CN105095833A - Network constructing method for human face identification, identification method and system - Google Patents

Abstract

The invention discloses a deeper layer network constructing method used for gender identification or age estimation based on human face. The method includes a step (101) dividing all training pictures into a plurality of groups; (102) extracting high layer features of a group of pictures based on a convolution neural network and thereby obtaining a first matrix composed of the high layer feature vectors, and extracting low layer and global features of the same group of the training images based on an artificial neural network and thereby obtaining a second matrix composed of the low layer feature vectors, obtaining a group of gender identification or age estimation results based on the extract first matrix, the second matrix and the defined judgment formula, wherein the values of a first weight matrix W1, a second weight matrix w2, an offset matrix b and an adjusting weight beta in the defined judgment formula are updated by utilizing an error back propagation algorithm and the final values of the parameters are obtained and the network construction is completed. Judgment of age and gender of a human face is performed based on the judgment formula determined according to the values of the parameters when the network construction is completed.

Description

For the network establishing method of recognition of face, recognition methods and system

Technical field

The present invention relates to computer vision and degree of depth learning art field, particularly a kind of network establishing method for recognition of face, recognition methods and system.

Background technology

Face concerning identifying computing machine and analyzing all very difficult object, just causes the extensive concern of researchers as one from the nineties in 20th century.And the effective human face analysis of success is in intelligent monitoring, the fields such as video index and people information statistics also exist again huge application prospect.The age of sex identification and face that human face analysis mainly comprises face is estimated, the mean absolute error of the accuracy rate of Gender Classification and age estimation is both critical index respectively.

The correlative study in the human face analysis field of current existence, be all determine that Feature Descriptor represents face based on artificial " craft ", combining classification device algorithm or regression algorithm launch.The early-stage preparations time of artificial selected feature often at substantial, there is subjectivity, and the feature choosing out often in a certain class data performance good, and when expanding to other data, performance can be decreased significantly, and generalization ability is more weak.And during practical application, the weak generalization ability of classic method is just not good in the lower picture of quality (intense light irradiation picture, there is background interference picture, wry face side face picture) upper performance, can not meet the demand of practical application.The research in current human face analysis field is limited to this bottleneck and slower development.On the other hand, in recent years, the method based on degree of depth study thoughts achieves great development, is the support that computer science provides on algorithm to intelligent direction development.The basic thought of degree of depth study is exactly the artificial neural network building deep layer, the study mechanism of simulation human brain, adopt the feature of mode " automatically " the learning objective object of unsupervised learning, the feature learnt has hierarchical structure, from detail to abstract concept, such feature has more essential portraying to data itself.The method application in a lot of fields of degree of depth study all achieves breakthrough success, and the full-automatic simultaneous interpretation system of the Handwritten Digit Recognition System of Duo Jia bank of the U.S., the Images Classification speech recognition integrated project GoogleBrain of Google, Microsoft is all that the method learnt based on the degree of depth realizes.The method of current degree of depth study mainly comprises deep layer sparse own coding algorithm, degree of depth belief network algorithm and convolutional neural networks algorithm.Wherein convolutional neural networks algorithm all achieves level advanced in the world in multiple fields of image procossing, such as Face datection, pedestrian detection and recognition of face etc.; And degree of depth belief network algorithm is mainly more in the application of field of speech recognition, effect is better; The sparse own coding algorithm of deep layer is more applied to the problem of Data Dimensionality Reduction class.

Convolutional neural networks essence is a kind of neural network structure of deep layer, and the ultimate principle of its algorithm and feature are that network structure is made up of two kinds of part and parcels: the convolutional layer that convolution processing unit is formed, and the down-sampling layer that down-sampling processing unit is formed.Convolutional layer and down-sampling layer form two-dimensional structure by neuron, process of convolution and down-sampling process are carried out respectively to the input picture being two-dimensional structure equally, then Convolution sums down-sampling is repeated, until " ideal " (the needing to determine according to research application) extracting image is after feature, then carry out classifying or returning or detect.When input picture is N × N size, first connect convolutional layer C ₁, C ₁in each neuron be only connected with one piece of local acceptance domain (also known as making convolution kernel, filter) of last layer.Suppose that the size of convolution kernel is m × m, so C ₁layer with the pixel of convolution kernel convolution input picture all possible (N-m+1) × (N-m+1) position of this m × m, will generate the local feature figure of one (N-m+1) × (N-m+1).All pixels in each block of input picture m × m size area and C ₁in layer, a neuron is connected, and identical weights (i.e. weights shared mechanism) are taked in this m × m connection.When using the connected mode of multiple different weights to generate multiple local features figure, C ₁layer just extracts the different local features of former figure.Then C ₁the local feature figure of layer is connected to the down-sampling layer S of lower one deck ₁.Suppose C ₁layer has F ₁open characteristic pattern, so corresponding C ₁layer also has F ₁open down-sampling figure, and and C ₁the characteristic pattern one_to_one corresponding of layer.S ₁each neuron in each figure of layer connects one piece of local field of this last layer, and does not have overlap, then calculates the value of mean value as sample level of all values in this block region.Suppose C ₁often open the size of characteristic pattern in layer for (N-m+1) × (N-m+1), local join domain size is n × n, so S ₁the size of each down-sampling figure of layer is (N-m+1)/n × (N-m+1)/n.So namely, achieve the down-sampling to last layer characteristic pattern, reduce spatial resolution.Then S ₁the convolutional layer C that layer is new with one again ₂layer connects, C ₂layer and S ₂layer connects, and so intersects repeatedly, determines the number of convolutional layer and the number of down-sampling layer according to actual needs.The output of last network is called output layer, and may be export different classifications according to the difference of problem, also may be the probability that output regression is estimated.

But, although the convolution of convolutional neural networks+down-sampling processing mode can extract the high level of target object, abstract feature, achieve good effect for during some classification problem, have ignored some low layers to the effective characteristic sum information of classification.And the local feature of mainly object that describes of the characteristic pattern that extracts of convolution+down-sampling process and local relevance, lack the performance to object global feature.When solving gender classification and face age estimation problem, exactly need to carry out comprehensive, multi-level feature extraction and description to face, accurate, comprehensive information can be obtained.

