CN104598888A - Human face gender recognition method - Google Patents

Abstract

The invention relates to the technical field of face recognition, in particular to a human face gender recognition method. The method includes the following steps of acquiring a human face image; detecting whether the human face image comprises a human face; converting the human face image into a gray histogram about pixel points; calculating appropriate amount features; converting the features into an approach basis function kernel; calculating an object gender determination value of the human face image through a utility function. The human face gender recognition method enables a computer to have human face image object gender determination capability and achieves automated intelligent management.

Description

A kind of recognition methods of face gender

Technical field

The present invention relates to technical field of face recognition, particularly a kind of recognition methods of face gender.

Background technology

Along with the continuous progress of society and each side are for an urgent demand of auto authentication fast and effectively, biometrics identification technology obtains development at full speed in nearly decades.As a kind of inherent attribute of people, and have very strong self stability and individual difference, biological characteristic becomes the most desirable foundation of auto authentication.Current biometrics identification technology mainly includes: fingerprint recognition, retina identification, iris recognition, Gait Recognition, hand vein recognition, recognition of face etc.Compared with other recognition methodss, recognition of face owing to itself having directly, friendly, feature easily, user to it without any mental handicape, to be therefore easy to accept by user, thus obtain research and apply widely.In addition, can also be for further analysis to the result of recognition of face, obtain the many extra abundant informations such as sex, expression, age about people, extend the application prospect of recognition of face.Such as, and face is one of important biological characteristic, facial image has contained a large amount of information, sex, age ethnic group, identity etc.Wherein, the sex recognition function of face is exactly attempt to give the facial image of computing machine according to input, judges the ability of user's sex.The development of human-computer interaction technology (HCI) makes computer vision, artificial intelligence, plays an increasingly important role in monitoring, GUI Human Machine Interface etc.Along with the progress of technology, the problem based on facial image pattern-recognition becomes the focus of Recent study gradually.Comprising all kinds of identification problems such as Face datection, face identification, face character (sex, age, expression, race etc.) identifications.Gender Classification based on face is exactly the process making computing machine can judge its sex according to the facial image of input.The ability of the sex identification of face seemingly people " inherent ", but allow computing machine carry out identifying and being not easy, even if having in a large number from the effort of the researchist of the every field such as computer vision, pattern-recognition, artificial intelligence, psychology.

Chinese patent application CN 101344916 discloses a kind of face identification system, and it comprises video input interface, links together with face image data collecting unit, for receiving face image data; Recognition of face arithmetic processor, for processing the face image data received, completes identification work; Microprocessor unit, links together with described recognition of face arithmetic processor, communicates with various application apparatus for being responsible for.

Chinese patent application CN 102831408A discloses a kind of face identification method, and it comprises the following steps: S1 judges to detect the certain gestures that whether there is user's setting in video sequence; If so, S2 is carried out; Otherwise, repeat step S1; S2 detects in video area whether have user's face, if so, carries out step S3, otherwise repeats step S2; S3 carries out real-time follow-up to the user's face detected; S4 extracts illumination-insensitive feature to user's facial image; S5 analog subscriber postural change rule; S6 face recognition process, judge whether the sample in facial image to be identified and face database belongs to a people together, if so, identifying terminates, and system performs user operation; If not, step S7 is carried out; S7 inquires whether user is first time use system; If so, newly-built Sample Storehouse, if not, whether inquiry user substitutes Sample Storehouse, if so, this user's facial image is replaced the original sample in face database, otherwise end operation.

From prior art, can also recognize: Chinese patent application CN 103729625A discloses a kind of method of recognition of face, and Chinese patent application CN 103761504A discloses a kind of face identification system.

But, above-mentioned prior art only can provide a kind of system or method of recognition of face, face image data can not be utilized further, more detailed analyzing and processing is provided, such as to the analytic function of the sex of user, skin, face characteristic similarity and age etc., to cater to the particular demands of user.

