Summary of the invention
In view of above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide be based on multiple dimensioned linear Differential Characteristics low-rank
The gender identification method and system of expression, it is intended to solve its recognition efficiency of existing recognition methods and accuracy rate has and need to be improved
The problem of.
Technical scheme is as follows:
A kind of gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein comprising steps of
A, using the Face datection algorithm based on Haar-like feature and Adaboost classifier, it is cut out testing image
In human face region;
B, the textural characteristics based on multiple dimensioned linear Differential Characteristics are extracted on the human face region being cut out;
C, under low-rank constraint condition, the eigenmatrix that textural characteristics are constituted is projected into feature vector subspace;
D, in feature vector subspace, training logistic regression classifier model utilizes the logistic regression classifier mould
Type carries out gender identification to human face region.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein the step A is specific
Include:
A1, integrogram is used to calculate the rectangular characteristic of testing image as Weak Classifier;
A2, the Weak Classifier for representing face is picked out using Adaboost algorithm, by weak point in the way of Nearest Neighbor with Weighted Voting
Class device is configured to strong classifier;
A3, the cascade filtering that several strong classifiers that training obtains finally are composed in series to a cascade structure.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein the step B is specific
Include:
B1, facial modeling is carried out using active shape model;
B2, the human face characteristic point position obtained according to active shape model ajust the human face region rotation of testing image,
The human face region is zoomed in and out, row interpolation of going forward side by side processing;
B3, centered on the human face characteristic point position of part, multiple images block is marked off in human face region;
B4, textural characteristics are extracted on each image block.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein in the step C,
It is projected as the following formula:
Wherein, W is the subspace projection matrix that projection obtains,For class scatter matrix,For Scatter Matrix in class,Nuclear norm after being characterized matrix projection to feature vector subspace, for limiting subspace projection matrix after projection
The size of order.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein in the step D,
It is weighted to obtain weighted value to the value on the different dimensions of textural characteristics by subspace projection matrix W, and with logic letter
Weighted value is compressed to 0 ~ 1 range by number.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein in the step D,
The gender of training sample is that woman's probability is
,For the low-rank representation of textural characteristics.
The gender identification method based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein in the step D,
Using gradient descent method, the loss function for minimizing subspace projection matrix W obtains the optimal solution of subspace projection matrix W.
A kind of gender identifying system based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein include:
Shear module is sheared for using the Face datection algorithm based on Haar-like feature and Adaboost classifier
Human face region in testing image out;
Characteristic extracting module, for extracting the line based on multiple dimensioned linear Differential Characteristics on the human face region being cut out
Manage feature;
Low-rank representation module, under low-rank constraint condition, the eigenmatrix that textural characteristics are constituted to be projected to feature
Vector subspace;
Identification module, in feature vector subspace, training logistic regression classifier model to be returned using the logic
Sorter model is returned to carry out gender identification to human face region.
The gender identifying system based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein the shear module
It specifically includes:
Computing unit, for using integrogram to calculate the rectangular characteristic of testing image as Weak Classifier;
Ballot unit, for picking out the Weak Classifier for representing face using Adaboost algorithm, according to Nearest Neighbor with Weighted Voting
Weak Classifier is configured to strong classifier by mode;
Series unit, for several strong classifiers that training obtains finally to be composed in series to the stacking point of a cascade structure
Class device.
The gender identifying system based on multiple dimensioned linear Differential Characteristics low-rank representation, wherein the feature extraction
Module specifically includes:
Positioning unit, for carrying out facial modeling using active shape model;
Unit for scaling, the human face characteristic point position for being obtained according to active shape model, by the face area of testing image
Domain rotation is ajusted, and is zoomed in and out to the human face region, row interpolation of going forward side by side processing;
Image block retrieval unit, for being marked off in human face region multiple centered on the human face characteristic point position of part
Image block;
Feature extraction unit, for extracting textural characteristics on each image block.
The utility model has the advantages that being cut out the present invention is based on the Face datection of Haar-like feature and Adaboost classifier algorithm
Human face region in testing image;And it is special that the texture based on multiple dimensioned linear Differential Characteristics is extracted on the human face region being cut out
Sign;Then under low-rank constraint condition, textural characteristics are projected into subspace, in the subspace, training logistic regression classification
Device model is identified using the gender that model carries out face picture.Its recognition accuracy of recognition methods of the invention is high, and can be effective
Improve recognition efficiency.
Specific embodiment
The present invention provides gender identification method and system based on multiple dimensioned linear Differential Characteristics low-rank representation, to make this hair
Bright purpose, technical solution and effect are clearer, clear, and the present invention is described in more detail below.It should be appreciated that herein
Described specific embodiment is only used to explain the present invention, is not intended to limit the present invention.
