CN105678253A - Semi-supervised age estimation device based on faces and semi-supervised age estimation method based on faces - Google Patents

Abstract

The invention discloses a semi-supervised age estimation method based on faces. A training method of the device comprises following steps: S1. obtaining face image data set and extracting image features; S2. performing age distribution initialization to each of age-labeled face images and using the images as a training set; S3. performing training to a LBFGS-LLD model by use of a current training set and performing age distribution prediction to all images; S4. calculating pseudo ages of face images without age labels; S5. grouping all the images by age, optimizing and solving variances corresponding to all age groups, and updating the age distributions of face images in corresponding age groups by use of the obtained variances; S6. using the updated images as a new training set and turning to S3 until a iteration termination condition being satisfied. The invention also discloses a semi-supervised age estimation method based on faces on the basis of the device. According to the device and method, only a few age-labeled face images are needed and in combination with more non-age-labeled face images, better age estimation precision is obtained.

Description

Semi-supervised face estimation of Age device and semi-supervised face age estimation method

Technical field

The present invention relates to a kind of face estimation of Age device, particularly relate to a kind of semi-supervised face estimation of Age device and semi-supervised face age estimation method, belong to machine learning and mode identification technology.

Background technology

Age is one of important attribute of people, and the behavior of people and preference are all different at different age brackets, and this shows that estimation of Age will have very important application prospect accurately. At present much age-related application, such as man-machine interaction, E-customer's relation management, safety management, surveillance monitor etc., is obtained for great development. And in the method for numerous estimation of Age, be perhaps one the most frequently used in daily life based on the estimation of Age of face.

One section of Chinese invention patent (CN102567719) discloses one " the human age automatic estimating method based on posterior probability neutral net ", utilize the posterior probability neutral net of training in advance as face estimation of Age model, its training stage use an age distribution (with vector representation) represent this face picture the probability distribution at likely age. Each element representation correspondence face picture real age in vector is the probability at this age. This distribution is based on an assumption that the appearance of people gradually changed with the age, and the speed of appearance change is different at each age bracket, therefore the pace of change of appearance can reflect this information of age, and the speed changed can be by being distributed and embodies. As it is shown in figure 1, people is teenager, as 0-20 year, and old, as 50-76 year, its appearance changes greatly, and corresponding age distribution is comparatively precipitous; And in the middle age, as 20-50 year, the change of its appearance is little, and corresponding age distribution is then comparatively mild. Therefore for each face picture, its age distribution should include all possible age bracket, 0-76 year in Fig. 1, the age that so we use in the training process of system is no longer an independent numerical value, but the distribution vector of all age brackets (note: in age distribution vector, all elements is all higher than being equal to 0, less than or equal to 1, and summation is 1).Meanwhile, this age distribution should be the highest in the probability at real age place, and secondly along with the gap at age Yu real age becomes larger, its probability tapers into, and therefore this distribution can be expressed as Gauss distribution. The average of this Gauss distribution is real age, and variance controls the appearance variation tendency of this age bracket.

The age of face picture can accurately be estimated by said method, but its face picture of using of training can only be the face picture of has age labelling. On the one hand, in real world, age distribution cannot get, it is possible to the has age obtained, and owing to the procurement cost at age is too big, major part face picture does not all have age indicator; On the other hand, in actual applications, obtaining the sufficient human face data with age indicator is an engineering taken time and effort, and along with the rise of the electronic equipment such as camera, smart mobile phone, obtains the substantial amounts of human face data without age indicator and become easily. Therefore, how to make full use of substantial amounts of unmarked human face data to carry out estimation of Age more accurately and just become most important.

Summary of the invention

The technical problem to be solved is in that to overcome prior art not enough, a kind of semi-supervised face estimation of Age device is provided, only need to use a small amount of has age labelling face picture to combine more without age indicator face picture, better estimation of Age precision can be obtained.

