment or the description of the Prior Art below; Obviously, the accompanying drawing in describing below only is some embodiments of the present invention, for those of ordinary skills; Under the prerequisite of not paying creative work property, can also obtain other accompanying drawing according to these accompanying drawings.

Fig. 1 is the process flow diagram that illustrates according to the age estimation method of the embodiment of the invention.

Fig. 2 is the process flow diagram that illustrates according to the relevant constituent analysis step of the embodiment of the invention.

Fig. 3 is the comparison synoptic diagram according to the average error result of the quadratic regression estimation of Age of the embodiment of the invention and prior art.

Fig. 4 is the test sample book exemplary plot.

Fig. 5 is the comparison synoptic diagram according to calculated performance on the FG-Net database of the embodiment of the invention and prior art.

Fig. 6 is the synoptic diagram that illustrates according to the estimation of Age equipment of the embodiment of the invention.

Fig. 7 is the synoptic diagram that illustrates according to the face identification system of the embodiment of the invention.

Embodiment

To combine the accompanying drawing in the embodiment of the invention below, the technical scheme in the embodiment of the invention is carried out clear, intactly description, obviously, described embodiment is the present invention's part embodiment, rather than whole embodiment.Based on the embodiment among the present invention, those of ordinary skills are not making the every other embodiment that is obtained under the creative work prerequisite, all belong to the scope of the present invention's protection.

Fig. 1 is the process flow diagram that illustrates according to the age estimation method 100 of the embodiment of the invention.

In this method 100, at S105, processing face images of users is to extract initial characteristic data.By being processed data set (for example, facial image) through calculating or measure generation one stack features value.Because this is one group of data that data set itself gets, therefore be called as initial characteristic data.

In addition, in S105, can carry out pre-service to user's facial image earlier.For example, with the image size normalization.

At S110, the subclass of selecting initial characteristic data is as low dimensional feature data.The purpose of feature selecting is exactly in order to reduce the dimension of feature space, keep as far as possible the effectively information of classification that can be used for.For facial image; Each width of cloth facial image data all constitutes a point of higher dimensional space, but have in numerous dimensions of higher dimensional space can not effectively embody characteristic of division, so carry out linear transformation; The deletion redundant information obtains the most effectively characteristic of division information.

Because primitive character quantity maybe be a lot, in order to adapt to engineering time, must reduce number of features as far as possible.Feature selecting is exactly that the numerical value that from primitive character, obtains is selected the subclass that is used to classify according to certain criterion, as the characteristic of division of dimensionality reduction (dropping to the m dimension from the n dimension).In other words, feature selecting is exactly to utilize the feature space of low dimension of existing characteristic parameter structure, and the useful information that contains in the primitive character is mapped on a few key character, ignores less important information.Just to a n dimension original feature vector

X＝[x ₁，x ₂，…，x _n，] ^T

Through certain conversion, carry out dimensionality reduction, produce the proper vector of low dimension

Y＝[y ₁，y ₂，…，y _m，] ^T，m＜n

Wherein Y contains the principal character of X.

For high dimensional data is mapped with many kinds of methods in low n-dimensional subspace n, principal component analysis (PCA) (PCA) is simple and effective a kind of.PCA is proposed by A.Turk and A.P.Pentland, and it is an expansion of Karhunen-Loeve transformation.Karhunen-Loeve transformation is the method for a kind of typical feature extraction and data representation, is the optimal mapping that under the meaning of square error minimum, obtains data compression, is applicable to probability distribution arbitrarily.Principal component analysis (PCA) is the best approach of carrying out feature extraction at present, has extensively been applied in the middle of a lot of fields such as pattern-recognition and computer vision.PCA is to locate a low n-dimensional subspace n that can satisfy the minimum variance criterion and expresses original high dimensional data.For given data set (initial characteristic data) $X={\left\{x_i\right\}}_{i=1}^n,$ If the definition projection matrix is p, so just find to make following formula value reach maximum subspace:

p＝arg?max?p ^TSp

Here, scatter matrix

$S=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x-\stackrel{\‾}{x}){(x_i-\stackrel{\‾}{x})}^T,$

Wherein x is { x _i} _I=1 ⁿAverage vector.

