CN105894050A - Multi-task learning based method for recognizing race and gender through human face image - Google Patents

Abstract

The invention provides a multi-task learning based method for recognizing race and gender through a human face image, relates to the technical fields such as the digital image processing field, the mode recognizing field, the computer vision field and the physiology field, and aims to solve the problems of recognizing of the race and the gender from a static human face or a video human face under a plurality of occasions. The multi-task learning method is a learning method for improving the learning performance through related task learning and has the advantages of recognizing the learning difference between the tasks, sharing related features of the tasks, improving the learning performance through the relevance, and reducing the high-dimension small sample over-learning problem. The multi-task learning method can be applied to the human face based race and gender recognizing; different semantics are treated as different tasks, and on that basis, the semantics-based multi-task feature selection is proposed. With the adoption of the method for recognizing the race and the gender, the generalization capacity of a learning system and the recognizing effect can be obviously improved.

Description

A kind of other recognition methods of facial image race and sex based on multi-task learning

Technical field

The present invention is a kind of other recognition methods of facial image race and sex based on multi-task learning, and the method relates to And the technical field such as Digital Image Processing, pattern recognition, computer vision, physiology.

Background technology

Facial image contains abundant information, from the angle of pattern recognition, can carry out race's identification, property Not Shi Bie, identification etc..

PCA (Principle Component Analysis, PCA), its basic thought is to pass through K-L The principal character of sample is extracted in conversion, is launched by the characteristic vector solving training sample covariance matrix Base, is represented the significance level of main constituent by eigenvalue descending sort.Kirby(Turk,M.,Pentland,A., Eigenface for Recognition[J].Journal of cognitive Neuroscience.Vol.3,No.1,1991, Etc. pp.1-17) in nineteen ninety PCA is used for solving recognition of face problem, Turk (Kirby, M., Sirovich, L.,Application of the Karhunen-Loeve procedure for the characterization of human faces[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,1990,Vol. 12, No.1, pp.103-108) etc. it is developed into eigenface (Eigenfaces), for front face identification, The attention of researcher is obtained from the method for this face subspace.

Support vector machine [42] (Support Vector Machine, SVM) (Cortes, V.V., Support vector Networks [J] .Machine Learning, Vol.20,1995, pp.273～297) it is by Cortes and Vapnik Propose to be used for solving handwriting recognition problem in nineteen ninety-five.It can be according to limited sample information at model Seek optimal trading off between complexity and learning capacity, be desirably to obtain best Generalization Ability.

The terminal decision function of SVM is only determined by the support vector of minority, and the complexity of calculating depends on propping up Hold vector number rather than the dimension of sample space, this not only can help we grasp the key link sample, Reject bulk redundancy sample, and method be simple, has preferable robustness, is mainly reflected in:

1. increase, delete non-supporting vector sample model is not affected；

2. support that vector sample set has certain robustness；

3. core is chosen relative insensitivity.

SVM is substantially binary classifier.But in the face of multi-class problem, one-to-many can be constructed, one to one and The methods such as SVM decision tree.SVM decision tree (SVM Decision Tree) is by SVM and binary decision tree Combine, constitute multistage classifier.

Multi-task learning method is a kind of learning method learning to improve learning performance by inter-related task, both The diversity of study between task can be distinguished, it is also possible to share correlated characteristic between task, carried by dependency High learning capacity, alleviates higher-dimension small sample and crosses problem concerning study.

Summary of the invention

The introducing that the abundant semanteme that facial image comprises is multi-task learning provides possible, and we are by different languages Justice is as different tasks, it is proposed that based on semantic multitask feature selecting algorithm, for given face Training set of images, is obtained after extracting feature by the data set of task ranking, tried to achieve optimum by iteration c Sparse coefficient W of task, the sparse coefficient of every string correspondence individual task of W, sparse value absolute value is the biggest, Show that this dimensional feature is the biggest to the contribution rate of place task recognition, can realize every kind after sorting by contribution margin The feature of task selects.

This method mainly comprises training part and part of detecting.Training part is to be obtained by training set and learning strategy To sex and the feature selection label of race, test process selects different task class marks, categorized device After be identified result, obtain the recognition accuracy of multitask feature selection through repetitive measurement.

