CN101615248A - Age estimation method, device and face recognition system - Google Patents

Abstract

The embodiment of the invention provides a kind of age estimation method and equipment.This method comprises: processing face images of users is to extract initial characteristic data; The subclass of selecting described initial characteristic data is as the low-dimensional characteristic, and the dimension of wherein said low-dimensional characteristic is less than the dimension of described initial characteristic data; To the constituent analysis of being correlated with of described low-dimensional characteristic, to obtain training data; According to the described training data that is obtained, training is used for the regretional analysis parameter of estimation of Age; Utilize described regretional analysis parameter to estimate described user's age.By the embodiment of the invention, can reduce the operand of the estimation of Age in the face identification system.

Description

Age estimation method, device and face recognition system

technical field

The present invention relates to a face recognition system, and more particularly, to an age estimation method and device in the face recognition system.

Background technique

To implement a robust face recognition system, it is inevitable to address the influence of factors such as illumination, expression, posture, and age. Age estimation is to estimate the age of the individual by using the characteristics of the face image according to the method of pattern classification. Age-based human-computer interaction has huge potential applications in everyday life. However, automatic age recognition technology has been rarely involved in research. A major reason is that several unique properties of the aging effect on human faces necessitate the use of some non-standard classification methods for age estimation. This also makes age estimation a challenging problem.

If the age recognition problem can be solved, the following goals are expected to be achieved:

Age-specific human-computer interaction: If the computer can accurately determine the age of the user, both the computing environment and the type of interaction can be adapted to the age of the user. Unlike standard human-computer interaction, such a system can be combined with Internet access security controls to keep young people away from unhealthy information.

Age-based face image indexing: automatic age recognition can be used to sort face images in the database, and it can also be used in electronic photo albums, so that users can easily sort their photos by setting the age range Sort and manage.

Realization of automatic face aging system: Automatic age recognition depends on the perception and classification of changes in facial features due to age changes, and the same method is also required for automatic face aging simulation systems.

In the prior art, there is an age recognition method based on EHMM (Embedded Hidden Markov Model). First, through the analysis of a large number of samples, the nonlinear relationship between age and the change of key features of the face is determined, and based on this relationship to establish a nonlinear model; then use PARI (partial ageingratio image, local aging ratio image) to estimate the texture of the aging face; finally, use the features extracted from the reconstructed image to train EHMM for face recognition.

In the course of realizing the present invention, the inventor finds that there are following problems in the above-mentioned prior art:

The implementation cost of the existing method is relatively high, and it is necessary to extract a large amount of contour information and texture information for the sample set of each age group, and the amount of calculation is very large.

Contents of the invention

Embodiments of the present invention provide an age estimation method and device in a face recognition system, and a face recognition system, so as to reduce the calculation amount of age estimation.

An embodiment of the present invention provides an age estimation method, including: processing the user's face image to extract original feature data; selecting a subset of the original feature data as low-dimensional feature data, wherein the dimension of the low-dimensional feature data smaller than the dimension of the original feature data; performing correlation component analysis on the low-dimensional feature data to obtain training data; according to the obtained training data, training regression analysis parameters for age estimation; and using the regression Analyze parameters to estimate the user's age.

The embodiment of the present invention also provides an age estimation device, including: an original feature extraction module, which processes the user's face image to extract original feature data; a feature selection module, which selects a subset of the original feature data as low-dimensional feature data , wherein the dimensions of the low-dimensional feature data are smaller than the dimensions of the original feature data; the correlation component analysis module performs correlation component analysis on the low-dimensional feature data to obtain training data; the training module, based on the correlation components The training data obtained by the analysis module trains regression analysis parameters for age estimation; and the estimation module uses the regression analysis parameters to estimate the user's age.

An embodiment of the present invention also provides a face recognition system, including the above-mentioned age estimation device.

In the embodiment of the present invention, the method of correlation component analysis is used to reduce the dimension of the face image, and the method of regression is used to estimate the age, which can reduce the amount of computation in the age estimation method.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without paying creative labor.