Summary of the invention

For overcoming the deficiencies in the prior art, the invention provides a kind of recognition methods of face gender, its special feature is to comprise the following steps successively:

S1 uses camera head to obtain facial image;

S2 detects this facial image and whether comprises face, and when determining that this facial image comprises face, jump procedure S3;

S3 performs photo-irradiation treatment to this facial image, and convert this facial image to grey level histogram about pixel, the computing formula of wherein this conversion is:

$g(x,y)=a+\frac{\ln \left(f(x,y)\right)+1}{b\ln c}$

Wherein, f (x, y) is the pixel element data of input facial image, and g (x, y) exports the pixel element data of facial image, and a, b, c are the parameters adjusting the position of curve and shape and introduce;

S4 carries out colour space transformation to described grey level histogram, and is converted to HSV data mode, and the pixel of its iconic model unified is N × N;

S5 utilizes the data of each pixel of this facial image, calculates an eigenvector:

${{\mathrm{AF}}_i}^{\mathrm{HSV}}=\underset{N}{\mathrm{\Σ}}\underset{N}{\mathrm{\Σ}}{A}_{N\×N^{\·}}^{\mathrm{HSV}}\×C_i^{\mathrm{HSV}},$

Wherein, for matrix of coefficients, for HSV picture element matrix;

S6 utilizes basis function kernel, is changed by described eigenvector:

$k(z_1,z_2)=\mathrm{Cexp}\left(\frac{-{\left|\right|z_1-z_2\left|\right|}^2}{{\mathrm{\σ}}^2}\right).;$

This eigenvector after conversion is input in default sorter by S7, then by following utility function, calculates the object sex discriminant value of this facial image:

$F=\frac{{(<{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_M>-<{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_F>)}^2}{\mathrm{\&sigma;}{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_M+\mathrm{\&sigma;}{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_F}$

And mapped one to one by other object sex discriminant value of described object sex discriminant value and the determinacy in database and mate, thus judge the sex of described facial image.

According to one embodiment of present invention, this step S2 utilizes network neural method, character references method or primary standard of color method whether to comprise face to detect this facial image.

According to one embodiment of present invention, this eigenvector after this step S7 prize-winning conversion is input in sorter according to following rule:

$f\left(\mathrm{AF}\right)=\mathrm{sgn}(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_{i}k(X_i,\mathrm{AF})+b).$

According to one embodiment of present invention, before this recognition methods of enforcement, first set up the color image data storehouse comprising various sufficient amount, this database comprises three independently image sets, namely runs, verifies and tests.

By technique scheme, the recognition methods according to face gender of the present invention can realize following technique effect:

(1) other ability of objectivity in computer discriminant facial image is given;

(2) automatic intelligent management is realized.

Accompanying drawing explanation

Fig. 1 shows the identification method for distinguishing process flow diagram according to the embodiment of the present invention.

Embodiment

Describe the present invention in detail below in conjunction with accompanying drawing and specific implementation method, be used for explaining the present invention in exemplary embodiment and description of the present invention, but not as a limitation of the invention.

Before implementing the recognition methods according to the embodiment of the present invention, the color image data storehouse of various sufficient amount can be set up in advance.In order to verify the validity of this recognition methods, the self-defining data storehouse in conjunction with FERET database can be adopted.FERET database carries out image data base the most frequently used in the other task of face, but the face lazy weight of its Different Individual comprised is to support this recognition methods, therefore also needs to collect and set up own image data base.Automatically the facial image face extracted from database is detected by AdaBoost Face datection algorithm, then the detection of mistake is manually removed, and the data set of gained is comprised 10/500 image segments (such as, wherein 5/250 for each classification).This data set is divided into three independently image sets: run, verify and test, wherein, run the structure that collection is used for generating feature and SVM classifier, checking collection is for verifying the kernel function optimized parameter avoiding excessively running, Selection effect in the selection process, test machine is used for the performance of the sorter of operation to carry out evaluating and using.

As shown in Figure 1, after obtaining facial image by camera head, first pre-service is carried out to obtained image, comprise and confirm face, facial pre-service.

In the present embodiment, confirm that the method for face can comprise: the method for network neural method (Neural Network), character references method (Feature-based) and primary standard of color method (Color-based).Wherein, network neural method uses many training set data (Training Data) to go training network neural, and these data somes are face, a part of Shi Fei face, system picked out comprise the image of face.Character references method utilizes some features of people face to detect face, and such as: the eyes on face, nose, face, and these organs have fixing relative position; And with regard to whole face, the chances are ovalize, and between face and background, have edge line roughly, etc., these features all can go out face for systems scan.And primary standard of color rule utilizes the color of people face to judge whether certain image comprises face: well-known, the color of face generally includes the colors such as such as orange, yellow, white, brown, dark brown; If image is GTG, the change of the GTG color of face can't be too large, and eyes, face, hair are more black parts; Thus, according to the feature of the organs such as eyes, eyebrow, face, nose and geometry site each other, face can be detected.