Referring to Fig. 1, Fig. 1 be present pre-ferred embodiments flow chart, as shown, itself comprising steps of
S101, using the Face datection algorithm based on Haar-like feature and Adaboost classifier, be cut out to mapping
Human face region as in;
S102, the textural characteristics based on multiple dimensioned linear Differential Characteristics are extracted on the human face region being cut out;
S103, under low-rank constraint condition, by textural characteristics constitute eigenmatrix project to feature vector subspace;
S104, in feature vector subspace, training logistic regression classifier model, utilize the logistic regression classifier
Model carries out gender identification to human face region.
Wherein, step S101 mainly completes Face datection.As shown in Fig. 2, the present invention uses Viola method for detecting human face,
It is a kind of method based on integrogram, cascade detectors and AdaBoost algorithm.
Specifically comprising step:
S201, integrogram is used to calculate the rectangular characteristic of testing image as Weak Classifier;
S202, the Weak Classifier for representing face is picked out using Adaboost algorithm, it will be weak in the way of Nearest Neighbor with Weighted Voting
Classifier is configured to strong classifier;
S203, the cascade filtering that several strong classifiers that training obtains finally are composed in series to a cascade structure.
Firstly, the rectangle for specifically calculating testing image using integrogram is special using Harr-like character representation face
Sign;Then, some rectangular characteristics (Weak Classifier) that can most represent face are picked out using Adaboost algorithm, i.e., it will be to mapping
As constituting subgraph to be measured at multi-scale image by the continuous scaling of ratio,;According still further to Nearest Neighbor with Weighted Voting mode by Weak Classifier structure
It makes as strong classifier;Finally, several strong classifiers that training obtains to be composed in series to the cascade filtering of a cascade structure, grade
Connection structure can effectively improve the detection speed of classifier.
Step S102 mainly completes multiple dimensioned linear Differential Characteristics (LDF) and extracts.Linear differential feature (LDF) is to utilize figure
The second differnce information of picture carries out reasonable coding.Specifically comprising step:
S301, facial modeling is carried out using active shape model;
The human face region of testing image is rotated pendulum by S302, the human face characteristic point position obtained according to active shape model
Just, which is zoomed in and out, row interpolation of going forward side by side processing;
S303, centered on the human face characteristic point position of part, multiple images block is marked off in human face region;
S304, textural characteristics are extracted on each image block.
Specifically, the process that multiple dimensioned LDF is extracted is as follows: first with active shape model (Active Shape
Model, ASM) facial modeling is carried out, such as use 68 point models.The construction of active shape model is divided into training and search
Two steps.
When training: establishing the position constraint of human face characteristic point, building obtains the local feature of human face characteristic point;
When search: the matching including iteration.The first step is trained: collecting N number of training sample building active shape mould first
Type;Hand labeled human face characteristic point;The coordinate of human face characteristic point in training sample is conspired to create into feature vector;Feature vector is carried out
Normalization and alignment (alignment uses Procrustes method);Feature vector after alignment does PCA processing.It then, is everyone
Face characteristic point constructs local feature, it is therefore an objective to which each human face characteristic point can find new position in each iterative search procedures
It sets.Local feature generally uses Gradient Features, to prevent illumination variation.Second step be search: first calculating training sample in eyes (or
Person's eyes and mouth) position, do scale and rotationally-varying, be aligned face;Then, each human face characteristic point after alignment is attached
Nearly search, matches each local feature (frequently with mahalanobis distance), obtains preliminary configuration;Again with average face (active shape mould
Type) amendment matching result;Iteration is until convergence.
The position of human face characteristic point is obtained according to active shape model, human face region is rotated based on 2 eye coordinates points and is put
Just, and according to two eye distances from carry out scaling to three scales, using bilinearity cubic interpolation.Then, with part human face characteristic point
Centered on position, multiple subregions are divided in human face region and forms image block and takes out, for different size human face region, are taken
Image block is sized to certain out, and the present invention selects center of 16 human face characteristic points as the image block to be obtained, these
Human face characteristic point is mainly some points of some points without choosing face edge inside face.On each image block, extract
The textural characteristics of linear differential feature.Linear differential feature has more characterization ability, and can enhance image matter to a certain extent
Amount.Firstly, calculating center pixel and the radius in "horizontal", "vertical", " positive diagonal line ", " counter-diagonal " this four rhumb line
It for the size relation of 2 adjacent pixel, calculates and not simply mutually subtracts each other, but subtract each other again after successively subtracting each other, that is, extract
It is the second differnce information in this rhumb line;Obtain difference and then according to 0 comparison result carry out two-value turn to 0 or
1;Finally, using obtain 01 it is encoded translated be decimal value as characteristic value, due to this algorithm four rhumb line
Include the direction of 8 neighborhoods, thus encode come out value between 0 ~ 15 rather than as LBP 0 ~ 255 between, doing histogram
When statistics, the dimension of histogram also can be relatively low, even if compared with using the dimension histogram 0 of the 59 of the LBP of " uniform pattern ",
Dimension is also lower.