FIG. 1 is a flowchart illustrating an age estimation method according to an embodiment of the present invention.

FIG. 2 is a flow chart showing the steps of correlation component analysis according to the embodiment of the present invention.

Fig. 3 is a schematic diagram of comparison of average error results of quadratic regression age estimation according to an embodiment of the present invention and the prior art.

Figure 4 is an example diagram of a test sample.

Fig. 5 is a schematic diagram of the comparison of computing performance on the FG-Net database according to the embodiment of the present invention and the prior art.

FIG. 6 is a schematic diagram showing an age estimation device according to an embodiment of the present invention.

Fig. 7 is a schematic diagram showing a face recognition system according to an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

FIG. 1 is a flowchart illustrating an age estimation method 100 according to an embodiment of the present invention.

In the method 100, at S105, the face image of the user is processed to extract original feature data. A set of feature values is generated by calculation or measurement from the processed data set (for example, face image). Since this is a set of data from the dataset itself, it is called raw feature data.

In addition, in S105, the user's face image may be preprocessed first. For example, normalize image size.

At S110, a subset of the original feature data is selected as low-dimensional feature data. The purpose of feature selection is to reduce the dimensionality of the feature space and retain as much information as possible for effective classification. For face images, each face image data constitutes a point in high-dimensional space, but some of the many dimensions of high-dimensional space cannot effectively reflect the classification features, so linear transformation is performed to delete redundant information , to get the most effective classification feature information.

Since the number of original features may be large, in order to accommodate the engineering time, the number of features must be reduced as much as possible. Feature selection is to use the values obtained from the original features to select a subset for classification according to certain criteria, as a classification feature for dimensionality reduction (from n-dimension to m-dimension). In other words, feature selection is to use the existing feature parameters to construct a lower-dimensional feature space, map the useful information contained in the original features to a few important features, and ignore the secondary information. That is, for an n-dimensional original feature vector

X=[x ₁ , x ₂ , . . . , x _n ,] ^T

Through a certain transformation, perform dimensionality reduction to generate low-dimensional feature vectors

Y=[y ₁ , y ₂ ,..., y _m ,] ^T , m<n

where Y contains the main features of X.

There are many methods for mapping high-dimensional data to low-dimensional subspaces, principal component analysis (PCA) is a simple and effective one. PCA was proposed by A.Turk and APPentland, it is an extension of KL transform. KL transformation is a typical method of feature extraction and data representation, and it is the best transformation to obtain data compression in the sense of minimum mean square error, and it is suitable for any probability distribution. Principal component analysis is currently the best method for feature extraction, and it has been widely used in many fields such as pattern recognition and computer vision. PCA is to find a low-dimensional subspace that can satisfy the minimum variance criterion to express the original high-dimensional data. For a given dataset (raw feature data) $x={\left\{x_i\right\}}_{i=1}^{\mathrm{no}},$ If the projection matrix is defined as p, then it is to find the subspace that can maximize the value of the following formula:

p = arg max p ^T Sp

Here, the scatter matrix

$\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ,</span>$

where x is the mean vector of { _xi } _i=1 ⁿ .

In S115, perform correlation component analysis on the low-dimensional feature data obtained in S110 to obtain training data.

Relevant Component Analysis (RCA; Relevant Component Analysis) can find and reduce global non-correlated changes within a dataset. The RCA method assigns larger weights to relevant dimensions and smaller weights to irrelevant dimensions through linear transformation, thereby changing the feature space represented by the data. The RCA method is a class of methods between principal component analysis (PCA) and linear discriminant analysis (LDA), which utilizes sample subsets with an equivalence relationship (the samples in each sample subset belong to the same but possibly unknown labels categories), estimate relevant and non-correlated dimensions through sample subsets with equivalence relationships, and reduce non-correlation through White Transformation.

The RCA algorithm is equivalent to using the equivalence information of some samples for discriminant dimension reduction and decorrelation.