A kind of other recognition methods of facial image race and sex based on multi-task learning, the method comprises the steps:

S1 picture pretreatment

Picture pretreatment is a link important in recognition of face flow process, and its main purpose is to weaken even to disappear Except in face picture to identifying unrelated information, filter noise jamming, strengthen the proportion of useful information, it is simple to Later stage extracts significantly more efficient feature and is used for classifying.The method of the picture pretreatment that the present invention uses specifically includes that

(1) histogram equalization: the method by known picture through certain convert so that it is grey level probability density in All hook distribution, weaken the impact of the even change of uneven illumination with this, thus strengthen image overall contrast.Nogata Figure equalization divides three steps:

A) rectangular histogram of original image is added up；

B) done conversion by the grey level histogram of original image according to score accumulation function and generate new grey level histogram；

S_k=T (r_k)=∑ P_r(r)

Wherein, S_kRepresent the gradation of image after conversion；R represents original image gray scale；K represents that gray level is numbered； T(r_k) represent transforming function transformation function；P_rR () represents the probability density function of image gray levels.

C) replace former rectangular histogram with new grey level histogram, now can make the probability phase that in picture, each pixel value occurs Closely.

(2) gray scale stretching: gray scale stretching is one of most basic gray-scale transformation method, by using piecewise linearity Arbitrary scope of original image gray scale is changed by function so that it is change to the interval specified, thus can To increase the image contrast in a certain interval.Gray scale stretching completes in two steps:

A) calculate rectangular histogram on the original image, determine that gray scale is drawn according to the histogram distribution situation of original image The flex point stretched.

B) use the piecewise linear function determined to be mapped to former grey scale pixel value specify numerical value, replace preimage with this Element value.

By the method, the appointment scope of the pixel value of face picture is mapped, thus strengthen gray value office Portion's contrast.

(3) image normalization: facial image normalization comprises two aspects:

A) face picture zoomed in and out or rotate and become in the same size, human face characteristic point position substantially phase With a collection of picture, the step for be referred to as the geometrical normalization of face picture.This is owing to face picture is being adopted During collection, as different in face shooting distance or the shooting of face that the difference of some physical locations causes Angle difference.

B) eliminate the picture pixels value brought in face picture gatherer process due to illumination variation etc. to differ greatly Impact.To this class difference, we can carry out the gray value normalization operation of image, weakens because of illumination with this Or the difference of the gray value that shooting angle difference causes.Can be at very great Cheng by image is normalized Improving the performance of face identification system on degree, improve recognizer discrimination, therefore tool has very important significance.

S2 multitask labelling

(1) select M width face picture as training sample set from training sample database, and carry out pretreatment.

(2) for c given relational learning task, training set is labeled as: { (X₁,y₁),…,(X_c,y_c), Wherein:

Represent the training sample of i-th task；

Represent the class mark of i-th task；

N: represent the number of i-th task training sample；

D: represent the dimension of training sample.

By the common study of c task, it is desirably to obtain weight matrix: W=[w₁,…,w_c]∈R^d*c,Weights coefficient for i-th task.

S3 training pattern

Assume J_i(w_i,X_i,y_i)=| w_iX_i-y_i| for the loss function of i-th task, general loss function has: Log-likelihood function, index likelihood function and Hinge function.

For i-th task, while minimizing experience error, normal form is used to obtain optimization problem as follows:

$\underset{w_i}{min}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\left|\right|w_i|{|}_1$

From the point of view of individual task, the problems referred to above are commonly called LASSO (least absolute shrinkage and Selection operator), i.e. rarefaction representation classification (sparse representation for classification, SRC) Method, LASSO is a convex optimization problem, no longer has analytic solutions, but can make w in solution procedure_iIn many Item levels off to 0, has openness, is l₀Regularization well approximates (l₀It is np hard problem).

The optimal solution individually solving c task is consistent with the global objective function solving Joint Task, and statement is such as Under:

$\underset{w}{min}\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{i=1}{\overset{c}{\Σ}}\left|\right|w_i|{|}_1$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible, at l₁Under the constraint of norm, it is separate between each task, in order to global characteristics is entered Row feature selection, after making following change, obtains l₂Object function expression-form under norm constraint:

$\underset{w}{\min }\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{k=1}{\overset{d}{\Σ}}\left|\right|w^k|{|}_2$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible l directly perceived₂The feature between each task is have shared under norm.From the point of view of whole W,First Calculate w^k2 norms, then sue for peace, referred to as l₁/l₂Norm.

Sparse coefficient W of c task of optimum is tried to achieve by iteration, every string correspondence individual task of W Sparse coefficient, sparse value absolute value is the biggest, shows that this dimensional feature is the biggest to the contribution rate of place task recognition, Can realize the feature of every kind of task is selected after sorting by contribution margin, finally give optimal models.

S4 model measurement

(1) select N width face picture as test sample collection from test sample storehouse, and carry out pretreatment.