At S115, to the low dimensional feature data that in S110, the obtain constituent analysis of being correlated with, to obtain training data.

Relevant constituent analysis (RCA; Relevant Component Analysis) can find and reduce the interior overall irrelevant variation of data set.The RCA method distributes bigger weights for relevant dimension through linear transformation, for uncorrelated dimension is distributed less weights, thereby changes the feature space of data representation.The RCA method is the class methods between principal component analysis (PCA) (PCA) and linear discriminant analysis (LDA); It has utilized the sample subclass (sample in each sample subclass belongs to classification identical but the possibility Unknown Label) with relation of equivalence; Estimate relevant dimension and irrelevant dimension through sample subclass, and change (White Transformation) through leucismus and reduce non-correlation with relation of equivalence.

The RCA algorithm is equivalent to utilize the equivalence information of part sample to do differentiation dimensionality reduction and decorrelation.

Fig. 2 is the process flow diagram that illustrates according to the relevant constituent analysis step S115 of the embodiment of the invention.

At S1151, at least one equivalent samples subclass of the low dimensional feature data of structure.Each equivalent samples subclass belongs to age-grade label.For a whole sample, can obtain the equivalence information of its part sample, promptly in these part samples, which sample belongs to identical classification.Need not know the label of classification; Can be expressed as the age label that need not know these facial images here, as long as but can confirm that some sample belongs to same age A1, some sample belongs to same age A2... (wherein A1, A2 the unknown).Can utilize this equivalence information, the original image collection is carried out dimensionality reduction and eliminates non-correlation.

At S1154, the equivalent samples subclass that is utilized among the S1151 to be constructed is differentiated dimensionality reduction to low dimensional feature data, to obtain by the further dimensionality reduction sample of dimensionality reduction.

In the process of dimensionality reduction; Covariance matrix as whole sample set (promptly between the class of the part sample with equivalence information capable of using; Estimation of covariance matrix value between class low dimensional feature data); Utilize cFDA (combined Fisher discriminant analysis, combination Fisher discriminatory analysis) to carry out dimensionality reduction then.

Particularly, for a given data set (low dimensional feature data) { X _i} _I=1 ^NAnd n the equivalent samples subclass that obtains from this data centralization $C_j={\left\{x_{\mathrm{Ij}}\right\}}_{i=1}^{n_j}$ $C_j={\left\{x_{\mathrm{Ij}}\right\}}_{i=1}^{n_j},$ Calculate the interior covariance matrix of class of subclass of equal value:

$\hat{C}=\frac{1}{N}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{i=1}{\overset{n_j}{\mathrm{\Σ}}}(x_{\mathrm{ji}}-m_j){(x_{\mathrm{ji}}-m_j)}^T$

Wherein N is the number of samples of data centralization, m _jIt is subclass $C_j={\left\{x_{\mathrm{Ij}}\right\}}_{i=1}^{n_j}$ In the mean value of sample.

Use

data set is differentiated dimensionality reduction, dimension reduction method is described below.

The dimension of representing low dimensional feature data with D.For one group of given equivalent samples subclass { C _j} _J=1 ⁿ, the order of covariance matrix in the class of calculating equivalent samples subclass $R={\mathrm{\Σ}}_{j=1}^n\left(\right|C_j|-1),$ Wherein | C _j| represent the size of j equivalent samples subclass.

If D＞R, then use PCA with the data dimension dimensionality reduction to α R, 0＜α＜1 wherein.

The covariance matrix S of computing whole _t, and with the estimated value of covariance matrix in the class of equivalent samples subclass as covariance in the class of whole sample set, i.e. S _w=C finds the solution following equality

$\tilde{A}=\underset{A}{\arg }\max \frac{{\mathrm{AS}}_t{A}^t}{{\mathrm{AS}}_w{A}^t}$

And obtain the expression (i.e. " dimensionality reduction sample data " X ') of data in new lower dimensional space with

.