FIG. 2 is a flowchart showing the relevant component analysis step S115 according to an embodiment of the present invention.

At S1151, construct at least one equivalent sample subset of low-dimensional feature data. Each equivalent sample subset belongs to the same age label. For an overall sample, the equivalence information of its partial samples can be obtained, that is, which samples belong to the same category among these partial samples. There is no need to know the labels of the categories, which can be expressed here as that it is not necessary to know the age labels of these face images, but as long as it can be determined that some samples belong to the same age A1, and some samples belong to the same age A2...( Among them, A1 and A2 are unknown). This equivalence information can be used to reduce the dimensionality and eliminate non-correlation of the original image set.

In S1154, use the equivalent sample subset constructed in S1151 to perform discrimination and dimensionality reduction on the low-dimensional feature data to obtain further dimensionality-reduced samples.

In the process of dimensionality reduction, the inter-class covariance matrix of some samples with equivalence information can be used as the estimated value of the inter-class covariance matrix of the overall sample set (ie, low-dimensional feature data), and then cFDA( combined Fisher discriminant analysis, combined Fisher discriminant analysis) for dimensionality reduction.

Specifically, for a given data set (low-dimensional feature data) {X _i } _i=1 ^N and n equivalent sample subsets obtained from the data set $C_j={\left\{x_{\mathrm{ij}}\right\}}_{i=1}^{{\mathrm{no}}_j}$ $C_j={\left\{x_{\mathrm{ij}}\right\}}_{i=1}^{{\mathrm{no}}_j},$ Compute the within-class covariance matrix for equivalent subsets:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

Where N is the number of samples in the data set, and m _j is the subset $C_j={\left\{x_{\mathrm{ij}}\right\}}_{i=1}^{{\mathrm{no}}_j}$ The mean of the samples in .

对数据集进行判别降维，降维方法如下所述。use

Discriminant dimensionality reduction is performed on the data set, and the dimensionality reduction method is as follows.

Let D denote the dimension of low-dimensional feature data. For a given set of equivalent sample subsets {C _j } _j=1 ⁿ , calculate the rank of the intra-class covariance matrix of the equivalent sample subset $R={\mathrm{\&Sigma;}}_{j=1}^{\mathrm{no}}\left(\right|C_j|-1),$ where |C _j | denotes the size of the jth equivalent sample subset.

If D>R, use PCA to reduce the dimensionality of the data to αR, where 0<α<1.

Calculate the overall covariance matrix S _t , and use the intra-class covariance matrix of the equivalent sample subset as the estimated value of the intra-class covariance of the overall sample set, that is, S _w =C, and solve the following equation

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; A</span>}}{\mathrm{<span\; class="notranslate">\; arg</span>}}\mathrm{<span\; class="notranslate">\; max</span>}\frac{{\mathrm{<span\; class="notranslate">\; AS</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; A</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}}{{\mathrm{<span\; class="notranslate">\; AS</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}{\mathrm{<span\; class="notranslate">\; A</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}}$

来获得数据在新的低维空间中的表示(即“降维样本数据”X′)。and use

To obtain the representation of the data in a new low-dimensional space (that is, "reduced dimensionality sample data"X').

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

In the process of reducing non-correlation, it can be defined that for the sample subset of the same age label, the direction with large variance is the non-correlation direction. White transformation can be used to transform the covariance matrix so that the weight of non-correlated directions is lower, which reduces the age non-correlation in the sample subset.

Specifically, the intra-class covariance matrix of the equivalent subset after dimensionality reduction can be recalculated

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

表示等价子集在低维子空间中的表示。in

represents an equivalence subset Representations in low-dimensional subspaces.

的白变换：calculate

The white transformation of :

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; :</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$

where W is the white transform result.

The samples in the dataset are transformed as follows

X _new = WX'

Where X' represents the data point after dimension reduction (dimension reduction sample data), and X _new is the training data used to train the regression analysis parameters for age estimation.