The present invention specifically solves above-mentioned technical problem by the following technical solutions:

A kind of semi-supervised face estimation of Age device, this device is trained by the following method and is obtained:

Step 1, acquisition face picture data set, and wherein each face picture is carried out image characteristics extraction; Described face picture data set includes the face picture of one group of has age labelling and one group of face picture without age indicator;

Step 2, each width face picture to has age labelling, be that it initializes the age distribution meeting Gauss distribution according to its labelling age, using these face picture as training set; The average of described age distribution is its labelling age, and variance is default initial variance;

Step 3, utilize current training set training indicia distribution learning model LBFGS-LLD, and with complete training LBFGS-LLD model all face picture in face picture data set are carried out age distribution prediction, obtain its prediction age distribution;

Step 4, calculate each width pseudo-age without the face picture of age indicator; The pseudo-age computational methods of arbitrary face picture are specific as follows: select face picture one group the highest with its comprehensive similarity for this face picture from the face picture of all has age labellings, then using the average at the labelling age of selected face picture as the pseudo-age of this face picture; The tolerance of described comprehensive similarity is the weighted sum of the characteristics of image similarity between two width face picture and age distribution similarity;

Step 5, all face picture in face picture data set were grouped according to its labelling age or pseudo-age, from each age group, then pick out the lineup's face picture less than default age deviation threshold of the deviation value between the average of its prediction age distribution and its labelling age or pseudo-age; To each age group, it is optimization aim to the maximum with comprehensive similarity sum between each face picture of picking out, the variance of one age distribution of Optimization Solution, the average of this age distribution is the age corresponding to this age group, and the variance obtained with Optimization Solution updates in this age group the age distribution corresponding to each face picture;

Step 6, using age distribution update after all face picture as new training set, go to step 3;Iterate, until meeting the stopping criterion for iteration preset.

A kind of semi-supervised face age estimation method based on above-mentioned semi-supervised face estimation of Age device, face picture to be estimated is carried out image characteristics extraction, and the characteristics of image extracted is inputted the LBFGS-LLD model that obtains of training stage, the average of the age distribution of LBFGS-LLD model output is the estimation age of this face picture to be estimated.

Compared to existing technology, the method have the advantages that

The present invention only needs to use a small amount of has age labelling picture, is utilized the face picture not having age indicator in a large number fully by age distribution; In learning process, the present invention utilizes has age labelling picture and age distribution thereof to estimate the pseudo-age without age indicator picture, and the continuous iteration of age distribution is updated. The present invention takes full advantage of the face picture without age indicator being easily obtained and estimation of Age model is trained, it is only necessary to uses a small amount of has age labelling face picture to combine more without age indicator face picture, can obtain better estimation of Age precision.

Accompanying drawing explanation

Fig. 1 is the age distribution example of face picture;

Fig. 2 is the schematic flow sheet of one specific embodiment of the present invention.

Detailed description of the invention

Below in conjunction with accompanying drawing, technical scheme is described in detail:

The semi-supervised face age estimation method of the present invention includes two stages: estimate training stage and the estimation of Age stage of model. Wherein, estimate that the training of model is as in figure 2 it is shown, specifically include following steps:

Step 1, acquisition face picture data set, and wherein each face picture is carried out image characteristics extraction; Described face picture data set includes the face picture of one group of has age labelling and one group of face picture without age indicator.

The present invention is for not only including the face picture of age indicator but also include the face picture without age indicator in the face picture data set of model training, thus the scale of face picture data set can significantly be expanded, make full use of the face picture without age indicator being easily obtained in a large number.

Described characteristics of image can adopt existing various facial image feature, for instance ActiveAppearanceModel (AAM), AgingpatternSubspace (AGES), AgeManifold, HOG, BIF feature etc. In the present embodiment, first extract the BIF feature of face picture, then utilize MFA algorithm that the BIF feature extracted is carried out dimensionality reduction. BIF is characterized by a kind of feature conventional in face age estimation method, its particular content is referred to document [G.Guo, G.Mu, Y.Fu, andT.S.Huang, " Humanageestimationusingbio-inspiredfeatures; " inProc.IEEEConf.ComputerVisionandPatternRecognition, Miami, FL, 2009, pp.112 119]. In order to reduce algorithm complex, improve algorithm real-time, carry out Feature Dimension Reduction further with MFA algorithm. MFA is a kind of conventional dimension reduction method, and its particular content is referred to document [S.Yan, D.Xu, B.Zhang, H.Zhang, Q.Yang, andS.Lin, " Graphembeddingandextensions:Ageneralframeworkfordimensio nalityreduction; " IEEETrans.PatternAnal.Mach.Intell., vol.29, no.1, pp.40 51,2007].