At S1157, reduce the non-correlation of dimensionality reduction sample data, to obtain training data.

In the process that reduces non-correlation, definable is for the sample subclass of same age label, and the direction that its variance is big is irrelevant direction.Can adopt leucismus to bring covariance matrix is carried out conversion, make that the weights of irrelevant direction are lower, so just reduce the age non-correlation in the sample subclass.

Particularly, can recomputate covariance matrix in the class of subclass of equal value behind the dimensionality reduction

$\tilde{C}=\frac{1}{N}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{i=1}{\overset{n_j}{\mathrm{\Σ}}}({\tilde{x}}_{\mathrm{ji}}-m_j){({\tilde{x}}_{\mathrm{ji}}-m_j)}^T$

Where

indicates an equivalent subset

In the low-dimensional subspace representation.

The leucismus that calculates

changes:

$\tilde{C}:W={\tilde{C}}^{-1/2}$

Wherein W is that leucismus changes the result.

The sample that data are concentrated carries out following conversion

X _new＝WX′

Wherein X ' expression is the data point (dimensionality reduction sample data) behind the dimensionality reduction, X _NewIt is the training data that is used for the regretional analysis parameter of sport career age estimation.

Return Fig. 1, after S115 obtained training data, at S120, according to the training data that in S115, obtains, training was used for the regretional analysis parameter of estimation of Age.

Regretional analysis (regression analysis) is a kind of statistical analysis technique of confirming complementary quantitative relationship between two or more parameter.

Represent that afterwards, estimation of Age has just become a multiple linear regression problem in the PCA space to sample set (low dimensional feature data) finding a low dimension PCA subspace.Following formula has provided concrete regression model:

$\mathrm{age}=f\left(M\right):\⇔\hat{L}=\widehat{f\left(y\right)}$

Here

is the age label that estimation obtains; F () is unknown regression equation, is an estimation of regression equation.For matrix form, following formula can be expressed as:

$L=\hat{Y}B+e,$

Var(e)＝σ ²I

The L here is an age label vector;

is a known matrix, and this matrix is made up of the PCA proper vector of sample set.Var () representes variance (Variance), and I is a unit matrix.B needs our the coefficient vector parameter in training stage study.Error vector e is a unobservable random variable, and it is that 0 variance is σ that this stochastic variable should satisfy average ²And irrelevance.

B is tried to achieve by following formula:

$\hat{B}={\left({\hat{Y}}^{T}Y\right)}^{-1}{\hat{Y}}^{T}L=\mathrm{HL}$

Just can obtain regression equation then:

$\hat{L}=\hat{Y}\hat{B}$

Residual vector $\hat{e}=L-\hat{L}$ Satisfy:

$E\left(\hat{e}\right)=0,$ $\mathrm{Var}\left(\hat{e}\right)={\mathrm{\σ}}^2(I-H)$

Take all factors into consideration result's accuracy and algorithm complex, can select quadratic equation to do regretional analysis.With the projection vector y of each sample image in the PCA subspace _iImport as characteristic.

${\hat{l}}_i={\hat{b}}_0+{\hat{b}}_1^T{y}_i+{\hat{b}}_2^T{y}_i^2$

Here,

Be image pattern x _iThe estimation age,

Be side-play amount,

Be to estimate the regression equation coefficient vector.y _iIt is the sample data behind the dimensionality reduction.

Therefore, can obtain:

$\hat{L}={[\hat{l},...,{\hat{l}}_n]}^T$

${\hat{B}}^T={[{\hat{b}}_0{{\hat{b}}_1}^{\left(1\right)}...{{\hat{b}}_1}^{\left(d\right)}{{\hat{b}}_2}^{\left(1\right)}...{{\hat{b}}_2}^{\left(d\right)}]}^T$

$\hat{Y}=\left[1_{n\×1}{[y_1...y_n]}^T{[y_1^2...y_2^2]}^T\right]$

Wherein 1 _{N * 1}Each element of representing a n * 1 dimension all is 1 matrix.