Returning to Fig. 1, after the training data is obtained in S115, in S120, the regression analysis parameters for age estimation are trained according to the training data obtained in S115.

Regression analysis is a statistical analysis method to determine the interdependent quantitative relationship between two or more variables.

After finding a low-dimensional PCA subspace to represent the sample set (low-dimensional feature data), age estimation becomes a multiple linear regression problem in PCA space. The following formula gives the specific regression model:

$\mathrm{<span\; class="notranslate">\; age</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; :</span><span\; class="notranslate">\; \&DoubleLeftRightArrow;</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}$

是估计得到的年龄标签，f(·)是未知的回归方程，是回归方程的一个估计。对于矩阵形式，上式可以表示为：here

is the estimated age label, f( ) is the unknown regression equation, and is an estimate of the regression equation. In matrix form, the above formula can be expressed as:

$\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; ,</span>$

Var(e)=σ ² I

是一个已知矩阵，该矩阵是由样本集的PCA特征向量组成。Var()表示方差(Variance)，I是单位矩阵。B是需要我们在训练阶段学习的系数向量参数。误差向量e是不可观测的随机变量，该随机变量应该满足均值为0方差为σ²以及不相关性。Here L is the age label vector,

Is a known matrix, which is composed of PCA eigenvectors of the sample set. Var() represents the variance (Variance), and I is the identity matrix. B is the coefficient vector parameter that we need to learn in the training phase. The error vector e is an unobservable random variable, which should have a mean of 0, a variance of σ ² and no correlation.

B is obtained by the following formula:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; HL</span>}$

Then the regression equation can be obtained:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; B</span>}}$

residual vector $\hat{e}=L-\hat{L}$ satisfy:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>$ $\mathrm{<span\; class="notranslate">\; Var</span>}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; )</span>$

Considering the accuracy of the results and the complexity of the algorithm, the quadratic equation can be selected for regression analysis. Take the projection vector y _i of each sample image in the PCA subspace as feature input.

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; l</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; +</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

是图像样本x_i的估计年龄，

是偏移量，

是估计回归方程系数向量。y_i是降维后的样本数据。here,

is the estimated age of image sample _xi ,

is the offset,

is the estimated regression equation coefficient vector. y _i is the sample data after dimensionality reduction.

Therefore, one can get:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; l</span>}}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; B</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}{{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>{{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; )</span>}{{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>{{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; 1</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ]</span>$

where 1 _n×1 represents an n×1-dimensional matrix in which each element is 1.

For the sample set, the above formula becomes the corresponding matrix form:

$\mathrm{<span\; class="notranslate">\; age</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; offset</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; b</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>$

Among them, age is the estimated age vector of the corresponding sample set, offset is the offset vector, b is the sample set, and w ₁ and w ₂ are regression coefficient matrices.

The aforementioned coefficient matrices w ₁ , w ₂ and the offset vector offset can be obtained through training with training samples. In this way we get the quadratic function value.

After the regression analysis parameters are obtained in S120, in S125, the age of the user is estimated by using the obtained regression analysis parameters. After the quadratic regression analysis parameters are obtained, the test sample is substituted into the formula as a function input, and the age estimate of each sample can be obtained.

The RCA-based feature dimensionality reduction method utilizes part of the category information of the sample set, that is, the equivalent sample subset, uses the equivalent sample subset to find the non-correlation direction, and uses white transformation to reduce the non-correlation of the sample set. The role of RCA dimensionality reduction is to combine the application of quadratic regression to reduce non-correlation and dimensionality reduction, which can not only save computing costs, but also improve the accuracy of regression estimation. This is because too many regression parameters will require more training samples, and it is difficult to meet the requirements of large sample size in practice. RCA solves this problem well. The order of each step of the method in the embodiment of the present invention can be adjusted according to actual needs.

Fig. 3 is a schematic diagram of comparison of average error results of quadratic regression age estimation according to an embodiment of the present invention and the prior art. Figure 4 is an example diagram of a test sample.

