CN105956571A - Age estimation method for face image - Google Patents

Abstract

The invention relates to an age estimation method for a face image. The age estimation method comprises the following five parts: (1) an image is blocked; (2) a matrix after blocking processing is learnt by using a PCA algorithm so as to acquire a convolution kernel for a convolution operation; (3) the acquired convolution kernel is learnt by using the PCA algorithm so as to carry out the convolution operation; (4) nonlinear processing is carried out behind a second convolution layer by using a mode of binarization; and (5) feature extraction is carried out by using a method of histogram statistics. According to the method, an age value is estimated by using nonlinear K-SVR (Kernel function Support Vector Regression) after feature extraction, and experiments show that the accuracy of age estimation can be greatly improved.

Description

A kind of age estimation method of facial image

Technical field

The present invention relates to face estimation of Age field, the age of a kind of facial image based on degree of depth learning model Method of estimation.

Background technology

In face age automatic estimating system, being generally divided into two stages, first stage is to extract age characteristics, the Two stages are to estimate the age, generally research it is important that how to extract optimal age characteristics.Existing estimation of Age system In, learn or extract the method for age characteristics to can be generally divided into manual extraction age characteristics and automatic learning age feature.Hands The dynamic method extracting age characteristics representative has LBP (Local Binary Patterns, LBP), SIFT (Scale- Invariant Feature Transform, SIFT), subspace model etc., the shortcoming that manual extraction feature exists is by people The impact selected for subjectivity.Although manually selecting feature may obtain extraordinary effect in some field, being primarily due to can To select the feature adapted to for a certain specific data type, but facing to new data or new under conditions of, manually Select feature mode to differ to be suitable for surely, be in this way difficult to adapt to the application in reality.

The degree of depth study that nearly this year rises, is also used for learning or extracting age characteristics.Degree of depth study is in machine learning A field, its motivation setting up, the simulation neutral net that is analyzed of human brain, the mechanism that it imitates human brain explains number According to.Degree of depth study structure contains many hidden layers, its by combination low-level feature formed more abstract high-rise represent attribute classification or Feature, to find that the distribution characteristics of data represents.But existing degree of depth study structure parameter when hidden layer number increases sharply increases, Operational performance is proposed higher requirement.This is also in the problem using degree of depth study to be faced and need to solve.

Summary of the invention

Present invention is primarily targeted at and overcome drawbacks described above of the prior art, propose a kind of based on degree of depth study The age estimation method of the facial image of DLPCANet model, the method can provide a kind of efficient learning age feature, have The too much problem of solution model parameter of effect, and realize Fast Learning and high accuracy estimation age.

The present invention adopts the following technical scheme that

A kind of age estimation method of facial image, it is characterised in that:

1) after the facial image of input being carried out pretreatment, then piecemeal, each block is gone equalization, processes successively The facial image of all inputs, thus obtain a big matrix comprising all pieces；

2) obtain the convolution kernel of ground floor convolution operation from big matrix learning with PCA algorithm；

3) the output characteristic maps after ground floor convolution operation carried out piecemeal and goes equalization to process, obtaining another big square Battle array, uses PCA algorithm to obtain the convolution kernel of second layer convolution operation from this big matrix learning；

4) feature maps of output after second layer convolution operation is carried out Nonlinear Processing；

5) output valve after conversion is divided into some pieces, each block rectangular histogram is added up, by all pieces of connections Become a vector, obtain the feature of input picture；

6) feature of input picture is estimated the age with non-linear support vector regression.

Preferably, in step 1) described in pretreatment be binaryzation, smooth and standardization, obtain gray level image.

Preferably, in step 1) in, it is assumed that input picture, after pretreatment, obtains the gray-scale map that size is m × n-pixel All gray level images obtained are carried out piecemeal by picture respectively, it is assumed that the size of each block is p₁×p₂, step in blocking process Cut down as s₁=s₂=1, then the i-th-th opens the piecemeal result of input picture and is expressed asWhereinTo A_iIn each block carry out equalization after be Processing all input pictures successively and obtain a big matrix, it is expressed asWhereinThe size representing each block is p₁p₂× 1,The size of representing matrix A is p₁p₂×Nm₁n₁, the span of i It is 1≤i≤N.