For sample set, above-mentioned formula becomes corresponding matrix form:

$\mathrm{age}=\mathrm{offset}+w_1^{T}b+w_2^T\left(b^2\right)$

Wherein age is the estimation age vector of corresponding sample set, and offset is the side-play amount vector, and b is a sample set, w ₁, w ₂It is the regression coefficient matrix.

Can obtain above-mentioned matrix of coefficients w through the training sample training ₁, w ₂And offset vector offset.We have just obtained the quadratic function value like this.

In S120, obtain after the regretional analysis parameter,, utilize the regretional analysis parameter that is obtained to come the age of estimating user at S125.After obtaining the quadratic regression analytical parameters, test sample book as function input substitution formula, just can be obtained the estimation of Age value of each sample.

Utilized the part classification information of sample set based on the feature dimension reduction method of RCA, i.e. equivalent samples subclass, utilization should equivalence sample subclass be found irrelevant direction, and utilized leucismus to bring the non-correlation that reduces sample set.The effect of RCA dimensionality reduction is the application that combines quadratic regression, reduces non-correlation and dimensionality reduction, not only can save the expense of computing, can also improve to return the precision of estimating.This is because the parameter that returns too much can need more training sample, and in reality, is difficult to satisfy the requirement of large sample amount, and RCA has solved this problem well.Each step of embodiment of the invention method can be adjusted order according to actual needs.

Fig. 3 is the comparison synoptic diagram according to the average error result of the quadratic regression estimation of Age of the embodiment of the invention and prior art.Fig. 4 is the test sample book exemplary plot.

Inventor of the present invention tests on the MORPH storehouse, and the data that will pass through after the PCA dimension-reduction treatment are divided into training sample and test sample book two parts according to 1: 1 ratio.Design for test is following:

Prior art adopts PCA+ to return.That is to say, use PCA that image set is carried out dimensionality reduction, carry out regression forecasting again, calculate the mean absolute error of tieing up the estimation of Age of 80 dimensions from 1 respectively.

According to one embodiment of present invention, adopt RCA+ to return: at first to use PCA with image dimensionality reduction to 80 dimension, adopt the cFDA dimensionality reduction again among the RCA then, carry out regression forecasting again, investigate the mean absolute error of the estimation of Age of from 1 to 80 dimension.(sample proportion that is used for subclass of equal value here is 30%, and when the cFDA in using RCA carried out dimensionality reduction, each age label was regarded as one type).

Particularly, the data that will pass through after the PCA dimension-reduction treatment are divided into training sample and test sample book two parts according to 1: 1 ratio.In test, at first original image is tieed up with PCA dimensionality reduction to 80.Wherein utilize PCA that original image is carried out dimension-reduction treatment.

Secondly, utilize RCA to reduce the non-correlation and the further dimensionality reduction of data.Here selected part has equivalent samples subclass of some images compositions of same age label, and constructs some equivalent samples subclass with all ages and classes label.Here we choose original sample and concentrate 30% sample composition equivalent samples subclass.Calculate covariance matrix in the class of this equivalence sample subclass (for fear of the linear discriminant dimensionality reduction of influence use to(for) regression result, with each age label as a type) then.

$\hat{C}=\frac{1}{N}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{i=1}{\overset{n_j}{\mathrm{\Σ}}}(x_{\mathrm{ji}}-m_j){(x_{\mathrm{ji}}-m_j)}^T$

And using the interior covariance of class that obtains from the sample subclass, we carry out linear discriminant dimensionality reduction (as above S1154 is said) to original image set.For the sample behind the dimensionality reduction, we use leucismus to change further minimizing non-correlation

$\tilde{C}:W={\tilde{C}}^{-1/2}$

At last, we adopt quadratic equation to carry out regression treatment to training sample.

${\hat{l}}_i={\hat{b}}_0+{\hat{b}}_1^T{y}_i+{\hat{b}}_2^T{y}_i^2$

Here,

Be image pattern x _iThe estimation age,

Be side-play amount,

Be to estimate the regression equation coefficient vector.y _iIt is the sample data behind the dimensionality reduction.