(2) test picture is sequentially inputted in the model trained, two tasks is solved simultaneously.

(3) category score solved is ranked up, takes maximum as final prediction classification.

(4) after drawing the other classification of race and sex of face picture, can be according to index from corresponding class labelling literary composition Part reads retrtieval carry out exporting explanation.

The most popular machine learning algorithm theory is all once only one task of study on unified model, Challenge is first resolved into the most independent subproblem, the sample in each subproblem, in training set Only reflect the information of individual task.Facial image contains the information such as race, sex, identity, corresponding different There is dependency between the identification mission of information, learning process is shared between each task certain relevant letter Breath.During multi-task learning method introducing facial image race and sex is not identified, using different semantic as different Task, proposes, based on semantic multitask feature selection, to be applied to race and sex and do not identify, can significantly improve The generalization ability of learning system and recognition effect.

Accompanying drawing explanation

Accompanying drawing herein is merged in description and constitutes the part of this specification, it is shown that meet the present invention Embodiment, and for explaining the principle of the present invention together with description.

Fig. 1 is the model training schematic flow sheet about this inventive method；

Fig. 2 is the model measurement schematic flow sheet about this inventive method；

Fig. 3 is the training sample pictures original according to a group shown in an exemplary embodiment；

Fig. 4 is according to the samples pictures collection after pretreatment of a group shown in an exemplary embodiment；

Fig. 5 is according to one group of test result schematic diagram shown in an exemplary embodiment；

Detailed description of the invention

Here will illustrate exemplary embodiment in detail, its example represents in the accompanying drawings.

When explained below relates to accompanying drawing, unless otherwise indicated, the same numbers in different accompanying drawings represents identical Or similar key element.

Embodiment described in following exemplary embodiment does not represent all realities consistent with the present invention Execute mode.On the contrary, they only with describe in detail in appended claims, the present invention some in terms of The example of consistent method.

The term used in the present invention is only merely for describing the purpose of specific embodiment, and is not intended to be limiting The present invention.

" a kind of ", " described " and " being somebody's turn to do " of singulative used in the present invention and appended claims It is also intended to include most form, unless context clearly shows that other implications.

The term that uses in the present invention " with ", "or" refers to comprise one or more project of listing being associated Any or all possible combination.

Term " first ", " second ", " the 3rd " etc. may be used to describe various information in the present invention, but these Information is not limited only to these terms.These terms are only used for same type of information is distinguished from each other out.Such as, Without departing from the present invention, the first information can also be referred to as the second information, similarly, Two information can also be referred to as the first information.Depend on linguistic context, word as used in this " if " can be by Explanation become " ... time " or " when ... time " or " in response to ... ".

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described in further detail:

A kind of other recognition methods of facial image race and sex based on multi-task learning, the method comprises the steps:

S1 picture pretreatment

Picture pretreatment is a link important in recognition of face flow process, and its main purpose is to weaken even to disappear Except in face picture to identifying unrelated information, filter noise jamming, strengthen the proportion of useful information, it is simple to Later stage extracts significantly more efficient feature and is used for classifying.The method of the picture pretreatment that the present invention uses specifically includes that

(1) histogram equalization: the method by known picture through certain convert so that it is grey level probability density in All hook distribution, weaken the impact of the even change of uneven illumination with this, thus strengthen image overall contrast.Nogata Figure equalization divides three steps:

A) rectangular histogram of original image is added up；

B) done conversion by the grey level histogram of original image according to score accumulation function and generate new grey level histogram；

S_k=T (r_k)=∑ P_r(r)

Wherein, S_kRepresent the gradation of image after conversion；R represents original image gray scale；K represents that gray level is numbered； T(r_k) represent transforming function transformation function；P_rR () represents the probability density function of image gray levels.

C) replace former rectangular histogram with new grey level histogram, now can make the probability phase that in picture, each pixel value occurs Closely.

(2) gray scale stretching: gray scale stretching is one of most basic gray-scale transformation method, by using piecewise linearity Arbitrary scope of original image gray scale is changed by function so that it is change to the interval specified, thus can To increase the image contrast in a certain interval.Gray scale stretching completes in two steps:

A) calculate rectangular histogram on the original image, determine that gray scale is drawn according to the histogram distribution situation of original image The flex point stretched.

B) use the piecewise linear function determined to be mapped to former grey scale pixel value specify numerical value, replace preimage with this Element value.

By the method, the appointment scope of the pixel value of face picture is mapped, thus strengthen gray value office Portion's contrast.