Based on above-mentioned, the method of Feature Descriptor is determined in traditional " craft " of application of human face analysis field, or directly apply convolutional neural networks (comprising the method for other degree of depth existing study), their performances and final effect all can not meet the needs of practical application.The present invention applies the brand-new method based on the study of the multiple features degree of depth and carries out human face analysis, build the network of training deep layer, " automatically " learn, extract face by different level, omnibearing feature, the structure description list forming a kind of multiple features (high-rise, low layer, abstract, concrete) is leted others have a look at face.Such multiple features structure achieves extraordinary effect when being applied to final Gender Classification and age estimation.

Still lack these class methods or system in the prior art.

Summary of the invention

The object of the invention is to, for the performance overcoming the classic method in the technology of existing human face analysis can not meet the defect that practical application needs, thus a kind of network establishing method for recognition of face is proposed and based on the face identification method of this structure network and system.

To achieve these goals, the invention provides a kind of deep layer network establishing method estimated for gender classification or age, described method comprises:

Step 101) all training pictures are divided into some groups;

Step 102) extract based on convolutional neural networks the high-level characteristic that a group is trained picture, and then obtain by vectorial the first matrix formed of high-level characteristic; Extract low layer and the global characteristics of same group of training picture simultaneously based on artificial neural network, and then obtain the second matrix of low-level feature vector composition;

The result of one group of sex identification or age estimation is obtained based on the first matrix extracted, the second matrix and following judgement formula:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, hfo represents the first matrix; Lfo represents the second matrix; For the first weight matrix w in first group of above-mentioned formula of training picture of input ₁, the second weight matrix w ₂, bias matrix b and regulate the initial value of weight beta to adopt random initializtion mode to obtain; For the w in all the other above-mentioned formula of training picture respectively organized of input ₁, w ₂, b and β acquisition methods be: utilize error backpropagation algorithm to calculate error function J (W, b that court verdict o and each group train the physical tags matrix Y of picture; β), then by calculate w ₁, w ₂, b and β be to error function J (W, b; Gradient β) and then undated parameter w ₁, w ₂, b and β value;

Step 103) input again one group training picture, and above-mentioned steps 102 is repeated to the training picture again inputted), until all groupings all carry out steps 102) process, complete and once train iteration;

Step 104) all training pictures are reclassified as some groups, and repeat above-mentioned steps 102 to each group that repartitions) and step 103), complete iteration again;

Through some grouping and iterative processing again, until when the final judgement o exported meets the condition of setting, obtain final parameter w ₁, w ₂, b and β value, complete network struction.

Optionally, comprise further when carrying out low-level feature abstract:

Step 102-11) each Zhang Xunlian picture of one group of input training picture is converted into vectorial form by the graph structure form of two dimension, then vector is normalized, obtain the original feature vector of each Zhang Xunlian picture;

Step 102-12) original vector of each the Zhang Xunlian picture obtained is inputted artificial neural network, and then obtain one group of reconstruction features vector, namely obtain the second described matrix; Wherein, described artificial neural network comprises L layer, and adopts full connected mode between layers, and each neuron of every one deck adopts sigmoid function to activate.

Optionally, the sex of picture or age identifying is trained specifically to comprise for an input:

Step 102-21) when extract one training picture high-level characteristic vector be H _fthe high-level characteristic vector of dimension, and low-level feature vector is L _fduring the proper vector tieed up, construct one and comprise " H _f+ L _f" individually neuronicly combine voting layer;

Step 102-22) when for sex identification, each neuron of combining voting layer of structure is connected with two output neurons of output layer respectively, and each output neuron carries out Sex Discrimination based on described judgement formula, exporting training picture is the probability of sex; When estimating for the age, combine voting each neuron of layer and be connected with S output neuron of output layer, wherein each output neuron correspondence is one-year-old.

Optionally, the error backpropagation algorithm of following formula is adopted to upgrade the first weight matrix w ₁value:

$\left\{\begin{array}{c}{\left(w_1\right)}_{\mathrm{new}}={\left(w_1\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_1\\ {\mathrm{\Δw}}_1=\mathrm{Od}*\mathrm{hfo}\end{array}\right.$

Wherein, (w ₁) _newrepresent the first weight matrix w after upgrading when error back propagation each time ₁value, (w ₁) _oldthe first weight matrix w before corresponding renewal ₁value, Od represents output layer sensitivity matrix, and this output layer sensitivity matrix adopts error function J (W, b; β) obtain in conjunction with court verdict o; α represents the learning rate of network, wherein the value of α is initialized as a larger value, then reduces gradually with the increase of training iterations;

The second weight matrix w is upgraded by following formula ₂value:

$\left\{\begin{array}{c}{\left(w_2\right)}_{\mathrm{new}}={\left(w_2\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_2\\ {\mathrm{\Δw}}_2=\mathrm{Od}*\mathrm{hfo}\end{array}\right.$

Wherein, (w ₂) _newrepresent the second weight matrix w after upgrading when error back propagation each time ₂value, (w ₂) _oldthe second weight matrix w before corresponding renewal ₂value.

Optionally, the value of described β training update method in iteration is each time:

$\left\{\begin{array}{c}{\mathrm{\β}}_{\mathrm{new}}={\mathrm{\β}}_{\mathrm{old}}-\mathrm{\α}\·\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}\\ 0\≤{\mathrm{\β}}_{\mathrm{new}},{\mathrm{\β}}_{\mathrm{old}}\≤1\end{array}\right.$

Wherein, β _newrepresent the value of the adjustment weight beta after upgrading when error back propagation each time, β _oldthe value of the adjustment weight beta before corresponding renewal;

Ask the part of local derviation can be obtained by following formula in above-mentioned formula:

$\left\{\begin{array}{c}\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}=\mathrm{mean}\left(B(:)\right)\\ B=(w_2*\mathrm{lfo})\·*f^{\text{'}}\left(o\right)\\ f^{\text{'}}\left(o\right)=o\·*(1-o)\end{array}\right.$

Wherein, f ' (o) represents court verdict o differentiate, and " mean (B (:)) " represents and to average computing to all elements in matrix B; Matrix B represents the matrix be made up of the value of the adjustment weight beta after upgrading when error back propagation each time, and the ranks number of this matrix B is identical with the ranks number of court verdict o.