The inventors of the present invention conducted experiments on the MORPH database, and divided the data after PCA dimension reduction processing into two parts, training samples and testing samples, according to a ratio of 1:1. For the experiment design as follows:

The prior art uses PCA+regression. That is, PCA is used to reduce the dimensionality of the image set, and then regression prediction is performed to calculate the mean absolute error of the age estimation from 1D to 80D respectively.

According to one embodiment of the present invention, RCA+regression is adopted: firstly, PCA is used to reduce the dimensionality of the image to 80 dimensions, and then the cFDA in RCA is used to reduce the dimensionality again, and then regression prediction is performed, and the average absolute value of age estimates from 1 to 80 dimensions is examined error. (The sample ratio used for the equivalent subset here is 30%, and each age label is regarded as one class when using cFDA in RCA for dimensionality reduction).

Specifically, the data after PCA dimensionality reduction processing is divided into two parts, training samples and testing samples, according to a ratio of 1:1. In the experiment, the original image is first reduced to 80 dimensions by PCA. Among them, PCA is used to reduce the dimensionality of the original image.

Second, use RCA to reduce the non-correlation of data and further reduce the dimensionality. Here, several images with the same age label are selected to form an equivalent sample subset, and several equivalent sample subsets with different age labels are constructed. Here we select 30% of the samples in the original sample set to form an equivalent sample subset. Then calculate the intra-class covariance matrix of the equivalent sample subset (in order to avoid the influence of linear discriminant dimensionality reduction on the regression results, each age label is regarded as a class).

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

And using the within-class covariance obtained from the sample subset, we perform linear discriminant dimensionality reduction on the original image set (as described in S1154 above). For the reduced samples, we use the white transform to further reduce the non-correlation

$\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; :</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; C</span>}}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$

Finally, we apply a quadratic regression to the training samples.

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; l</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; +</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

是偏移量，

是估计回归方程系数向量。y_i是降维后的样本数据。here, is the estimated age of image sample _xi ,

is the offset,

is the estimated regression equation coefficient vector. y _i is the sample data after dimensionality reduction.

For the sample set, the above equation becomes the corresponding matrix form:

$\mathrm{<span\; class="notranslate">\; age</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; offset</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; b</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>$

age is the estimated age vector of the corresponding sample set, offset is the offset vector, b is the sample set, w ₁ and w ₂ are regression coefficient matrices.

The aforementioned coefficient matrices w ₁ , w ₂ and the offset vector offset can be obtained through training with training samples. In this way, the value of the quadratic function is obtained. Substituting test samples as function inputs into the formula yields age estimates for each sample.

The results are shown in Figure 3, the X-axis represents the feature dimension taken, and the Y-axis represents the mean absolute error (year).

According to the results in Fig. 3, it can be seen that in the case of using only the prior art of PCA, when the dimension of the low-dimensional subspace is low, the mean absolute error increases significantly. However, the RCA method according to the embodiment of the present invention can maintain a relatively low and stable mean absolute error. And the RCA method achieves better results than PCA in lower dimensions.

In addition, in order to compare the difference in calculation performance and estimation efficiency between the method of the embodiment of the present invention and the method of the prior art, the inventors of the present invention conducted the following experiments.

Experiments were carried out on FG-Net and compared with the method of PCA+SVR (Support Vector Regression, Support Vector Regression). In the experiment, each algorithm was divided into a test set by a 1:1 method. and the training set. The test results are shown in Figure 5. Fig. 5 is according to the embodiment of the present invention and the comparative schematic diagram of computing performance on the FG-Net database of prior art, and wherein X-axis is the feature dimension after dimensionality reduction, and Y-axis is the time of image training and testing in the experiment, and the unit is Second. In contrast, the RCA+QR (Quadratic Regression, quadratic regression) method adopted in the embodiment of the present invention has a relatively obvious improvement in calculation performance, and has a relatively close result in estimation performance, considering practical application For the requirement of the accuracy of age estimation, the inventors of the present invention found that the RCA+QR method can greatly improve the calculation performance on the basis of keeping the estimated mean absolute error low.