(3) image normalization: facial image normalization comprises two aspects:

A) face picture zoomed in and out or rotate and become in the same size, human face characteristic point position substantially phase With a collection of picture, the step for be referred to as the geometrical normalization of face picture.This is owing to face picture is being adopted During collection, as different in face shooting distance or the shooting of face that the difference of some physical locations causes Angle difference.

B) eliminate the picture pixels value brought in face picture gatherer process due to illumination variation etc. to differ greatly Impact.To this class difference, we can carry out the gray value normalization operation of image, weakens because of illumination with this Or the difference of the gray value that shooting angle difference causes.Can be at very great Cheng by image is normalized Improving the performance of face identification system on degree, improve recognizer discrimination, therefore tool has very important significance.

S2 multitask labelling

(1) select M width face picture as training sample set from training sample database, and carry out pretreatment.

(2) for c given relational learning task, training set is labeled as: { (X₁,y₁),…,(X_c,y_c),

Wherein:

Represent the training sample of i-th task；

Represent the class mark of i-th task；

N: represent the number of i-th task training sample；

D: represent the dimension of training sample.

By the common study of c task, it is desirably to obtain weight matrix: W=[w₁,…,w_c]∈R^d*c,Weights coefficient for i-th task.

S3 training pattern

Assume J_i(w_i,X_i,y_i)=| w_iX_i-y_i| for the loss function of i-th task, general loss function has: Log-likelihood function, index likelihood function and Hinge function.

For i-th task, while minimizing experience error, normal form is used to obtain optimization problem as follows:

$\underset{w_i}{\min }{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\left|\right|w_i|{|}_1$

From the point of view of individual task, the problems referred to above are commonly called LASSO (least absolute shrinkage and Selection operator), i.e. rarefaction representation classification (sparse representation for classification, SRC) Method, LASSO is a convex optimization problem, no longer has analytic solutions, but can make w in solution procedure_iIn many Item levels off to 0, has openness, is l₀Regularization well approximates (l₀It is np hard problem).

The optimal solution individually solving c task is consistent with the global objective function solving Joint Task, and statement is such as Under:

$\underset{w}{\min }\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{i=1}{\overset{c}{\Σ}}\left|\right|w_i|{|}_1$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible, at l₁Under the constraint of norm, it is separate between each task, in order to global characteristics is entered Row feature selection, after making following change, obtains l₂Object function expression-form under norm constraint:

$\underset{w}{\min }\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{k=1}{\overset{d}{\Σ}}\left|\right|w^k|{|}_2$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible l directly perceived₂The feature between each task is have shared under norm.From the point of view of whole W,First Calculate w^k2 norms, then sue for peace, referred to as l₁/l₂Norm.

Sparse coefficient W of c task of optimum is tried to achieve by iteration, every string correspondence individual task of W Sparse coefficient, sparse value absolute value is the biggest, shows that this dimensional feature is the biggest to the contribution rate of place task recognition, Can realize the feature of every kind of task is selected after sorting by contribution margin, finally give optimal models.

S4 model measurement

(1) select N width face picture as test sample collection from test sample storehouse, and carry out pretreatment.

(2) test picture is sequentially inputted in the model trained, two tasks is solved simultaneously.

(3) category score solved is ranked up, takes maximum as final prediction classification.

(4) after drawing the other classification of race and sex of face picture, can be according to index from corresponding class labelling literary composition Part reads retrtieval carry out exporting explanation.

Claims (1)

1. the other recognition methods of facial image race and sex based on multi-task learning, it is characterised in that: should Method comprises the steps:

S1 picture pretreatment

Picture pretreatment is a link important in recognition of face flow process, and its main purpose is to weaken even to disappear Except in face picture to identifying unrelated information, filter noise jamming, strengthen the proportion of useful information, it is simple to Later stage extracts significantly more efficient feature and is used for classifying；The method of the picture pretreatment that the present invention uses specifically includes that

(1) histogram equalization: the method by known picture through certain convert so that it is grey level probability density in All hook distribution, weaken the impact of the even change of uneven illumination with this, thus strengthen image overall contrast；Nogata Figure equalization divides three steps:

A) rectangular histogram of original image is added up；

B) done conversion by the grey level histogram of original image according to score accumulation function and generate new grey level histogram；

S_k=T (r_k)=∑ P_r(r)

Wherein, S_kRepresent the gradation of image after conversion；R represents original image gray scale；K represents that gray level is numbered； T(r_k) represent transforming function transformation function；P_rR () represents the probability density function of image gray levels；