Step 2, each width face picture to has age labelling, be that it initializes the age distribution meeting Gauss distribution according to its labelling age, using these face picture as training set; The average of described age distribution is its labelling age, and variance is default initial variance.

Conventional face estimation of Age only has a unique age indicator for every face picture, and the present invention for every the labelling age face picture give an age distribution, this age distribution is Gauss distribution, its average is real age (the labelling age), and variance is default initial value σ⁰.Then using the has age labelling face picture that is distributed with initial age as training set. Wherein initial variance σ⁰Value can set flexibly, rule of thumb, be set to effect when 3 better. The circular of age distribution can use formula (1) to represent:

$d_{ij}^0=\frac{1}{{\mathrm{\σ}}^0\sqrt{2\mathrm{\π}}{Z}_i}\exp (-\frac{{(y_j-{\mathrm{\μ}}_i)}^2}{2{\left({\mathrm{\σ}}^0\right)}^2})---\left(1\right)$

Wherein σ⁰For initial variance, x_iIt is i-th markd face picture example, y_jRepresent jth age indicator, u_iBeing labelling age corresponding to i-th face picture, d is corresponding age distribution.

Step 3, utilize current training set training indicia distribution learning model LBFGS-LLD, and with complete training LBFGS-LLD model all face picture in face picture data set are carried out age distribution prediction, obtain its prediction age distribution.

The present invention utilizes indicia distribution learning model LBFGS-LLD to train estimation model. Assuming to be currently kth time iteration, its target is the KL divergence in the age distribution of minimum model prediction and training set between the age distribution of labelling, thus obtaining the parameter of optimum. Its optimization object function is:

$L\left({\mathrm{\θ}}^k\right)={\mathrm{\Σ}}_i{\mathrm{log\Σ}}_j\exp \left({\left({\mathrm{\θ}}^k\right)}^{T}g\right(x_i\left)\right)-{\mathrm{\Σ}}_i{\mathrm{\Σ}}_j{d}_{ij}^k{\left({\mathrm{\θ}}^k\right)}^{T}g\left(x_i\right)---\left(2\right)$

Wherein θ^kThe model parameter vector solved and (θ is needed for kth time iteration^k)^TFor its transposed vector, i is image index, and j is age indicator, x_iFor i-th facial image example; d_ijFor about x_iAge distribution value on jth age indicator, g (x_i) for representing x_iImage feature vector. More detailed content about LBFGS-LLD model is referred to document [X.Geng, C.Yin, andZ.-H.Zhou.FacialAgeEstimationbyLearningfromLabelDistr ibutions.IEEETransactionsonPatternAnalysisandMachineInte lligence (IEEETPAMI), 2013,35 (10): 2401-2412]. At the parameter θ obtaining optimum^kAfter, with the LBFGS-LLD model taking this optimized parameter, all face picture in face picture data set are carried out estimation of Age, it was predicted that the prediction age distribution that every face picture is corresponding.

Step 4, calculate each width pseudo-age without the face picture of age indicator; The pseudo-age computational methods of arbitrary face picture are specific as follows: select face picture one group the highest with its comprehensive similarity for this face picture from the face picture of all has age labellings, then using the average at the labelling age of selected face picture as the pseudo-age of this face picture; The tolerance of described comprehensive similarity is the weighted sum of the characteristics of image similarity between two width face picture and age distribution similarity.