Table 1 is the result of running on the FG-Net library

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

Among them, LAR4, LAR8, LAR16, and LAR32 are robust methods for local adjustment based on SVR, respectively, and their calculation time complexity is higher than that of SVR. Here, only the computational time complexity of SVR and RCA+QR is compared.

Note that the actual ages of the three samples shown in Figure 4 are 16, 39 and 62 respectively. The estimated ages obtained using the method of the embodiment of the present invention are 18.2, 32.0 and 56.9. The estimation accuracy of aging is low, because the training library used is mostly old life photos with poor quality, blurry texture, and interference such as unnatural expressions and postures.

By adopting the age estimation method according to the embodiment of the present invention, the calculation amount of age estimation can be reduced.

FIG. 6 is a schematic diagram illustrating an age estimation device 200 according to an embodiment of the present invention.

The age estimation device 200 includes an original feature extraction module 205 , a feature selection module 210 , a correlation component analysis module 215 , a training module 220 and an estimation module 225 .

The age estimation device 200 may perform the method 100 described above. Specifically, the original feature extraction module 205 processes the user's face image to extract original feature data. The feature selection module 210 selects a subset of the original feature data as low-dimensional feature data, wherein the dimension of the low-dimensional feature data is smaller than that of the original feature data. The correlation component analysis module 215 performs correlation component analysis on the low-dimensional feature data to obtain training data. The training module 220 trains regression analysis parameters for age estimation according to the training data obtained by the correlation component analysis module 215 . The estimation module 225 uses the regression analysis parameters trained by the training module 220 to estimate the user's age.

Each module of the age estimating device 200 and its operation may correspond to the above-mentioned age estimating method 100 , and will not be repeated here.

With the age estimating device according to the embodiment of the present invention, the calculation amount of age estimation can be reduced.

FIG. 7 is a schematic diagram showing a face recognition system 300 according to an embodiment of the present invention. The human face device system 300 may include the above-mentioned age estimation device 200 . Those skilled in the art can understand that, according to actual needs, the face device system 300 may also include other devices (not shown), such as devices for reading face images, devices for storing face images, and the like.

By adopting the face recognition system according to the embodiment of the present invention, the calculation amount of age estimation can be reduced.

Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description in the embodiment, and can also be changed and located in one or more devices different from the embodiment. The modules in the above embodiments can be combined into one module, and can also be further split into multiple sub-modules.

An embodiment of the present invention proposes an age estimation method based on correlation component analysis and quadratic regression, which estimates samples in a face image set. In the embodiments of the present invention, when performing dimensionality reduction on face images, correlation component analysis is used to reduce non-correlation of face image sets and age discrimination information is considered in dimensionality reduction. For face images of the same age, lower weights are assigned to directions with large changes (non-correlated dimensions) through white transformation, so as to ensure that face images with the same age maintain a small dispersion in low-dimensional space. On the other hand, the embodiments of the present invention apply regression analysis to age estimation. Specifically, firstly, the dimensionality reduction of the sample is performed by principal component analysis. Second, use RCA to reduce the non-correlation of the sample set, and further select the appropriate discriminant dimensionality reduction space. Apply quadratic regression in a dimensionality reduction subspace to build regression and prediction functions.

The embodiment of the present invention can be applied to estimating the age of a person, providing technical support for marking in video image storage and organizing image databases, and can also narrow the scope of retrieval for suspect retrieval, etc., and improve retrieval efficiency. In addition, the embodiments of the present invention can provide a technical basis for face recognition, age classification, and the like.

It should be noted that the embodiment of the present invention describes that PCA is used for dimensionality reduction, but those skilled in the art understand that other dimensionality reduction methods can also be applied to the present invention, and thus are also within the scope of the present invention.

Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the relationship between hardware and software Interchangeability. In the above description, the composition and steps of each example have been generally described according to their functions. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present invention.

The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be implemented by hardware, software modules executed by a processor, or a combination of both. Software modules can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other Any other known storage medium.

Although some embodiments of the present invention have been shown and described, those skilled in the art will understand that various modifications can be made to these embodiments without departing from the principles and spirit of the invention, and such modifications shall fall within into the scope of the present invention.

Claims (13)

1, the age estimation method in a kind of face identification system is characterized in that, comprising:

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

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

To the constituent analysis of being correlated with of described low-dimensional characteristic, to obtain training data;

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

Utilize the described user's of described regretional analysis parameter estimation age.

2, age estimation method as claimed in claim 1 is characterized in that, the subclass of the described initial characteristic data of described selection comprises as the low-dimensional characteristic: adopt principal component analysis (PCA) to select the subclass of described initial characteristic data as described low-dimensional characteristic.

3, age estimation method as claimed in claim 1 is characterized in that, described the constituent analysis of being correlated with comprises to obtain training data to described low-dimensional characteristic:

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

Utilize the equivalent samples subclass of being constructed, described low-dimensional characteristic is differentiated dimensionality reduction, to obtain by the dimensionality reduction sample data of further dimensionality reduction; Reduce the non-correlation of described dimensionality reduction sample data, to obtain described training data.

4, age estimation method as claimed in claim 3 is characterized in that, the equivalent samples subclass that described utilization is constructed is differentiated dimensionality reduction to described low-dimensional characteristic and comprised:

Determine the interior covariance matrix of class of described equivalent samples subclass;

As Estimation of covariance matrix value in the class of described low-dimensional characteristic, utilize combination Fisher discriminatory analysis that described low-dimensional characteristic is differentiated dimensionality reduction covariance matrix in the class of described equivalent samples subclass.

5, age estimation method as claimed in claim 3 is characterized in that, the non-correlation of the described dimensionality reduction sample data of described minimizing comprises:

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

6, age estimation method as claimed in claim 3 is characterized in that, the non-correlation of the described dimensionality reduction sample data of described minimizing comprises:

Determine the interior covariance matrix of class of described dimensionality reduction sample data;

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

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

7, the estimation of Age equipment in a kind of 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 described initial characteristic data is as the low-dimensional characteristic, and the dimension of wherein said low-dimensional characteristic is less than the dimension of described initial characteristic data;

Relevant component analysis module is used for to the constituent analysis of be correlated with of described low-dimensional characteristic, 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 described training data that described relevant component analysis module obtained;

Estimation module is used to utilize described regretional analysis parameter to estimate described user's age.

8, estimation of Age equipment as claimed in claim 7 is characterized in that, described feature selection module adopts principal component analysis (PCA) to select the subclass of described initial characteristic data as described low-dimensional characteristic.

9, estimation of Age equipment as claimed in claim 7 is characterized in that, described relevant component analysis module comprises:

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

Differentiate the dimensionality reduction unit, be used to utilize the equivalent samples subclass of being constructed, described low-dimensional characteristic is 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 described dimensionality reduction sample data, to obtain described training data.

10, estimation of Age equipment as claimed in claim 8, it is characterized in that, the interior covariance matrix of the class of described equivalent samples subclass is determined in described differentiation dimensionality reduction unit, as Estimation of covariance matrix value in the class of described low-dimensional characteristic, utilize combination Fisher discriminatory analysis that described low-dimensional characteristic is differentiated dimensionality reduction covariance matrix in the class of described equivalent samples subclass.

11, estimation of Age equipment as claimed in claim 8 is characterized in that, described 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.

12, estimation of Age equipment as claimed in claim 8 is characterized in that, described decorrelation module is carried out leucismus to described dimensionality reduction sample data and changed, to obtain described training data.

13. a face identification system is characterized in that, comprises as each described estimation of Age equipment of claim 7 to 12.

Images