Based on the deep layer network of above-mentioned structure, present invention also offers a kind of for the age of face or the recognition methods of sex, described method comprises:

For extracting the high-level characteristic of face picture to be identified based on convolutional neural networks;

For extracting low layer and the global characteristics of face picture to be identified based on artificial neural network;

For the low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be the value that deep layer network establishing method is determined, hfo represents the high-level characteristic vector of the face to be identified of extraction, and lfo represents the low-level feature vector of the face to be identified of extraction, and o represents the court verdict at sex or age.

Following steps are adopted to extract low-level feature:

For the face picture to be identified of input is carried out flaky process and after normalization, is obtained the original feature vector of face;

For original feature vector is inputted artificial neural network, by multilayer neuronal structure, input vector is rebuild, obtain a L _fthe proper vector of dimension is as the low-level feature vector extracted.

In addition, the invention provides a kind of for the age of face or the recognition system of sex, described system comprises:

High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks;

Low-level feature abstract module, for extracting low layer and the global characteristics of face picture to be identified based on artificial neural network;

Based on the judging module of neural network, for the low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be for training picture several times iteration after obtain (the final value of each parameter obtained when namely above-mentioned network struction completes), hfo represents the high-level characteristic vector of the face to be identified of extraction, lfo represents the low-level feature vector of the face to be identified of extraction, and o represents the court verdict at sex or age.

Optionally, above-mentioned low-level feature abstract module comprises further:

Flaky process module, for carrying out flaky process by the face picture to be identified of input and after normalization, obtain the original feature vector of face;

Reconstruction features vector acquisition module, for original feature vector is inputted artificial neural network, is rebuild input vector by multilayer neuronal structure, obtains a L _fthe proper vector of dimension is as the low-level feature vector extracted.

Optionally, above-mentioned judging module comprises further:

Combining voting layer module, for merging the high-level characteristic and low-level feature that extract, exporting a kind of multiple features structure;

Output layer module, for adopting some output neurons to carry out sex judgement or age judgement, each described neuron is adjudicated based on described judgement formula.

Compared with determining the method for Feature Descriptor with traditional " craft ", technical advantage of the present invention is:

Deep layer network provided by the invention can " automatically " study face feature, and the feature learnt has hierarchy, such feature has more essential portraying to data itself, thus the such feature of final utilization carries out classifying and regression estimates time effect also better.Further, compared with the method learnt with the existing degree of depth, deep layer network of the present invention can not only learn the higher level of abstraction feature of face, have learned the low layer global characteristics of face simultaneously, to let others have a look at face in conjunction with two kinds of comprehensive, multi-level description lists, when classification final like this and regression estimates, the method that Performance Ratio of the present invention is only extracted the existing degree of depth study of independent high-level characteristic wants better.The method that this degree of depth based on multiple features of the present invention learns, when being applied to human face analysis field, not only has extremely strong learning ability, also has extremely strong generalization ability.In the high quality standards face picture of test common data sets, when the network face picture that quality is lower and the actual face picture that watch-dog gathers, all achieve the performance exceeding prior art.Therefore the present invention can meet the needs of actual persons face analysis application.

Accompanying drawing explanation

Fig. 1 is gender classification process flow diagram provided by the invention;

Fig. 2 estimates process flow diagram at the face age provided by the invention;

The structural representation of the human face analysis system that Fig. 3 provides for the invention process example.

Embodiment

Now the invention will be further described by reference to the accompanying drawings.

The present invention gives a kind of human face analysis method based on the study of the multiple features degree of depth, comprising:

Step 1), Face datection and pre-service are carried out to picture.

Step 2), to step 1) face picture that obtains carries out human face analysis, is input to gender classification deep layer network respectively and the face age estimates deep layer network.

Step 3), the sex of the face picture of the gender classification deep layer network prediction of output, man or female; The age round values of the face picture of face age estimation deep layer network output estimation, how old.

One, the structure of gender classification network:

In technique scheme, described step 2) provide a kind of face gender identification method, as shown in Figure 1, i.e. a kind of deep layer network system for gender classification, comprising:

Step 2-1) using step 1) in the face picture that obtains as the input of network, the output of network is gender prediction's value of people in picture (man or female).Network forms primarily of 4 partial function modules, comprises high-level characteristic extraction module, low-level feature abstract module, fusion feature cascading judgement output module and parameter training module.

In such scheme, described step 2-1) specifically comprise the steps:

Step 2-1-1), high-level characteristic extracts: directly adopt the high-level characteristic extraction that the convolutional layer of the convolutional neural networks in existing degree of depth learning method and down-sampling Rotating fields realize input training picture.In the specific implementation, 3 layers of convolutional layer C are adopted ₁, C ₂and C ₃, and two-layer down-sampling layer S ₁, S ₂combination, entirely connect between layers.

Step 2-1-2), low-level feature abstract: synchronously with high-level characteristic extraction module to process inputting training image.

First, the training picture of the face of input is converted into the form (be called flaky process, flat operates) of vector by the graph structure form of two dimension.Again the vector obtained is normalized after flat operation is carried out to input picture, obtains the original feature vector of face.

Then, original feature vector is connected to reconstruction network to obtain reconstruction features vector.Rebuild network to build based on artificial neural network principle, altogether L layer.The neuron of every one deck exports can regard proper vector as, the neuron of lower one deck carries out recompile to the vector that last layer exports, again represent and export after describing again, take the full mode connected between layers, each neuron adopts sigmoid function to activate.First original feature vector is connected to the H rebuilding network ₁layer, H ₁layer is containing h ₁individual neural unit, suppose that input picture size is N × N, so face original feature vector is N × N dimension, through H ₁h is become after layer ₁dimensional feature vector.Then then H is connected to ₂layer, H ₂layer is containing h ₂individual neural unit, so proper vector becomes h further ₂dimension.By that analogy, according to actual needs, finally H is connected to _nlayer, obtains a h _nthe proper vector of dimension.Connection mathematical formulae is between layers expressed as:

a ^l+1＝sigm(W ^l·a ^l+b ^l)(1)

Wherein " sigm () " represents the matrix form (namely carrying out the activation of sigmoid function to each element in matrix) of sigmoid function, a ^l+1and a ^lrepresent the matrix form (being proper vector herein) of the output of (l+1) layer and l layer respectively, W ^lfor connecting (l+1) neuronic weight matrix between layer and l layer, b ^lrepresent the bias matrix of l layer.