Due to the age estimated without age indicator face picture is relatively unreliable in step 3, in order to obtain the more reliable estimation age, the present invention is further to the age recalculating its correspondence without age indicator face picture, and is called the pseudo-age. The computational methods at pseudo-age particularly as follows: select face picture one group the highest with its comprehensive similarity for this face picture from the face picture of all has age labellings, then using the average at the labelling age of selected face picture as the pseudo-age of this face picture. The present invention utilizes the weighted sum of the characteristics of image similarity between two width face picture and age distribution similarity to measure without the similarity between age indicator face picture and has age labelling face picture. Wherein, the characteristics of image similarity between two width face picture can adopt existing Euclidean distance, manhatton distance, correlation coefficient, comentropy isometry form; Age distribution similarity between two width face picture can adopt Euclidean distance, Sorensen distance, KL divergence, Jeffrey divergence isometry form, present invention preferably employs KL divergence (also known as relative entropy).Both weights can sets itself according to actual needs.

The present embodiment is utilize k nearest neighbor method the comprehensive similarity expression formula according to formula (3) to search out m from all has age labelling face picture and open the k nearest neighbor set N without age indicator face picture_m:

$\left|\right|x_m-x_n|{|}_2^2+{\mathrm{C\Σ}}_{j}p\left(y_j\right|x_m;{\mathrm{\θ}}^k)ln\frac{p\left(y_j\right|x_m;{\mathrm{\θ}}^k)}{p\left(y_j\right|x_n;{\mathrm{\θ}}^k)}---\left(3\right)$

Wherein x_m, x_nRespectively m, n open the image feature vector of face picture, and C is balance factor, y_jFor jth age, θ^kParameter vector during for kth time iteration, p (y_j|x_m; θ^k), p (y_j|x_n; θ^k) respectively by the model tried to achieve in step 3 m, n opened the age distribution of face picture prediction. Noting, now m pictures is the face picture without age indicator, and the face picture that the n-th pictures is has age labelling. Then m opens the pseudo-age through type (4) without age indicator face picture and determines:

${\mathrm{\μ}}_m^k=\frac{1}{K}{\mathrm{\Σ}}_{x_n\∈N_m}{\mathrm{\μ}}_n---\left(4\right)$

Wherein,For m being opened the pseudo-age that face picture is estimated, μ in kth time iteration (i.e. current iteration)_nIt is the authentic signature age of the n-th pictures, N_mThe m determined for through type (3) opens the K without age indicator face picture and opens neighbour's picture, and K is neighbour's number.

Step 5, all face picture in face picture data set were grouped according to its labelling age or pseudo-age, from each age group, then pick out the lineup's face picture less than default age deviation threshold of the deviation value between the average of its prediction age distribution and its labelling age or pseudo-age; To each age group, it is optimization aim to the maximum with comprehensive similarity sum between each face picture of picking out, the variance of one age distribution of Optimization Solution, the average of this age distribution is the age corresponding to this age group, and the variance obtained with Optimization Solution updates in this age group the age distribution corresponding to each face picture.

All face picture in face picture data set being grouped according to its labelling age or pseudo-age, each face picture with same tag age or pseudo-age is divided into same group; Then from each age group, pick out one group of picture that confidence level is higher. The concrete selection method adopted in the present embodiment by: select age (probability is the highest) of with good grounds forecast of distribution absolute value with real labelling age or the difference at pseudo-age less than mean absolute error (MeanAbsoluteError, be called for short MAE) face picture, wherein the computational methods of MAE are

$MAE=\frac{1}{l+u}{\mathrm{\Σ}}_i{e}_i---\left(5\right)$

The wherein quantity of the face picture of l, u respectively has age labelling and face picture without age indicator, e_iIt it is the absolute value of the age forecast error of i-th face picture.