Detect in this image after comprising face, usual needs carry out pre-service to this image, this is because often there is the even problem of uneven illumination in pending facial image, and this directly can have influence on the feature extraction precision of face, therefore photo-irradiation treatment must be carried out to improve picture quality to the image of input.Usually, grey level histogram can be used for the statistical relationship in representative digit image between each gray level and frequency of its appearance.For partially secretly, the image that partially bright, brightness range is not enough or contrast is not enough carries out histogram specification, and the histogram distribution of input picture can be made to be transformed into approximate specific histogram.Transforming function transformation function can select the function of the forms such as such as Gauss, Rayleigh, logarithm, index.Adopt log-transformation form in the present embodiment:

$g(x,y)=a+\frac{\ln \left(f(x,y)\right)+1}{b\ln c}$

In formula, f (x, y) is input picture, and g (x, y) is output image, the parameter that a, b, c are the adjustment position of curve and shape and introduce.The low tonal range of image can be made to be expanded by this conversion, high tonal range can be compressed, and the intensity profile of image is tending towards even.Image after greyscale transformation is carried out to the medium filtering of such as 3 × 3 again, remove the noise information introduced in image, to improve picture quality.

After completing facial pre-treatment step, perform linear discriminant.Linear discriminant is the method for carrying out dimensionality reduction based on sample class, and pattern-recognition has a wide range of applications.The method needs to find (or one group) axis of projection particularly, when ensureing that variance is minimum, can be separated well by dissimilar sample after making sample project to this space.Wherein, the amount of difference between metrics class average can be called between class scatter matrix, and the amount of measuring variance around these averages is called Scatter Matrix in class.The target differentiated finds such one (or one group) axis of projection, can maximize between class scatter minimizing in class while divergence.

This recognition methods is devoted to can by the object face in analysing digital image, to distinguish the sex of object.Such system can be applied to different fields, the digital signage network of such as robot, man-machine interaction, consensus data's collection, video monitoring, online audience measurement and other application programs many etc.In addition, the differentiation of sex may be used for the pre-treatment step of recognition of face.In machine learning techniques, use Gender Classification to have ubiquity, it allows the solution of sex identification and knowledge apply in any other image understanding or object classification task.

This recognition methods is as a kind of new Sex Discrimination method, because it is based on the RBF kernel of Nonlinear Support Vector Machines (SupportVector Machine) sorter, wherein SVM is a kind of method of supervised learning, can be widely used in statistical classification and regretional analysis.SVM belongs to vague generalization linear classifier, also can think the special case putting forward clo husband standardization (TikhonovRegularization) method.The feature of this sorter can minimize experience error simultaneously and maximize Geometry edge region, and therefore SVM is also referred to as maximal margin region classifier.SVM is information extraction from image segments, and these information is moved in the feature space of a lower dimension, and this is the algorithm generated according to the self-adaptive features of linear discriminant analysis (LDA principle) by optimizer device.For constructing full automatic face analysis system, the differentiation of sex will be used in the Face datection sorter of AdaBoost, and for selecting the analysis of candidate to connect.Then, pre-service is carried out to the image segments detected, to adjust its light characteristic, and it is transformed to unified yardstick.

After extracting face characteristic/image segments, use, comprising to obtain required discriminant value according to the method for the embodiment of the present invention: calculate appropriate feature, convert thereof into close basis function kernel, the object sex discriminant value being calculated this facial image by utility function.

In the present embodiment, colour space transformation is carried out to pretreated face image data, and/or carry out image scaling, and carry out calculating with adaptation function collection thus carry out svm classifier, to obtain preliminary kernel transition.Then, the face image data of input is converted to HSV data mode from three look plane (R, G, B) data modes.Further, by the face image data unified standard after this conversion be the form model of such as N × N.

In the present embodiment, utilize the data of each pixel of facial image after conversion, calculate an eigenvector:

${{\mathrm{AF}}_i}^{\mathrm{HSV}}=\underset{N}{\mathrm{\&Sigma;}}\underset{N}{\mathrm{\&Sigma;}}{A}_{N\&times;N^{\&CenterDot;}}^{\mathrm{HSV}}\&times;C_i^{\mathrm{HSV}},$

Wherein, for matrix of coefficients, i, j represent the row and column of each pixel of this grey level histogram respectively, and above-mentioned computing formula represents the row/column element of each pixel and matrix of coefficients be multiplied.