This h _nnamely the proper vector of dimension is the output of rebuilding network, and also namely reconstruction features is vectorial.Reconstruction features vector eliminates redundant information invalid to Gender Classification in face original feature vector, decreases partial noise interference, can portray low layer and the global characteristics of face preferably.

Step 2-1-3), cascading judgement exports: combined extracting to high-level characteristic carry out the judgement of final sex together with low-level feature, export gender prediction's value (man or female).

Vector form is converted into after the high-level characteristic figure that high-level characteristic Extraction parts obtains carries out flat operation, the reconstruction features vector obtained with the low-level feature abstract formation that links together combines voting layer, combines voting layer and is still connected with last layer based on artificial neural network principle.Suppose that high-level characteristic extracts and finally obtain the characteristic pattern that G opens q × q size, a G × q × q can be obtained after being so converted into vector and tie up high-level characteristic vector; Reconstruction features vector is h _ndimension; So combine in voting layer namely containing (G × q × q+h _n) individual neuron, high-level characteristic vector sum reconstruction features vector is merged, forms a kind of vector of multiple features version.Namely the output of combining voting layer be the (G × q × q+h of the multiple features structure that our whole network extraction arrives _n) dimensional feature vector.

Combine voting layer and be entirely connected to final output neuron again, output neuron has two (men and women two classes), output be the Probability p of a certain class of net result _i, output neuron adopts sigmoid function to activate, and so the probability of each class can be expressed as:

$p_i=\mathrm{sigm}(\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{ik}}^o{x}_i^{\mathrm{jv}}+b_k^o)---\left(2\right)$

Wherein represent and combine the neuronic output of voting layer kth, represent and combine that voting layer i-th neuron and output layer kth are individual is neuronicly connected weights, for output layer correspondence is biased.

Because multiple training picture (parameter training part has respective description) of each input of whole network, therefore the matrix form of the judgement Output rusults of the network of the present invention's definition is:

o＝sigm(w ₁*hfo+w ₂*lfo+b)(3)

Wherein, o represents the court verdict (court verdict of a sample is shown in each list) that network exports; w ₁represent the weight matrix that output layer is connected with the output of high-level characteristic Extraction parts, i.e. the first weight matrix; " * " representing matrix multiplication, hfo represents the output (output of a sample is shown in each list) of high-level characteristic Extraction parts; w ₂represent the weight matrix that output layer is connected with the output of low-level feature abstract part, i.e. the second weight matrix; Lfo represents the output matrix (output of a sample is shown in each list) of low-level feature abstract part; B represents output layer bias matrix.

Consider that two kinds of features are on the impact of end product, in the reconstruction features vector representing low-level feature, add one regulate weight beta, 0≤β≤1, regulates, the suppression that affect to a certain extent of low-level feature on the judgement of net result.The court verdict that then final network exports is:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)(4)

Step 2-1-4), adopt training method to get parms w ₁, w ₂, b and β final value: artificial neural network, the basic theories of degree of depth study is divided into training two parts of parameter in the design of network and network.When (namely obtaining above-mentioned high-level characteristic, low-level feature and judgement formula) after the structure designing network, training is needed to determine that in network, the value of each parameter (namely determines w ₁, w ₂, b and β value, and the value of parameter in convolutional layer and down-sampling layer, the value of each neuronic parameter in artificial neural network in L layer), the face picture of this network handles identification then could be used to carry out the application such as actual classification and recurrence.

Described training method takes error backpropagation algorithm, considers a large amount of training sample support of degree of depth Learning demands simultaneously, for reducing calculated load, trains in conjunction with stochastic gradient descent strategy.All T are opened training picture random division and be some groups, and every B opens one group (B wants to divide exactly T), " T/B " group altogether.To own in order in the above-mentioned network designed of " T/B " group input, and then extract high-level characteristic low-level feature and after carrying out sex judgement, complete and once train iteration; Then again all T are opened training picture random division and be some groups, remain every B and open one group, altogether " T/B " group.Each is all random division, after guaranteeing each division, each picture organized is not identical with the last time, still will own in order in the above-mentioned network designed of " T/B " group input, and then extract high-level characteristic low-level feature and after carrying out sex judgement, complete new once training iteration.Altogether carry out E training iteration and finally could obtain parameter w ₁, w ₂, b and β value.

Undated parameter w ₁, w ₂, b detailed process as follows:

First, random initializtion parameter w ₁, w ₂, b and β value, then input first group training picture obtain last court verdict o, in o, the judgement Output rusults of a sample is shown in each list.Then calculate the error of output layer, computing formula is as follows:

$\mathrm{MSE}=\frac{1}{2}{\left|\right|Y-o\left|\right|}_2---\left(5\right)$

Wherein, MSE represents the square error between court verdict o and actual sample label (classification is men and women herein) matrix Y, and MSE is matrix representation; Y is the label matrix of input amendment, if the face picture of input is the male sex, so Y is just [10] ^tmatrix, if women, is [01] then ^tmatrix; O is the court verdict that network exports, " || || ₂" corresponding element asks the matrix representation after the difference of two squares between representing matrix.

Then, this error is utilized can to calculate parameter w in formula (4) ₁, w ₂, b and β is for final error function J (W, b; Gradient β), utilizes Gradient Descent principle to upgrade above-mentioned parameter w ₁, w ₂, the value of b and β.Wherein error function J (W, b; Matrix form β) is MSE.For this reason, the sensitivity of output layer need be calculated:

Wherein, Od represents the matrix form of output layer sensitivity, dot product (corresponding element is multiplied, and dimension is consistent) between representing matrix.Wherein f ' (o) represents output function differentiate, and activation function adopts sigmoid function, and its derivative form is f ' (x)=f (x) (1-f (x)).Adopt following 3 formula can obtain parameter w further according to sensitivity ₁, w ₂and the value of biased b, namely to w ₁, w ₂and biased b upgrades:

$\left\{\begin{array}{c}{\left(w_1\right)}_{\mathrm{new}}={\left(w_1\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_1\\ {\mathrm{\Δw}}_1=\mathrm{Od}*\mathrm{hfo}\end{array}\right.---\left(7\right)$

Wherein, (w ₁) _newrepresent the first weight matrix w after upgrading when error back propagation each time ₁value, (w ₁) _oldthe first weight matrix w before corresponding renewal ₁value; α represents the learning rate of network, and the present invention takes learning rate changing Strategies Training, and namely the value of α is initialized as a larger value, then reduces gradually with the increase of training iterations, ensures the convergence of whole network.