After obtaining the picture that confidence level is higher, need to redefine the variance updating Gauss distribution corresponding to each age, its method is: to each age group, it is optimization aim to the maximum with comprehensive similarity sum between each face picture of picking out, the variance of one age distribution of Optimization Solution, the average of this age distribution is the age corresponding to this age group. Shown in its mathematical expression such as formula (6):

${\mathrm{\σ}}_{\mathrm{\μ}}^k={\mathrm{argmin}}_{{\mathrm{\σ}}_{\mathrm{\μ}}}{\mathrm{\Σ}}_{x_r\∈S_{\mathrm{\μ}}^k}{\mathrm{\Σ}}_j{d}_{x_r,y_j}ln\frac{d_{x_r,y_j}}{p\left(y_j\right|x_r;{\mathrm{\θ}}^k)}---\left(6\right)$

WhereinFor the picture set that the confidence level picked out in kth time iteration μ age group is higher, p (y_j|x_r; θ^k) it is to picture x by parameter current model_rThe age distribution of prediction,Computational methods be:

$d_{rj}=\frac{1}{{\mathrm{\σ}}_{\mathrm{\μ}}^k\sqrt{2\mathrm{\π}}{Z}_{\mathrm{\μ}}}\exp (-\frac{{(y_j-\mathrm{\μ})}^2}{2{\left({\mathrm{\σ}}_{\mathrm{\μ}}^k\right)}^2})---\left(7\right)$

Wherein Z_μNormalization item for distribution.

So, to each age group, all obtain a corresponding new Gauss distribution variance, update, by this variance, the age distribution that in this age group, each face picture is corresponding.

Step 6, using age distribution update after all face picture as new training set, go to step 3;Iterate, until meeting the stopping criterion for iteration preset.

In order to study is to better age distribution adaptively, algorithm needs successive ignition, therefore the variance to complete prediction age distribution update after all face picture as new training set, then go to step 3 and carry out next iteration; Until meeting the stopping criterion for iteration preset, exit. LBFGS-LLD model now is final face age forecast model. The stopping criterion for iteration model in the present embodiment forecast error on checking collection reaches minimum. To the method for estimation at checking collection upper face picture age it was wherein

$\mathrm{\μ}={\mathrm{argmax}}_{y_j}P\left(y_j\right|x;{\mathrm{\θ}}^k)---\left(8\right)$

Wherein x is the characteristics of image that face picture to be predicted is concentrated in checking, θ^kParameter vector during for kth time iteration.

When the face age forecast model that recycling trains carries out the prediction of face age, first face picture to be estimated is carried out image characteristics extraction, and the characteristics of image extracted is inputted the LBFGS-LLD model that obtains of training stage, LBFGS-LLD model can export an age distribution, and the average of this age distribution is the estimation age of face picture to be estimated.

In order to verify the effect of the inventive method, itself and existing several face age estimation methods having carried out contrast verification experiment, the data set used by experiment is MORPH data base, and it there are about 55000 face picture. Whole image datas are randomly divided into ten parts of equalization, take a copy of it every time and do test set by the method adopting ten times of cross validations, portion makees checking collection, remain eight parts and do training set, altogether training ten times, take the average result of the ten times evaluation criterion as the inventive method performance. Simultaneously in order to simulate semi-supervised application scenarios, the age indicator removal of divided data in the middle part of training set is represented and does not have markd data. Table 1 shows the inventive method and several existing method testing results.