For sample set, above-mentioned equality becomes corresponding matrix form:

$\mathrm{age}=\mathrm{offset}+w_1^{T}b+w_2^T\left(b^2\right)$

Age is the estimation age vector of corresponding sample set, and offset is the side-play amount vector, and b is a sample set, w ₁, w ₂It is the regression coefficient matrix.

Can obtain above-mentioned matrix of coefficients w through the training sample training ₁, w ₂And offset vector offset.So just obtained the quadratic function value.Test sample book as function input substitution formula, just can be obtained the estimation of Age value of each sample.

The result is as shown in Figure 3, and the X axle is represented the intrinsic dimensionality got, and the Y axle is represented mean absolute error (year).

According to Fig. 3 result, can see under the situation of the prior art of only using PCA that when the dimension of hanging down n-dimensional subspace n was lower, mean absolute error had tangible increase.And can keep lower stable mean absolute error according to the RCA method of the embodiment of the invention.And the RCA method has obtained the result who is superior to PCA when hanging down dimension.

In addition, for the method for the comparison embodiment of the invention and the difference of method on calculated performance and estimated efficiency of prior art, inventor of the present invention has carried out following test.

On FG-Net, test, compare with the method for PCA+SVR (Support Vector Regression, support vector regression), the method that in the test each algorithm was all adopted 1: 1 is divided into test set and training set to the FG-Net storehouse.Test findings and shown in Figure 5.Fig. 5 is the comparison synoptic diagram according to calculated performance on the FG-Net database of the embodiment of the invention and prior art, and wherein the X axle is the intrinsic dimensionality behind the dimensionality reduction, and the Y axle is second for the time of image training and test in the test, unit.By contrast; The method of the RCA+QR (Quadratic Regression, quadratic regression) that the embodiment of the invention adopts has apparent in view improvement on calculated performance, and with it more approaching result is arranged on estimated performance; Consider the requirement of practical application for the degree of accuracy of estimation of Age; Inventor of the present invention finds that the method that adopts RCA+QR significantly improves the performance of calculating on the basis that can keep the low mean absolute error of estimating.

Table 1 is the operation result on the FG-Net storehouse

Method	SVM	SVR	LAR4	LAR8	LAR16	LAR32	RCA+QR
								Mean absolute error	7.16	5.16	5.07	5.07	5.12	6.03	5.19

LAR4 wherein, LAR8, LAR16 and LAR32 are respectively the robustness methods of the part adjustment on the SVR basis, carried out, and the time complexity of its calculating will be higher than SVR.Only compared SVR and the RCA+QR complexity on computing time here.

Notice that the actual age of three samples shown in Figure 4 is respectively 16 years old, 39 years old and 62 years old.The estimation age of utilizing the method for the embodiment of the invention to obtain is 18.2,32.0 and 56.9.Accuracy of estimation for old-age group is lower, and its reason is mostly the training storehouse of being adopted is that living photo of the remote past is second-rate, and texture is fuzzy, and the non-interference of expression and attitude etc. is naturally arranged.

Employing can reduce the operand of estimation of Age according to the age estimation method of the embodiment of the invention.

Fig. 6 is the synoptic diagram that illustrates according to the estimation of Age equipment 200 of the embodiment of the invention.

This estimation of Age equipment 200 comprises primitive character extraction module 205, feature selection module 210, relevant component analysis module 215, training module 220 and estimation module 225.

Estimation of Age equipment 200 can be carried out above-described method 100.Particularly, primitive character extraction module 205 processing face images of users are to extract initial characteristic data.Feature selection module 210 selects the subclass of initial characteristic data as low dimensional feature data, and wherein the dimension of low dimensional feature data is less than the dimension of initial characteristic data.The constituent analysis of be correlated with of relevant 215 pairs of low dimensional feature data of component analysis module is with the acquisition training data.Training module 220, according to the training data that is obtained by relevant component analysis module 215, training is used for the regretional analysis parameter of estimation of Age.Estimation module 225 utilizes the regretional analysis parameter that is obtained by training module 220 training to come the age of estimating user.

Each module of estimation of Age equipment 200 with and the operation can repeat no more at this corresponding to above-mentioned age estimation method 100.