$\left\{\begin{array}{c}{\left(w_2\right)}_{\mathrm{new}}={\left(w_2\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_2\\ {\mathrm{\Δw}}_2=\mathrm{Od}*\mathrm{hfo}\end{array}\right.---\left(8\right)$

Wherein, (w ₂) _newrepresent the second weight matrix w after upgrading when error back propagation each time ₂value, (w ₂) _oldthe second weight matrix w before corresponding renewal ₂value.

$\left\{\begin{array}{c}{\left(b\right)}_{\mathrm{new}}={\left(b\right)}_{\mathrm{old}}-\mathrm{\α}*\mathrm{\Δb}\\ \mathrm{\Δb}=\mathrm{Od}\end{array}\right.---\left(9\right)$

Wherein, (b) _newrepresent the value of the bias matrix b after upgrading when error back propagation each time, (b) _oldthe value of the bias matrix b before corresponding renewal.

The detailed process of undated parameter β is as follows:

Because β is a real number, take the more new formula of gradient descent method renewal β as follows:

$\left\{\begin{array}{c}{\mathrm{\β}}_{\mathrm{new}}={\mathrm{\β}}_{\mathrm{old}}-\mathrm{\α}\·\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}\\ 0\≤{\mathrm{\β}}_{\mathrm{new}},{\mathrm{\β}}_{\mathrm{old}}\≤1\end{array}\right.---\left(10\right)$

Error function asks the formula of local derviation can be turned to further by chain type rule to β:

$\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}=\frac{\∂J}{\∂o}\·\frac{\∂o}{\∂\mathrm{\β}}---\left(11\right)$

β and matrix (w ₂* lfo) be multiplied the matrix A and matrix (w that can regard a ranks number identical with matrix as ₂* lfo) dot product, wherein in A, element value is all β.Like this, formula (11) finally can turn to:

$\left\{\begin{array}{c}\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}=\mathrm{mean}\left(B(:)\right)\\ B=(w_2*\mathrm{lfo})\·*f^{\text{'}}\left(o\right)\\ f^{\text{'}}\left(o\right)=o\·*(1-o)\end{array}\right.---\left(12\right)$

Wherein " mean (B (:)) " represents and to average computing to all elements in matrix B, matrix B represents the matrix be made up of the value of the adjustment weight beta after upgrading when error back propagation each time, and the ranks number of this matrix B is identical with the ranks number of court verdict o.

Utilize error backpropagation algorithm in a word, the network parameter w in formula (4) can be completed ₁, w ₂, the renewal of b and β.Also comprise the parameter in convolutional layer in whole network, the parameter in the parameter in down-sampling layer and L layer artificial neural network needs the value being determined them by training.Institute's using method remains error backpropagation algorithm.Continue error MSE forward direction, now combine a voting layer part and be connected with high-level characteristic Extraction parts, a part is connected with reconstruction network, and so MSE is divided into two parts error, continues forward direction respectively in these two modules.In the error back propagation process of high-level characteristic Extraction parts, the method of the convolutional layer in the existing degree of depth learning art of direct employing and down-sampling layer error back propagation undated parameter, upgrades parameter in convolutional layer and the down-sampling layer { value of CS} ({ CS} represents the set of matrices of all parameters in convolutional layer and down-sampling layer).In the error back propagation process of low-level feature abstract part, the method of the error back propagation undated parameter in the existing artificial neural network technology of direct employing, upgrades parameter in the L layer artificial neural network { value of LN} ({ LN} represents the set of matrices of all parameters in L layer artificial neural network).This completes the parameter training process of a picture group sheet, when then inputting the second picture group sheet training, the network parameter w determined with the first picture group sheet ₁, w ₂, the value of b and β, and { CS} and { the value calculating court verdict o of LN}, then repetitive error back-propagation process, w in renewal network ₁, w ₂, b and β value, and { CS} and the { value of LN}.By that analogy, the value of the parameter more than training of each picture group sheet determined after one picture group sheet training calculates court verdict, and then from output, the error of calculation backpropagation, to input, upgrade the value of all parameters of whole network.All input after network completes training until " T/B " organizes training sample, just complete and once train iteration.

After completing all E training iteration, namely network now can be used for actual sex identification, and input face picture, network will export the predicted value of sex.

Two, the structure of face age estimation network:

In technique scheme, described step 2) provide a kind of method that face age estimates, as shown in Figure 2, i.e. a kind of deep layer network system estimated for the face age, comprising:

Step 2-2) using step 1) in the face training picture that obtains as the input of network, the output of network is the age estimated value (integer) of people in picture.Network forms primarily of 4 partial function modules, comprises high-level characteristic extraction module, low-level feature abstract module, fusion feature cascading judgement output module and parameter training module.

In such scheme, described step 2-1) comprising:

Step 2-2-1), high-level characteristic extracts: directly adopt the high-level characteristic extraction that the convolutional layer of the convolutional neural networks in existing degree of depth learning method and down-sampling Rotating fields realize for training picture.In the specific implementation, 3 layers of convolutional layer C are adopted ₁, C ₂and C ₃, and three layers of down-sampling layer S ₁, S ₂and S ₃combination, entirely connect between layers.

Step 2-2-2), low-level feature abstract: synchronous with high-level characteristic extraction module input picture to be processed.

First, the training picture of the face of input is converted into the form (be called flaky process, flat operates) of vector by the graph structure form of two dimension, then the vector obtained is normalized, obtain face original feature vector.

Then, original feature vector is connected to reconstruction network to obtain reconstruction features vector.Rebuild network to build based on artificial neural network principle, altogether L layer.The neuron of every one deck exports can regard proper vector as, the neuron of lower one deck carries out recompile to the vector that last layer exports, again represent and export after describing again, take the full mode connected between layers, each neuron adopts sigmoid function to activate.First face original feature vector is connected to the H rebuilding network ₁layer, H ₁layer is containing h ₁individual neural unit, suppose that input picture size is N × N, so face original feature vector is N × N dimension, through H ₁h is become after layer ₁dimensional feature vector.Then then H is connected to ₂layer, H ₂layer is containing h ₂individual neural unit, so proper vector becomes h further ₂dimension.By that analogy, according to actual needs, finally H is connected to _nlayer, obtains a h _nthe proper vector of dimension.The concrete calculating of connection between layers can be obtained by formula (1).