Preferably, in step 2) in, the described volume obtaining ground floor convolution operation with PCA algorithm from big matrix learning Long-pending core, specific as follows, it is assumed that the convolution kernel number of the 1st layer is L₁, PCA algorithm is so that following objective functions reconstructed error is Little, its object function is:

$\underset{V\∈R^{p_1{p}_2\×L_1}}{min}\left|\right|A-{\mathrm{VV}}^{T}A|{|}_F^2,s.t.V^{T}V=I_{L_1}$

WhereinFor unit matrix, its size is L₁×L₁, solve this object function and be matrix A A^TCarry out eigenvalue Decompose, choose L₁Characteristic vector corresponding to the eigenvalue of individual maximum is as follows as its expression formula of convolution kernel of convolution operation:

$\begin{array}{ccc}{W}_l^i={\mathrm{mat}}_{p_1{p}_2}\left(p_l\right(A^{T}A\left)\right)\∈R^{p_1\×p_2},& i=1,2,& l=1,2,\mathrm{....},L_1\end{array}$

WhereinRepresent handleIt is transformed into matrixForm, p_l(XX^T) represent solution matrix AA^T Main characteristic vector.

Preferably, in step 3) in, it is assumed that the convolution operation of second convolutional layer isWhereinTable Show the output valve of first convolutional layer, W_l ²It it is convolution kernel.

Preferably, in step 4) in, described Nonlinear Processing includes first binaryzation, is reconverted into decimal value, table Reaching formula isWherein δ represents binary conversion treatment,Span be

Preferably, in step 6) in, use the Selection of kernel function RBF in non-linear support vector regression.

From the above-mentioned description of this invention, compared with prior art, there is advantages that

The method of the present invention, carries out feature extraction based on degree of depth learning model, with non-linear support vector regression K-SVR (Kernel function Support Vector Regression) estimates age value, can effectively solve the problem that model parameter Too much problem, and realize Fast Learning and high accuracy estimation age.

Accompanying drawing explanation

Fig. 1 is the degree of depth learning model flow chart of the present invention, wherein patch-mean removal represent equalization, Convolutional Layer C1 represents that first convolutional layer, Convolutional Layer C2 represent second convolution Layer, Binary quantization and mapping represent binaryzation and mapping, Concatenated image Block-wise histogram represents the statistics with histogram of all pieces in connection image.

Fig. 2 is that input picture piecemeal (patch) processes schematic diagram.

Fig. 3 is the window form trying to achieve characteristic vector and being transformed into convolution kernel.

Fig. 4 is first convolutional layer operation chart.

Fig. 5 is second convolutional layer operation chart.

Fig. 6 is the schematic diagram of non-linear layer operation.

Fig. 7 is that feature extraction layer carries out piecemeal schematic diagram.

Fig. 8 is the sample schematic diagram of two age public data storehouse MORPH and FG-NET.

Fig. 9 is to estimate age value two public data storehouse MORPH and FG-NET parts.

Detailed description of the invention

Below by way of detailed description of the invention, the invention will be further described.

In FIG, it is known that, the present invention is age estimation method based on degree of depth learning model DLPCANet, this degree of depth Practise model and comprise three layers, be convolutional layer (The Convolutional Layer), non-linear layer (The Nonlinear respectively Process Layer), feature extraction layer (Feature Pooling Layer).

Referring to figs. 1 through Fig. 7, the age estimation method of a kind of facial image of the present invention, comprise the steps:

1) facial image of input being carried out pretreatment, including binaryzation, smooth and standardization, obtaining size is m × n picture All gray level images obtained are carried out piecemeal, as illustrated in fig. 2, it is assumed that the size of each block is by the gray level image of element respectively p₁×p₂, the paces in blocking process are s₁=s₂=1, blocking process can also regard that with a window size be s as₁= s₂The row and column of image is scanned by=1.Then the i-th-th opens the piecemeal result of input picture and is expressed asWhereinTo A_iIn each block go After equalization it isProcessing all input pictures successively and obtain a big matrix, it represents For:

$A=\[{\stackrel{\‾}{A}}_{i,1},{\stackrel{\‾}{A}}_{i,2},...\·,{\stackrel{\‾}{A}}_{i,m_1{n}_1}\]\∈R^{p_1{p}_2\×{\mathrm{Nm}}_1{n}_1}---\left(1\right)$

WhereinThe size representing each block is p₁p₂× 1,The size of representing matrix A is p₁p₂× Nm₁n₁, the span of i is 1≤i≤N.

2) obtain the convolution kernel of ground floor convolution operation from big matrix A learning with PCA algorithm.Wherein convolution kernel number is just It is that PCA algorithm tries to achieve characteristic vector.As it is shown on figure 3, be the window-shaped characteristic vector solved being transformed into convolution kernel W Formula.The convolution kernel number assuming the 1st layer is L₁, PCA algorithm is so that following objective functions reconstructed error is minimum, its object function For:

$\underset{V\∈R^{p_1{p}_2\×L_1}}{min}\left|\right|A-{\mathrm{VV}}^{T}A|{|}_F^2,s.t.V^{T}V=I_{L_1}---\left(2\right)$

WhereinFor unit matrix, its size is L₁×L₁, solving this object function is exactly to matrix A A in fact^TCarry out spy Value indicative is decomposed, and chooses L₁Characteristic vector corresponding to the eigenvalue of individual maximum is as the convolution kernel of convolution operation, and its expression formula is such as Under:

$\begin{array}{ccc}{W}_l^i={\mathrm{mat}}_{p_1{p}_2}\left(p_l\right(A^{T}A\left)\right)\∈R^{p_1\×p_2},& i=1,2,& l=1,2,\mathrm{....},L_i\end{array}---\left(3\right)$

WhereinRepresent handleIt is transformed into matrixForm, p_l(XX^T) represent solution matrix AA^T Main characteristic vector.