Employing can reduce the operand of estimation of Age according to the estimation of Age equipment of the embodiment of the invention.

Fig. 7 is the synoptic diagram that illustrates according to the face identification system 300 of the embodiment of the invention.This people's face device systems 300 can comprise above-mentioned estimation of Age equipment 200.It will be understood by those skilled in the art that according to actual needs people's face device systems 300 can also comprise other equipment (not shown), as be used to read facial image equipment, be used to equipment of storing facial image etc.

Employing can reduce the operand of estimation of Age according to the face identification system of the embodiment of the invention.

It will be appreciated by those skilled in the art that the module in the device among the embodiment can be distributed in the device of embodiment according to the embodiment description, also can carry out respective change and be arranged in the one or more devices that are different from present embodiment.The module of the foregoing description can be merged into a module, also can further split into a plurality of submodules.

The embodiment of the invention has proposed a kind of age estimation method based on relevant constituent analysis and quadratic regression, estimates for the sample that facial image is concentrated.The embodiment of the invention is used relevant constituent analysis when carrying out dimensionality reduction for facial image, reduced the non-correlation of facial image collection and in dimensionality reduction, considered the age discriminant information.For age-grade facial image, bring to changing the bigger lower weights of direction (irrelevant dimension) distribution through leucismus, thereby having the facial image of same age, assurance keeps less dispersion at lower dimensional space.On the other hand, the embodiment of the invention is applied to regretional analysis in the estimation of Age.Particularly, at first, utilize principal component analysis (PCA) that sample is carried out dimensionality reduction.Secondly, utilize RCA to reduce the non-correlation of sample set, and further select suitable differentiation dimensionality reduction space.In the dimensionality reduction subspace, use quadratic regression, set up recurrence and anticipation function.

The embodiment of the invention can be applicable to people's age is estimated, for making marks and put image library in order in the video image storage technical support is provided, retrieves to wait for suspect and also can dwindle range of search, the raising recall precision.In addition, the embodiment of the invention can provide technical foundation for recognition of face, character classification by age etc.

It should be noted that the embodiment of the invention described employing PCA and carried out dimensionality reduction, but it will be understood by those skilled in the art that other dimension reduction methods also can be applicable to the present invention, thereby also within the scope of the invention.

Those of ordinary skills can recognize; The unit and the algorithm steps of each example of describing in conjunction with embodiment disclosed herein; Can realize with electronic hardware, computer software or the combination of the two; For the interchangeability of hardware and software clearly is described, the composition and the step of each example described prevailingly according to function in above-mentioned explanation.These functions still are that software mode is carried out with hardware actually, depend on the application-specific and the design constraint of technical scheme.The professional and technical personnel can use distinct methods to realize described function to each certain applications, but this realization should not thought and exceeds scope of the present invention.

The software module that the method for describing in conjunction with embodiment disclosed herein or the step of algorithm can use hardware, processor to carry out, perhaps the combination of the two is implemented.Software module can place the storage medium of any other form known in random access memory (RAM), internal memory, ROM (read-only memory) (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or the technical field.

Although illustrated and described some embodiments of the present invention, it will be understood by those skilled in the art that under the situation that does not break away from principle of the present invention and spirit can carry out various modifications to these embodiment, such modification should fall in the scope of the present invention.

Claims (11)

1. the age estimation method in the face identification system is characterized in that, comprising:

Processing face images of users is to extract initial characteristic data;

The subclass of selecting said initial characteristic data is as low dimensional feature data, and the dimension of wherein said low dimensional feature data is less than the dimension of said initial characteristic data;

To the constituent analysis of being correlated with of said low dimensional feature data, to obtain training data;

According to the said training data that is obtained, training is used for the regretional analysis parameter of estimation of Age;

Utilize the said user's of said regretional analysis parameter estimation age,

It is wherein said that the constituent analysis of being correlated with comprises to obtain training data to said low dimensional feature data:

Construct at least one equivalent samples subclass of said low dimensional feature data, wherein each said equivalent samples subclass belongs to age-grade label;

Utilize the equivalent samples subclass of being constructed, said low dimensional feature data are differentiated dimensionality reduction, to obtain by the dimensionality reduction sample data of further dimensionality reduction;

Reduce the non-correlation of said dimensionality reduction sample data, to obtain said training data.