This h _nnamely the proper vector of dimension is the output of rebuilding network, and also namely reconstruction features is vectorial.Reconstruction features vector eliminates in face original feature vector estimates invalid redundant information to the age, decreases partial noise interference, can portray low layer, the global characteristics of face preferably.

Step 2-2-3), combine voting layer and output: the high-level characteristic of combined extracting and low-level feature carry out the final age and estimate, export age predicts value (how old).

Being connected to one containing M neuronic full articulamentum by extracting the high-level characteristic figure obtained, obtaining the high-level characteristic vector of a M dimension.Then high-level characteristic vector and the reconstruction features vector formation that links together combine voting layer.Suppose that reconstruction features vector is h _ndimension, namely so both are united forms one containing (M+h _n) individually neuronicly combine voting layer.Namely the output of combining voting layer be the (M+h of the multiple features structure that our whole network extraction arrives _n) dimensional feature vector.

Combine voting layer and be entirely connected to final output neuron again.Output neuron has S (each correspondence is one-year-old).Output layer still adopts sigmoid function to activate, and so the probability of each class can be expressed as formula (2).Because each input plurality of pictures (parameter training part has respective description) of whole network, therefore the output matrix form of the network of the present invention's definition can be expressed as formula (3).

Consider that two kinds of features are on the impact of end product, in the reconstruction features vector representing low-level feature, add one regulate weight beta, 0≤β≤1, regulates, the suppression that affect to a certain extent of low-level feature on the judgement of net result.Then final network exports and can be obtained by formula (4).

Step 2-2-4), adopt training method to get parms w ₁, w ₂, b and β final value: artificial neural network, the basic theories of degree of depth study is divided into training two parts of parameter in the design of network and network.When (namely obtaining above-mentioned high-level characteristic, low-level feature and judgement formula) after the structure designing network, training is needed to determine that in network, the value of each parameter (namely determines w ₁, w ₂, b and β value, and the value of parameter in convolutional layer and down-sampling layer, the value of each neuronic parameter in artificial neural network in L layer), the face picture of this network handles identification then could be used to carry out the application such as actual classification and recurrence.

Described training method takes error backpropagation algorithm, considers a large amount of training sample support of degree of depth Learning demands simultaneously, for reducing calculated load, trains in conjunction with stochastic gradient descent strategy.All T are opened training picture random division and be some groups, and every B opens one group (B wants to divide exactly T), " T/B " group altogether.To own in order in the above-mentioned network designed of " T/B " group input, and then extract high-level characteristic low-level feature and after carrying out sex judgement, complete and once train iteration; Then again all T are opened training picture random division and be some groups, remain every B and open one group, altogether " T/B " group.Each is all random division, after guaranteeing each division, each picture organized is not identical with the last time, still will own in order in the above-mentioned network designed of " T/B " group input, and then extract high-level characteristic low-level feature and after carrying out sex judgement, complete new once training iteration.Altogether carry out E training iteration and finally could obtain parameter w ₁, w ₂, b and β value.Undated parameter w ₁, w ₂, b detailed process as follows:

First, random initializtion parameter w ₁, w ₂, b and β value, then input first group training picture obtain last court verdict o, in o, the judgement Output rusults of a sample is shown in each list.Then formula (5) is used to calculate the error of output layer.It should be noted that the sample label matrix Y in formula (5) is S dimensional vector form in the age is estimated, if the age inputting face picture corresponding is 1 years old, so Y is then [10...0] ^t; If the age is 2 years old, Y is [01...0] ^t; If S year, be [00...1] then ^t.

Then, the error utilizing formula (5) to obtain can calculate the parameter w in formula (4) ₁, w ₂, b and β is for final error function J (W, b; Gradient β), utilizes Gradient Descent principle to upgrade above-mentioned parameter w ₁, w ₂, the value of b and β.Wherein error function J (W, b; Matrix form β) is MSE.For this reason, the sensitivity of output layer need be calculated, formula (6) can be had to obtain.According to sensitivity, adopt formula (7), formula (8) and formula (9) can obtain parameter w further ₁, w ₂and the value of biased b, namely to w ₁, w ₂and biased b upgrades.

Because β is a real number, when taking gradient descent method to upgrade β, formula (10) can be utilized to obtain the value of the β after renewal.Error function asks the formula of local derviation can be transformed further by formula (11) by chain type rule to β.Finally, β and matrix (w ₂* lfo) be multiplied the matrix A and matrix (w that can regard a ranks number identical with matrix as ₂* lfo) dot product, wherein in A, element value is all β.Like this, formula (11) finally can turn to formula (12), thus calculates the updated value of β.

Utilize error backpropagation algorithm in a word, the network parameter w in formula (4) can be completed ₁, w ₂, the renewal of b and β.Also comprise the parameter in convolutional layer in whole network, the parameter in the parameter in down-sampling layer and L layer artificial neural network needs the value being determined them by training, and institute's using method remains error backpropagation algorithm.Continue error MSE forward direction, now combine a voting layer part and be connected with high-level characteristic Extraction parts, a part is connected with reconstruction network, and so MSE is divided into two parts error, continues forward direction respectively in these two modules.In the error back propagation process of high-level characteristic Extraction parts, the method of the convolutional layer in the existing degree of depth learning art of direct employing and down-sampling layer error back propagation undated parameter, upgrades parameter in convolutional layer and the down-sampling layer { value of CS} ({ CS} represents the set of matrices of all parameters in convolutional layer and down-sampling layer).In the error back propagation process of low-level feature abstract part, the method of the error back propagation undated parameter in the existing artificial neural network technology of direct employing, upgrades parameter in the L layer artificial neural network { value of LN} ({ LN} represents the set of matrices of all parameters in L layer artificial neural network).This completes the parameter training process of a picture group sheet, when then inputting the second picture group sheet training, the network parameter w determined with the first picture group sheet ₁, w ₂, the value of b and β, and { CS} and { the value calculating court verdict o of LN}, then repetitive error back-propagation process, w in renewal network ₁, w ₂, b and β value, and { CS} and the { value of LN}.By that analogy, the value of the parameter more than training of each picture group sheet determined after one picture group sheet training calculates court verdict, and then from output, the error of calculation backpropagation, to input, upgrade the value of all parameters of whole network.All input after network completes training until " T/B " organizes training sample, just complete and once train iteration.