Fig. 4 is first convolutional layer operation chart, it is assumed that input picture isPCA Algorithm Learning is used to obtain the The convolution kernel of one layer convolution operation is W_l ¹, l=1,2 ...., L₁, use different convolution kernels to go with input picture and roll up Time long-pending, it is possible to obtain different features maps.So the convolution output of first convolutional layer can be write as:

$O_i^1={\{I_i*W_l^1\}}_{l=1}^{L_1},i=1,2,\mathrm{....},N---\left(4\right)$

Wherein, * represents convolution operation, output characteristic mapsSize and input I_iEqually, reason is to carry out convolution During operation, input picture is carried out edge zero padding operation so that the size after convolution is the same with input.

3) the output characteristic maps after ground floor convolution operation carried out piecemeal and go equalization, obtaining another big matrix, PCA algorithm is used to obtain the convolution kernel of second layer convolution operation from this big matrix learning.Assume the convolution of second convolutional layer Operation isWhereinRepresent the output valve of first convolutional layer, W_l ²It it is convolution kernel.Concrete: Fig. 5 table Show second convolutional layer operation chart, the output of first convolutional layerAs the input of second convolutional layer, and ground floor Convolution equally does same operation, by inputCan be expressed as after carrying out piecemeal and going equalization:

${\stackrel{\‾}{Y}}_i^l=\[{\stackrel{\‾}{y}}_{i,l,1},{\stackrel{\‾}{y}}_{i,l,2},\mathrm{....},{\stackrel{\‾}{y}}_{i,l,m_1{n}_1}\]\∈R^{p_1{p}_2\×m_1{n}_1}---\left(5\right)$

Wherein l=1,2 ... .L₂, process the input of all second layers in the same way, can obtain

$Y=\[Y^1,Y^2,....,Y^{L_1}\]\∈R^{p_1{p}_2\×{\mathrm{Nm}}_1{n}_1{L}_1}---\left(6\right)$

WhereinWith the spy corresponding to the dominant eigenvalue of PCA Algorithm for Solving matrix Y Levy the vector convolution kernel W as second convolutional layer_l ².So the convolution of second convolutional layer is output as:

$D_i^2={\{O_i^1*W_l^2\}}_{l=1}^{L_2}---\left(7\right)$

4) Fig. 6 represents non-linear layer operation chart, and feature maps of output after second layer convolution operation is carried out non-thread Property process, including first binaryzation, be reconverted into decimal value, expression formula is

$B_i^l=\underset{l=1}{\overset{L_2}{\Σ}}{2}^{l-1}\mathrm{\δ}(O_i^1*W_l^2)---\left(8\right)$

Wherein δ represents binary conversion treatment,Span be

5) output valve after conversion is divided into some pieces, each block rectangular histogram is added up, by all pieces of connections Become a vector, obtain the feature of input picture.Particularly as follows: with reference to Fig. 7, be characterized abstraction, layer and carry out piecemeal schematic diagram, non-thread The output of property layerAs the input value of feature extraction layer,Being divided into Z-block, its size is h₁×h₂.Each block is used straight The method of side's figure statistics is added up, and then connects into a vectorial form all of piece, and it is expressed asProcess successivelyL=1,2 ...., L₁, input picture I may finally be obtained_iFeature be:

$f_i=\[BLKHist\left(B_i^1\right),BLKHist\left(B_i^2\right),\mathrm{....},BLKHist\left(B_i^{L_1}\right)\]\∈R\left(2^{L_2}\right)L_{1}Z---\left(9\right)$

Process the sample of all inputs in the same way, the feature of input sample may finally be extracted, so using Non-linear support vector regression.

6) by the feature of input picture with non-linear support vector regression K-SVR (Kernel function Supportive Vector Regression, K-SVR) estimate the age.Use the core letter in non-linear support vector regression Number selects RBF RBF (Radial-Basis Function).