C) replace former rectangular histogram with new grey level histogram, now can make the probability phase that in picture, each pixel value occurs Closely；

(2) gray scale stretching: gray scale stretching is one of most basic gray-scale transformation method, by using piecewise linearity Arbitrary scope of original image gray scale is changed by function so that it is change to the interval specified, thus can To increase the image contrast in a certain interval；Gray scale stretching completes in two steps:

A) calculate rectangular histogram on the original image, determine that gray scale is drawn according to the histogram distribution situation of original image The flex point stretched；

B) use the piecewise linear function determined to be mapped to former grey scale pixel value specify numerical value, replace preimage with this Element value；

By the method, the appointment scope of the pixel value of face picture is mapped, thus strengthen gray value office Portion's contrast；

(3) image normalization: facial image normalization comprises two aspects:

A) face picture zoomed in and out or rotate and become in the same size, human face characteristic point position substantially phase With a collection of picture, the step for be referred to as the geometrical normalization of face picture；This is owing to face picture is being adopted During collection, as different in face shooting distance or the shooting of face that the difference of some physical locations causes Angle difference；

B) eliminate the picture pixels value brought in face picture gatherer process due to illumination variation etc. to differ greatly Impact；To this class difference, we can carry out the gray value normalization operation of image, weakens because of illumination with this Or the difference of the gray value that shooting angle difference causes；Can be at very great Cheng by image is normalized Improving the performance of face identification system on degree, improve recognizer discrimination, therefore tool has very important significance；

S2 multitask labelling

(1) select M width face picture as training sample set from training sample database, and carry out pretreatment；

(2) for c given relational learning task, training set is labeled as: { (X₁,y₁),…,(X_c,y_c),

Wherein:

Represent the training sample of i-th task；

Represent the class mark of i-th task；

N: represent the number of i-th task training sample；

D: represent the dimension of training sample；

By the common study of c task, it is desirably to obtain weight matrix: W=[w₁,…,w_c]∈R^d*c,Weights coefficient for i-th task；

S3 training pattern

Assume J_i(w_i,X_i,y_i)=| w_iX_i-y_i| for the loss function of i-th task, general loss function has: Log-likelihood function, index likelihood function and Hinge function；

For i-th task, while minimizing experience error, normal form is used to obtain optimization problem as follows:

$\underset{w_i}{min}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\left|\right|w_i|{|}_1$

From the point of view of individual task, the problems referred to above are commonly called LASSO (least absolute shrinkage and Selection operator), i.e. rarefaction representation classification (sparse representation for classification, SRC) Method, LASSO is a convex optimization problem, no longer has analytic solutions, but can make w in solution procedure_iIn many Item levels off to 0, has openness, is l₀Regularization well approximates (l₀It is np hard problem)；

The optimal solution individually solving c task is consistent with the global objective function solving Joint Task, and statement is such as Under:

$\underset{W}{min}\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{i=1}{\overset{c}{\Σ}}\left|\right|w_i|{|}_1$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible, at l₁Under the constraint of norm, it is separate between each task, in order to global characteristics is entered Row feature selection, after making following change, obtains l₂Object function expression-form under norm constraint:

$\underset{W}{\min }\underset{i=1}{\overset{c}{\Σ}}{J}_i(w_i,X_i,y_i)+\mathrm{\λ}\underset{k=1}{\overset{d}{\Σ}}\left|\right|w^k|{|}_2$

$W=\left(\begin{array}{cccc}{w}_1^1& w_2^1& ...& w_c^1\\ w_1^2& w_2^2& ...& w_c^2\\ ...& ...& ...& ...\\ w_1^d& w_2^d& ...& w_c^d\end{array}\right)$

Visible l directly perceived₂The feature between each task is have shared under norm；From the point of view of whole W,First Calculate w^k2 norms, then sue for peace, referred to as l₁/l₂Norm；

Sparse coefficient W of c task of optimum is tried to achieve by iteration, every string correspondence individual task of W Sparse coefficient, sparse value absolute value is the biggest, shows that this dimensional feature is the biggest to the contribution rate of place task recognition, Can realize the feature of every kind of task is selected after sorting by contribution margin, finally give optimal models；

S4 model measurement

(1) select N width face picture as test sample collection from test sample storehouse, and carry out pretreatment；

(2) test picture is sequentially inputted in the model trained, two tasks is solved simultaneously；

(3) category score solved is ranked up, takes maximum as final prediction classification；

(4) after drawing the other classification of race and sex of face picture, can be according to index from corresponding class labelling literary composition Part reads retrtieval carry out exporting explanation.