After completing all E training iteration, namely network now can be used for actual age estimation, and input face picture, network will export the estimated value at age.

Above-mentioned formula (3), (4), (7), (8), (10), (12) are the new formula that the present invention proposes.

Three, carry out the face age based on above-mentioned structure network and estimate and sex identification:

After employing said method structure obtains sex recognition network and age estimation network, is carried out sex identification or age estimation in the network that picture to be identified input has been built, concrete identifying is:

The high-level characteristic of face picture to be identified is extracted based on convolutional neural networks;

Low layer and the global characteristics of face picture to be identified is extracted based on artificial neural network;

The low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be for training picture several times iteration after obtain (the final value of each parameter obtained when namely above-mentioned network struction completes), hfo represents the high-level characteristic vector of the face to be identified of extraction, lfo represents the low-level feature vector of the face to be identified of extraction, and o represents the court verdict at sex or age.

Low-level feature abstract comprises further:

The face picture to be identified of input is carried out flaky process and after normalization, obtained the original feature vector of face;

Original feature vector is inputted artificial neural network, by multilayer neuronal structure, input vector is rebuild, obtain a L _fthe proper vector of dimension is as the low-level feature vector extracted.

Concrete judging process is:

Merge the high-level characteristic and low-level feature that extract, export a kind of multiple features structure; Adopt some output neurons to carry out sex judgement or age judgement, each described neuron is adjudicated based on described judgement formula.

The present invention gives a kind of system example specifically implemented based on the human face analysis method of multiple features degree of depth study, as shown in Figure 3." Face datection and picture pre-service " wherein in Fig. 3 directly adopts prior art means to realize, and Face datection is based on the mark spot check test and appraisal point method realization verified in conjunction with face regional area; Picture pre-service comprises the gray processing process of colour picture, picture size adjustment and the process of picture histogram equalization.After pre-service, system on human face picture adopts following strategy to identify:

High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks;

Low-level feature abstract module, for extracting low layer and the global characteristics of face picture to be identified based on artificial neural network;

Judging module, for the low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be for training picture several times iteration after obtain (the final value of each parameter obtained when namely above-mentioned network struction completes), hfo represents the high-level characteristic vector of the face to be identified of extraction, lfo represents the low-level feature vector of the face to be identified of extraction, and o represents the court verdict at sex or age.

Above-mentioned low-level feature abstract module comprises further:

Flaky process module, for carrying out flaky process by the face picture to be identified of input and after normalization, obtain the original feature vector of face;

Reconstruction features vector acquisition module, for original feature vector is inputted artificial neural network, is rebuild input vector by multilayer neuronal structure, obtains a L _fthe proper vector of dimension is as the low-level feature vector extracted.

Above-mentioned judging module comprises further:

Combining voting layer module, for merging the high-level characteristic and low-level feature that extract, exporting a kind of multiple features structure; Final sex or the judgement at age is carried out to output layer.

Output layer module, for adopting some output neurons to carry out sex judgement or age judgement, each described neuron is adjudicated based on described judgement formula.

In a word, the invention provides a kind of human face analysis method and system based on the study of the multiple features degree of depth.This system first carries out Face datection to picture and pre-service obtains face picture, then analyzes face picture.Human face analysis specifically comprises gender classification and the face age is estimated.System the network having trained two deep layers is respectively used to gender classification and the face age is estimated, using the input of face picture as network, through the extraction of high-level characteristic, cascading judgement after the extraction of low-level feature and fusion feature, two deep layer networks finally export gender prediction's value and the age estimated value of face picture respectively.Native system effectively raises the accuracy rate of Gender Classification and reduces the mean absolute error that the age estimates, simultaneity factor has stronger generalization ability, can show good performance on the picture of low-quality network picture and camera actual acquisition.

The variable representing matrix of the black matrix in technique scheme.

It should be noted last that, above embodiment is only in order to illustrate technical scheme of the present invention and unrestricted.Although with reference to embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that, modify to technical scheme of the present invention or equivalent replacement, do not depart from the spirit and scope of technical solution of the present invention, it all should be encompassed in the middle of right of the present invention.

Claims (10)

1., for the deep layer network establishing method that gender classification or age are estimated, described method comprises:

Step 101) all training pictures are divided into some groups;

Step 102) extract based on convolutional neural networks the high-level characteristic that a group is trained picture, and then obtain by vectorial the first matrix formed of high-level characteristic; Extract low layer and the global characteristics of same group of training picture simultaneously based on artificial neural network, and then obtain the second matrix of low-level feature vector composition;

The result of one group of sex identification or age estimation is obtained based on the first matrix extracted, the second matrix and following judgement formula:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, hfo represents the first matrix; Lfo represents the second matrix; For the first weight matrix w in first group of above-mentioned formula of training picture ₁, the second weight matrix w ₂, bias matrix b and regulate the initial value of weight beta to adopt random initializtion mode to obtain; For the w in all the other above-mentioned formula of training picture respectively organized of input ₁, w ₂, b and β the acquisition methods of value be: utilize error backpropagation algorithm to calculate error function J (W, b that court verdict o and each group train the physical tags matrix Y of picture; β), then by calculate w ₁, w ₂, b and β be to error function J (W, b; Gradient β) and then undated parameter w ₁, w ₂, b and β value;

Step 103) input again one group training picture, and above-mentioned steps 102 is repeated to the training picture again inputted), until all groupings all carry out steps 102) process, complete and once train iteration;

Step 104) all training pictures are reclassified as some groups, and repeat above-mentioned steps 102 to each group that repartitions) and step 103), complete iteration again;

Through some grouping and iterative processing again, until when the final judgement o exported meets the condition of setting, obtain final parameter w ₁, w ₂, b and β value, complete network struction.