2. age estimation method as claimed in claim 1 is characterized in that, the subclass of the said initial characteristic data of said selection comprises as low dimensional feature data: the subclass that said initial characteristic data is selected in the employing principal component analysis (PCA) is as said low dimensional feature data.

3. age estimation method as claimed in claim 1 is characterized in that, the equivalent samples subclass that said utilization is constructed is differentiated dimensionality reduction to said low dimensional feature data and comprised:

Confirm the interior covariance matrix of class of said equivalent samples subclass;

As Estimation of covariance matrix value in the class of said low dimensional feature data, utilize combination Fisher discriminatory analysis that said low dimensional feature data are differentiated dimensionality reduction covariance matrix in the class of said equivalent samples subclass.

4. age estimation method as claimed in claim 1 is characterized in that, the non-correlation of the said dimensionality reduction sample data of said minimizing comprises:

For the equivalent samples subclass that belongs to age-grade label, defining the big direction of its variance is irrelevant direction.

5. age estimation method as claimed in claim 1 is characterized in that, the non-correlation of the said dimensionality reduction sample data of said minimizing comprises:

Confirm the interior covariance matrix of class of said dimensionality reduction sample data;

Covariance matrix in the class of said dimensionality reduction sample data is carried out leucismus change, obtain leucismus and change the result;

Change result and said dimensionality reduction sample data based on said leucismus, obtain said training data.

6. the estimation of Age equipment in the face identification system is characterized in that, comprising:

The primitive character extraction module is used for processing face images of users to extract initial characteristic data;

Feature selection module, the subclass that is used to select said initial characteristic data is as low dimensional feature data, and the dimension of wherein said low dimensional feature data is less than the dimension of said initial characteristic data;

Relevant component analysis module is used for to the constituent analysis of be correlated with of said low dimensional feature data, with the acquisition training data;

Training module is used for training the regretional analysis parameter that is used for estimation of Age according to by the said training data that said relevant component analysis module obtained;

Estimation module is used to utilize said regretional analysis parameter to estimate said user's age,

Wherein said relevant component analysis module comprises:

Tectonic element is used to construct at least one equivalent samples subclass of said low dimensional feature data, and wherein each said equivalent samples subclass belongs to age-grade label;

Differentiate the dimensionality reduction unit, be used to utilize the equivalent samples subclass of being constructed, said low dimensional feature data are differentiated dimensionality reduction, to obtain by the dimensionality reduction sample data of further dimensionality reduction;

The decorrelation unit is used to reduce the non-correlation of said dimensionality reduction sample data, to obtain said training data.

7. estimation of Age equipment as claimed in claim 6 is characterized in that, the subclass that said initial characteristic data is selected in said feature selection module employing principal component analysis (PCA) is as said low dimensional feature data.

8. estimation of Age equipment as claimed in claim 6; It is characterized in that; The interior covariance matrix of the class of said equivalent samples subclass is confirmed in said differentiation dimensionality reduction unit; As Estimation of covariance matrix value in the class of said low dimensional feature data, utilize combination Fisher discriminatory analysis that said low dimensional feature data are differentiated dimensionality reduction covariance matrix in the class of said equivalent samples subclass.

9. estimation of Age equipment as claimed in claim 6 is characterized in that, said decorrelation module is for the equivalent samples subclass that belongs to age-grade label, and defining the big direction of its variance is irrelevant direction.

10. estimation of Age equipment as claimed in claim 6; It is characterized in that; Said decorrelation module is confirmed the interior covariance matrix of the class of said dimensionality reduction sample data, said dimensionality reduction sample data is carried out leucismus change, to obtain said training data; Change result and said dimensionality reduction sample data based on said leucismus, obtain said training data.

11. a face identification system is characterized in that, comprises like each described estimation of Age equipment of claim 6 to 10.

Images