Step S103 mainly completes the low-rank representation of textural characteristics.
Specifically, multiple face pictures of the same person are considered, if linear for the extraction of each width face picture poor
The textural characteristics of dtex sign obtain a row vector, constitute the textural characteristics of multiple face pictures of multiple people as row vector
Eigenmatrix should be then low-rank matrix in the matrix theory.But due in reality, every width face picture can be by
To the influence of different factors, for example illumination, the noises such as block, translate.The eigenmatrix that these noise on human face pictures are constituted
Influence can be regarded as the effect of a noise matrix.So the present invention is when carrying out subspace projection, in addition low-rank constrains item
Part such as can be effectively reduced illumination, block at the influence of noises.Subspace projection algorithm of the invention is based on linear discriminant analysis
(Linear Discriminant Analysis, LDA), with conventional linear discriminant analysis model the difference is that: traditional
On the basis of model, in addition low-rank constraint condition.Optimizing expression is as follows:
Wherein, W is the subspace projection matrix that projection obtains,For class scatter matrix,For Scatter Matrix in class,Nuclear norm (the formula bottom right is " * " number) after being characterized matrix projection to feature vector subspace, after limiting projection
The size of subspace projection rank of matrix.By the optimization to above-mentioned goal expression, available subspace projection matrix.
Step S104 mainly completes logistic regression classifier model.
Specifically, a kind of logistic regression classifier model (Logic Regression) method based on discriminate, it
It is assumed that the example of class is linear separability, by the parameter of direct estimation discriminate, final prediction model is obtained.In the present invention
In mainly consider be based on binary classification logistic regression prediction model, i.e., classifier identification class label be 0(male) and 1(female).It patrols
It collects and returns in sorter model, the value on the different dimensions of textural characteristics is weighted by subspace projection matrix W
To weighted value, and with logic (Sigmoid) function weighted value is compressed to 0 ~ 1 range, be positive functional value as the sample
The probability of sample.Logic (Sigmoid) function is
In Logic Regression Models, the gender of training sample is that woman's probability is
,For the low-rank representation of textural characteristics.
Solution for Logic Regression Models exactly finds suitable subspace projection matrix W, so that in training sample
Men and women's picture is maximumlly classified correctly.The loss function of the problem is
Using gradient descent method, the loss function for minimizing subspace projection matrix W obtains the optimal of feature weight vector W
The logistic regression classifier model for gender identification can be obtained in solution.
Based on the above method, the present invention also provides the genders based on multiple dimensioned linear Differential Characteristics low-rank representation to identify system
System comprising:
Shear module is sheared for using the Face datection algorithm based on Haar-like feature and Adaboost classifier
Human face region in testing image out;
Characteristic extracting module, for extracting the line based on multiple dimensioned linear Differential Characteristics on the human face region being cut out
Manage feature;
Low-rank representation module, under low-rank constraint condition, the eigenmatrix that textural characteristics are constituted to be projected to feature
Vector subspace;
Identification module, in feature vector subspace, training logistic regression classifier model to be returned using the logic
Sorter model is returned to carry out gender identification to human face region.
Further, the shear module specifically includes:
Computing unit, for using integrogram to calculate the rectangular characteristic of testing image as Weak Classifier;
Ballot unit, for picking out the Weak Classifier for representing face using Adaboost algorithm, according to Nearest Neighbor with Weighted Voting
Weak Classifier is configured to strong classifier by mode;
Series unit, for several strong classifiers that training obtains finally to be composed in series to the stacking point of a cascade structure
Class device.
Further, the characteristic extracting module specifically includes:
Positioning unit, for carrying out facial modeling using active shape model;
Unit for scaling, the human face characteristic point position for being obtained according to active shape model, by the face area of testing image
Domain rotation is ajusted, and is zoomed in and out to the human face region, row interpolation of going forward side by side processing;
Image block retrieval unit, for being marked off in human face region multiple centered on the human face characteristic point position of part
Image block;
Feature extraction unit, for extracting textural characteristics on each image block.
It has been described in detail in the method for technical detail in front about above-mentioned modular unit, so it will not be repeated.
It should be understood that the application of the present invention is not limited to the above for those of ordinary skills can
With improvement or transformation based on the above description, all these modifications and variations all should belong to the guarantor of appended claims of the present invention
Protect range.