2. the deep layer network establishing method estimated for gender classification or age according to claim 1, is characterized in that, comprise further when carrying out low-level feature abstract:

Step 102-11) each Zhang Xunlian picture of one group of input training picture is converted into vectorial form by the graph structure form of two dimension, then vector is normalized, obtain the original feature vector of each Zhang Xunlian picture;

Step 102-12) original vector of each the Zhang Xunlian picture obtained is inputted artificial neural network, and then obtain one group of reconstruction features vector, namely obtain the second described matrix; Wherein, described artificial neural network comprises L layer, and adopts full connected mode between layers, and each neuron of every one deck adopts sigmoid function to activate.

3. the deep layer network establishing method estimated for gender classification or age according to claim 1, is characterized in that, trains the sex of picture or age identifying specifically to comprise for an input:

Step 102-21) when extract one training picture high-level characteristic vector be H _fthe high-level characteristic vector of dimension, and low-level feature vector is L _fduring the proper vector tieed up, construct one and comprise " H _f+ L _f" individually neuronicly combine voting layer;

Step 102-22) when for sex identification, each neuron of combining voting layer of structure is connected with two output neurons of output layer respectively, and each output neuron carries out Sex Discrimination based on described judgement formula, exporting training picture is the probability of sex; When estimating for the age, combine voting each neuron of layer and be connected with S output neuron of output layer, wherein each output neuron correspondence is one-year-old.

4. the deep layer network establishing method estimated for gender classification or age according to claim 1, is characterized in that, adopt the error backpropagation algorithm of following formula to upgrade the first weight matrix w ₁value:

$\left\{\begin{array}{c}{\left(w_1\right)}_{\mathrm{new}}={\left(w_1\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_1\\ {\mathrm{\Δw}}_1=\mathrm{Od}*\mathrm{hfo}\end{array}\right.$

Wherein, (w ₁) _newrepresent the first weight matrix w after upgrading when error back propagation each time ₁value, (w ₁) _oldthe first weight matrix w before corresponding renewal ₁value, Od represents output layer sensitivity matrix; α represents the learning rate of network;

The second weight matrix w is upgraded by following formula ₂value:

$\left\{\begin{array}{c}{\left(w_2\right)}_{\mathrm{new}}={\left(w_2\right)}_{\mathrm{old}}-\mathrm{\α}*{\mathrm{\Δw}}_2\\ {\mathrm{\Δw}}_2=\mathrm{Od}*\mathrm{lfo}\end{array}\right.$

Wherein, (w ₂) _newrepresent the second weight matrix w after upgrading when error back propagation each time ₂value, (w ₂) _oldthe second weight matrix w before corresponding renewal ₂value.

5. the deep layer network establishing method estimated for gender classification or age according to claim 1, is characterized in that, the value of described β training update method in iteration is each time:

$\left\{\begin{array}{c}{\mathrm{\β}}_{\mathrm{new}}={\mathrm{\β}}_{\mathrm{old}}-\mathrm{\α}\·\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}\\ 0\≤{\mathrm{\β}}_{\mathrm{new}},{\mathrm{\β}}_{\mathrm{old}}\≤1\end{array}\right.$

Wherein, β _newrepresent the value of the adjustment weight beta after upgrading when error back propagation each time, β _oldthe value of the adjustment weight beta before corresponding renewal;

Ask the part of local derviation can be obtained by following formula in above-mentioned formula:

$\left\{\begin{array}{c}\frac{\∂J(W,b;\mathrm{\β})}{\∂\mathrm{\β}}=\mathrm{mean}\left(B(:)\right)\\ B=(w_2*\mathrm{lfo})\·*f^{\text{'}}\left(o\right)\\ f^{\text{'}}\left(o\right)=o\·*(1-o)\end{array}\right.$

Wherein, f ' (o) represents court verdict o differentiate, and " mean (B (:)) " represents and to average computing to all elements in matrix B; Matrix B represents the matrix be made up of the value of the adjustment weight beta after upgrading when error back propagation each time, and the ranks number of this matrix B is identical with the ranks number of court verdict o.

6. for the age of face or a recognition methods for sex, the first weight matrix w that the structure network that the method is recorded based on a claim any in claim 1-5 is finally determined ₁, the second weight matrix w ₂, β and b value, described method comprises:

For extracting the high-level characteristic of face picture to be identified based on convolutional neural networks;

For extracting low layer and the global characteristics of face picture to be identified based on artificial neural network;

For the low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be the value that deep layer network establishing method is determined, hfo represents the high-level characteristic vector of the face to be identified of extraction, and lfo represents the low-level feature vector of the face to be identified of extraction, and o represents the court verdict at sex or age.

7. according to claim 6ly to it is characterized in that for the age of face or the recognition methods of sex, adopt following steps to extract low-level feature:

For the face picture to be identified of input is carried out flaky process and after normalization, is obtained the original feature vector of face;

For original feature vector is inputted artificial neural network, by multilayer neuronal structure, input vector is rebuild, obtain a L _fthe proper vector of dimension is as the low-level feature vector extracted.

8. for the age of face or a recognition system for sex, it is characterized in that, described system comprises:

High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks;

Low-level feature abstract module, for extracting low layer and the global characteristics of face picture to be identified based on artificial neural network;

Based on the judging module of neural network, for the low-level feature of extraction and high-level characteristic are inputted following judgement formula, carry out sex or age judgement, export court verdict:

o＝sigm(w ₁*hfo+β×w ₂*lfo+b)

Wherein, the first weight matrix w in above-mentioned formula ₁, the second weight matrix w ₂, β and b be for training picture several times iteration after obtain, hfo represent the face to be identified of extraction high-level characteristic vector, lfo represent the face to be identified of extraction low-level feature vector, o represents the court verdict at sex or age.

9. the age of face according to claim 8 or the recognition system of sex, is characterized in that, described low-level feature abstract module comprises further:

Flaky process module, for carrying out flaky process by the face picture to be identified of input and after normalization, obtain the original feature vector of face;

Reconstruction features vector acquisition module, for original feature vector is inputted artificial neural network, is rebuild input vector by multilayer neuronal structure, obtains a L _fthe proper vector of dimension is as the low-level feature vector extracted.

10. the age of face according to claim 8 or the recognition system of sex, is characterized in that, described judging module comprises further:

Combining voting layer module, for merging the high-level characteristic and low-level feature that extract, exporting a kind of multiple features structure;

Output layer module, for adopting some output neurons to carry out sex judgement or age judgement, each described neuron is adjudicated based on described judgement formula.