In the present embodiment, utilize basis function kernel, described eigenvector changed:

$k(z_1,z_2)=\mathrm{Cexp}\left(\frac{-{\left|\right|z_1-z_2\left|\right|}^2}{{\mathrm{\&sigma;}}^2}\right)..$

In the present embodiment, after being changed by kernel, the eigenvector obtained is as input data, and be input in Linear SVM sorter, wherein this input rule is:

$f\left(\mathrm{AF}\right)=\mathrm{sgn}(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_{i}k(X_i,\mathrm{AF})+b)$

In above formula, after obtaining above-mentioned eigenvector, and under the support of auxiliary data, wherein this auxiliary data comprises a stack features factor { X _i, auxiliary parameter group { y _i, { α _iand deviate b, these auxiliary datas and eigenvector are carried out computing in the lump, pass through utility function:

$F=\frac{{(<{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_M>-<{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_F>)}^2}{\mathrm{\&sigma;}{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_M+\mathrm{\&sigma;}{\left\{{{\mathrm{AF}}_i}^{\mathrm{HSV}}\right\}}_F}$

Wherein, utility function F calculates the mean square as feature/discriminant value, namely object sex discriminant value mentioned above, after drawing the mean square of feature/discriminant value, with differentiated in database that the mean square group that the face of sex carries out feature/discriminant value that same computing draws maps coupling one to one in advance, as then the sex representated by mean square of known characteristic of correspondence/discriminant value is man or female.

Above the technical scheme that the embodiment of the present invention provides is described in detail, apply specific case herein to set forth the principle of the embodiment of the present invention and embodiment, the explanation of above embodiment is only applicable to the principle helping to understand the embodiment of the present invention; Meanwhile, for one of ordinary skill in the art, according to the embodiment of the present invention, embodiment and range of application all will change, and in sum, this description should not be construed as limitation of the present invention.

Claims (4)

1. a recognition methods for face gender, is characterized in that, comprises the following steps successively:

S1, use camera head obtain facial image;

S2, detect described facial image and whether comprise face, and when determining that described facial image comprises face, jump procedure S3;

S3, perform photo-irradiation treatment to described facial image, utilize histogram specification to change described facial image, the computing formula of wherein said conversion is:

$g(x,y)=a+\frac{\ln \left(f(x,y)\right)+1}{b\ln c}$

Wherein, f (x, y) is the pixel element data of input facial image, and g (x, y) exports the pixel element data of facial image, and a, b, c are the parameters adjusting the position of curve and shape and introduce;

S4, carry out colour space transformation and convergent-divergent to described output facial image, and be converted to HSV data mode, the pixel of its iconic model unified is N × N;

The data of each pixel of facial image after S5, utilization conversion, calculate an eigenvector:

${\mathrm{AF}}_i^{\mathrm{HSV}}=\underset{N}{\mathrm{\&Sigma;}}\underset{N}{\mathrm{\&Sigma;}}{A}_{N\&times;N}^{\mathrm{HSV}}.\&times;C_i^{\mathrm{HSV}}.$

Wherein, for matrix of coefficients, for HSV picture element matrix;

S6, utilize basis function kernel, described eigenvector changed:

$k(z_1,z_2)=\mathrm{C\; exp}\left(\frac{-{\left|\right|z_1-z_2\left|\right|}^2}{{\mathrm{\&sigma;}}^2}\right).;$

S7, by conversion after described eigenvector be input in default sorter, then by following utility function, calculate the object sex discriminant value of described facial image:

$F=\frac{{(\&lang;{\left\{{\mathrm{AF}}_i^{\mathrm{HSV}}\right\}}_M\&rang;-\&lang;{\left\{{\mathrm{AF}}_i^{\mathrm{HSV}}\right\}}_F\&rang;)}^2}{\mathrm{\&sigma;}{\left\{{\mathrm{AF}}_i^{\mathrm{HSV}}\right\}}_M+\mathrm{\&sigma;}{\left\{{\mathrm{AF}}_i^{\mathrm{HSV}}\right\}}_F}$

And mapped one to one by other object sex discriminant value of described object sex discriminant value and the determinacy in database and mate, thus judge the sex of described facial image.

2. recognition methods according to claim 1, is characterized in that, described step S2 utilizes network neural method, character references method or primary standard of color method whether to comprise face to detect described facial image.

3. recognition methods according to claim 1, is characterized in that, is input in sorter in described step S7 by the described eigenvector after conversion according to following rule:

$f\left(\mathrm{AF}\right)=\mathrm{sgn}(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_{i}k(X_i,\mathrm{AF})+b).$

4. recognition methods according to claim 1, it is characterized in that, before the described recognition methods of enforcement, first set up the color image data storehouse comprising various sufficient amount, described database comprises three independently image sets, namely runs collection, checking collection and test set.