Table 1 test result contrasts

This experiment adopts mean absolute error MAE as estimation of Age measure of effectiveness index. Wherein, KPLS is document [G.GuoandG.Mu.Simultaneousdimensionalityreductionandhuman ageestimationviakernelpartialleastsquaresregression.InCo mputerVisionandPatternRecognition (CVPR), 2011IEEEConferenceon, pages657 664.IEEE, 2011] method in; OHRank is document [K.-Y.Chang, C.-S.Chen, andY.-P.Hung.Ordinalhyperplanesrankerwithcostsensitiviti esforageestimation.InComputerVisionandPatternRecognition (CVPR), 2011IEEEConferenceon, pages585 592.IEEE, 2011] method in; LDL is document [X.Geng, C.Yin, andZ.-H.Zhou.Facialageestimationbylearningfromlabeldistr ibutions.PatternAnalysisandMachineIntelligence, IEEETransactionson, 35 (10): 2,401 2412,2013] method in; ALDL is document [X.Geng, Q.Wang, andY.Xia.FacialAgeEstimationbyAdaptiveLabelDistributionL earning.InPatternRecognition (ICPR), 201422ndInternationalConferenceon, pages4465 4470.IEEE, 2014] method in. The method being more than all traditional supervision face estimation of Age, in order to the advantage of the inventive method is better described, this experiment also compares traditional semi-supervised method LP, its particular content is referred to document [F.WangandC.Zhang.Labelpropagationthroughlinearneighborho ods.KnowledgeandDataEngineering, IEEETransactionson, 20 (1): 55 67,2008].OurMethod in table 1 represents the inventive method. All algorithms all carry out ten times of cross validations on MORPH data base, and when wherein verifying, markd face picture has 100 every time, it does not have the face picture of labelling has 49000, and result is the meansigma methods of ten the results.

The inventive method is compared existing method in the accuracy of estimation of Age and is greatly improved from the results shown in Table 1.

Claims (6)

1. a semi-supervised face estimation of Age device, it is characterised in that this device is trained by the following method and obtained:

Step 1, acquisition face picture data set, and wherein each face picture is carried out image characteristics extraction; Described face picture data set includes the face picture of one group of has age labelling and one group of face picture without age indicator;

Step 2, each width face picture to has age labelling, be that it initializes the age distribution meeting Gauss distribution according to its labelling age, using these face picture as training set; The average of described age distribution is its labelling age, and variance is default initial variance;

Step 3, utilize current training set training indicia distribution learning model LBFGS-LLD, and with complete training LBFGS-LLD model all face picture in face picture data set are carried out age distribution prediction, obtain its prediction age distribution;

Step 4, calculate each width pseudo-age without the face picture of age indicator; The pseudo-age computational methods of arbitrary face picture are specific as follows: select face picture one group the highest with its comprehensive similarity for this face picture from the face picture of all has age labellings, then using the average at the labelling age of selected face picture as the pseudo-age of this face picture; The tolerance of described comprehensive similarity is the weighted sum of the characteristics of image similarity between two width face picture and age distribution similarity;

Step 5, all face picture in face picture data set were grouped according to its labelling age or pseudo-age, from each age group, then pick out the lineup's face picture less than default age deviation threshold of the deviation value between the average of its prediction age distribution and its labelling age or pseudo-age; To each age group, it is optimization aim to the maximum with comprehensive similarity sum between each face picture of picking out, the variance of one age distribution of Optimization Solution, the average of this age distribution is the age corresponding to this age group, and the variance obtained with Optimization Solution updates in this age group the age distribution corresponding to each face picture;

Step 6, using age distribution update after all face picture as new training set, go to step 3; Iterate, until meeting the stopping criterion for iteration preset.

2. semi-supervised face estimation of Age device as claimed in claim 1, it is characterised in that the method for described image characteristics extraction is specific as follows: first extract the BIF feature of face picture, then utilizes MFA algorithm that the BIF feature extracted is carried out dimensionality reduction.

3. semi-supervised face estimation of Age device as claimed in claim 1, it is characterised in that the value of described initial variance is 3.

4. semi-supervised face estimation of Age device as claimed in claim 1, it is characterized in that, described age deviation threshold is determined in accordance with the following methods: to all face picture in this age group, calculate the deviation value between its prediction average of age distribution and its labelling age or pseudo-age respectively, then using the meansigma methods of described deviation value as the age deviation threshold corresponding to this age group.

5. semi-supervised face estimation of Age device as claimed in claim 1, it is characterised in that utilize KL divergence to measure the age distribution similarity between two width face picture.

6. based on the face age estimation method of face estimation of Age device semi-supervised described in any one of Claims 1 to 5, it is characterized in that, face picture to be estimated is carried out image characteristics extraction, and the characteristics of image extracted is inputted the LBFGS-LLD model that obtains of training stage, the average of the age distribution of LBFGS-LLD model output is the estimation age of this face picture to be estimated.