Illustrate

With reference to Fig. 8, face age public data storehouse MORPH and FG-NET, the facial image of the two lane database has Corresponding age information.MORPH data base comprises 55332 facial images, also comprises the most agnate people, age simultaneously From 16 years old to 77 years old.FG-NET data base has 1002 facial images, and the image of the inside has highly variable (such as illumination, appearance Gesture, expression etc.), age several years was from 0 years old to 69 years old.

MORPH data base is randomly selected 5000 images as 5000 images of training set as test set, and for Little FG-NET data base then randomly selects the image of 80% as training set, and remaining 20% as test set.All of figure As based on becoming the gray level image that size is 28*28 after pretreatment.Finally predict the outcome and be respectively as follows: at MORPH lane database Age mean absolute error reached for 4.72 (ages), and the age mean absolute error at FG-NET lane database reached for 4.66 (years Age), wherein being calculated as of mean absolute errorT_kRepresent real age value, P_kThe year estimated Age is worth.Fig. 9 is to estimate age value two public data storehouse MORPH and FG-NET parts.

Above are only the detailed description of the invention of the present invention, but the design concept of the present invention is not limited thereto, all utilize this Design carries out the change of unsubstantiality to the present invention, all should belong to the behavior invading scope.

Claims (7)

1. the age estimation method of a facial image, it is characterised in that:

1) after the facial image of input being carried out pretreatment, then piecemeal, each block is gone equalization, processes all successively The facial image of input, thus obtain a big matrix comprising all pieces；

2) obtain the convolution kernel of ground floor convolution operation from big matrix learning with PCA algorithm；

3) the output characteristic maps after ground floor convolution operation carried out piecemeal and go equalization to process, obtaining another big matrix, PCA algorithm is used to obtain the convolution kernel of second layer convolution operation from this big matrix learning；

4) feature maps of output after second layer convolution operation is carried out Nonlinear Processing；

5) output valve after conversion is divided into some pieces, each block rectangular histogram is added up, connects into one by all pieces Individual vector, obtains the feature of input picture；

6) feature of input picture is estimated the age with non-linear support vector regression.

The age estimation method of a kind of facial image the most as claimed in claim 1, it is characterised in that: in step 1) described in Pretreatment is binaryzation, smooth and standardization, obtains gray level image.

The age estimation method of a kind of facial image the most as claimed in claim 1, it is characterised in that: in step 1) in, it is assumed that Input picture, after pretreatment, obtains the gray level image that size is m × n-pixel, is entered respectively by all gray level images obtained Row piecemeal, it is assumed that the size of each block is p₁×p₂, the paces in blocking process are s₁=s₂=1, then the i-th-th opens input figure The piecemeal result of picture is expressed asWhereinRight A_iIn each block carry out equalization after beProcess all input pictures successively to obtain To a big matrix, it is expressed asWhereinRepresent the size of each block For p₁p₂× 1,The size of representing matrix A is p₁p₂×Nm₁n₁, the span of i is 1≤i≤N.

4. the age estimation method of a kind of facial image as described in claim 1 or 3, it is characterised in that: in step 2) in, institute The convolution kernel obtaining ground floor convolution operation with PCA algorithm from big matrix learning stated, specific as follows, it is assumed that the volume of the 1st layer Long-pending core number is L₁, PCA algorithm is so that following objective functions reconstructed error is minimum, and its object function is:

$\underset{V\∈R^{p_1{p}_2\×L_1}}{min}\left|\right|A-{\mathrm{VV}}^{T}A|{|}_F^2,s.t.V^{T}V=I_{L_1}$

WhereinFor unit matrix, its size is L₁×L₁, solve this object function and be matrix A A^TCarry out Eigenvalues Decomposition, Choose L₁Characteristic vector corresponding to the eigenvalue of individual maximum is as follows as its expression formula of convolution kernel of convolution operation:

$\begin{array}{ccc}{W}_l^i={\mathrm{mat}}_{p_1{p}_2}\left(p_l\right(A^{T}A\left)\right)\∈R^{p_1\×p_2},& i=1,2,& l=1,2,\mathrm{....},L_1\end{array}$

WhereinRepresent handleIt is transformed into matrixForm, p_l(XX^T) represent solution matrix AA^TMaster Characteristic vector.

The age estimation method of a kind of facial image the most as claimed in claim 4, it is characterised in that: in step 3) in, it is assumed that The convolution operation of second convolutional layer isWhereinRepresent the output valve of first convolutional layer,It it is volume Long-pending core.

The age estimation method of a kind of facial image the most as claimed in claim 5, it is characterised in that: in step 4) in, described Nonlinear Processing include first binaryzation, be reconverted into decimal value, expression formula isWherein δ Represent binary conversion treatment,Span be

The age estimation method of a kind of facial image the most as claimed in claim 1, it is characterised in that: in step 6) in, use Selection of kernel function RBF in non-linear support vector regression.