CN103294998B - A kind of face visual feature representation method based on attribute space - Google Patents

Abstract

The invention discloses a kind of face visual feature representation method based on attribute space, the method comprises: carry out digital simulation, pre-service and data evaluation and test to 64 kinds of face overall situations or local attribute, extract " bottom " feature of data; Utilize gauss hybrid models to estimate attribute data distribution, form face character space; Input human face data is projected in attribute space, obtains the face visualization feature quantized; By Z-score method, standardization is carried out to visualization feature, can more effective input data be identified.The inventive method can obtain the feature differentiated needed for face more accurately, can be widely used in other pattern-recognition and advanced intelligent domain model.

Description

A kind of face visual feature representation method based on attribute space

Technical field

The invention belongs to image procossing and mode identification technology, relate to a kind of face visual feature representation method based on attribute space.

Background technology

Along with the fast development of computing machine and image processing techniques, public safety problem receives that society is increasing to be paid close attention to, and various identity recognizing technology emerges in an endless stream.Recognition of face makes it have huge advantage with its feature such as direct, convenient, friendly, hidden.Recognition of face relates to pattern-recognition, computer vision, intelligent man-machine interactively, computer graphics, multiple subject such as cognitive science.As living things feature recognition gordian technique, recognition of face has good application prospect in fields such as information security, security protection, finance.

People have done a lot of good try in field of face identification and have achieved great successes, but utilize computing machine to carry out automatically, efficiently recognition of face still has many difficulties, main manifestations is: face is complicated, changes various, facial image and can be subject to comprising illumination, expression, attitude, shooting the visual angle even impact of the factor such as camera specification.

Summary of the invention

The object of the invention is to propose a kind of method studying face visualization feature in face character space, to excavate the effect of face visualization feature to recognition of face.

The technical solution used in the present invention is:

The invention provides a kind of face visual feature representation method based on attribute space, comprise the following steps:

Step 1: face visualization feature is counted as face character, obtains the face overall situation and local attribute that can be used for recognition of face; Some basic face characters, if sex, race etc. is the outstanding attribute for differentiating face, but only can't carry out recognition of face with these attributes completely effectively, still need to obtain some other attributes and are supplemented;

Step 2: carry out digital simulation to the face character that step 1 is chosen, to data prediction, extracts " bottom " feature;

Digital simulation carries out in the following ways, and the people in evaluation and test group first carries out primary election to the data of every attribute, data is issued other members of group afterwards, only has group member all to agree to, these data just can be added in this type of attribute data;

For global property, as sex, age etc., the advanced row Face datection of attribute data, is remedied in standard coordinate system with the key point coordinate returned by face; If x ＇ and y ＇ is the coordinate figure of conversion preceding pixel, x and y is the coordinate figure of pixel after conversion; A, b, c, d, e, f are affined transformation coefficients; Affined transformation computing formula is as follows:

$\left\{\begin{array}{c}x={\mathrm{ax}}^{\′}+{\mathrm{by}}^{\′}+c\\ y={\mathrm{dx}}^{\′}+{\mathrm{ey}}^{\′}+f\end{array}\right.$

For local attribute, as the width of eyes, the shape etc. of eyebrow, attribute data can be divided into 8 pieces of face subregions, in subregion, obtains corresponding visualization feature;

Finally, pre-service completes, and extracts sift as " bottom " feature to all properties data;

Step 3: utilize support vector machine (SVM) to test and assess to attribute data, obtains the optimization simulation of attribute and data; When the data selected are after step 2 processes, every attribute is provided to positive example (same attribute data) and the counter-example (different attribute data) of tape label, for testing attribute and data rationality;

Step 4: mixed Gauss model (GMM) obtains the visualization feature of input face; Utilize mixed Gauss model (GMM) analog nature Data distribution8, form attribute space; Calculate the posterior probability of input human face data various attribute data distribution in attribute space; The corresponding face character of posterior probability is counted as the visualization feature of face, for the face feature vector F (I) of input, according to obtaining face character V _i=1...n, calculating corresponding visualization feature is:

$V\left(I\right)=\{V_1\left(F\left(I\right)\right),...,V_n\left(F\left(I\right)\right)\}$

Step 5: adopt Z-Score method to carry out standardization to the visualization feature be quantized; Standardization formula is as follows:

$V_i^{\′}\left(F\left(I\right)\right)=\frac{V_i\left(F\left(I\right)\right)-E\left(V\left(I\right)\right)}{\mathrm{std}\left(V\left(I\right)\right)}$

In formula, average, standard deviation that E (V (I)), std (V (I)) they are data; After standardization completes, the COS distance of vector is utilized to obtain inputting the similarity of data A, B:

$\mathrm{Sim}(A,B)=\frac{\mathrm{\ΣV}\_A\left(I\right)*V\_B\left(I\right)}{\sqrt{(\mathrm{\ΣV}\_A{\left(I\right)}^2)(\mathrm{\ΣV}\_B{\left(I\right)}^2)}}$

Finally use Sim (A, B) to carry out recognition of face and checking, complete recognition of face and checking.

A kind of face visual feature representation method based on attribute space provided by the invention, its advantage and good effect are:

1, the method is predicted each attribute affecting recognition of face, carries out digital simulation to attribute, utilizes mixed Gauss model to model attributes, forms attribute space.

2, the method is based on pattern-recognition and integration processing theory, excavate the effect of face visualization feature in recognition of face, in attribute space, face " bottom " feature is changed, reduce the susceptibility of factors vary in recognition of face such as attitude, illumination, expression, this achievement can also be generalized in other pattern recognition model and advanced intelligent domain model.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the inventive method.

Fig. 2 is people face part attribute test result figure.

Fig. 3 is the mapping schematic diagram of face in attribute space.

Embodiment

Basic thought of the present invention is by carrying out modeling to face character, is associated by the face of input with attribute space, utilizes mixed Gauss model to excavate the visualization feature of input face, in high-level characteristic, carries out recognition of face and checking.

According to above thought, flow process of the present invention as shown in Figure 1, further illustrates method of the present invention below in conjunction with technical scheme and accompanying drawing.

First analyze the impact of various face character on recognition of face, obtain the face overall situation and local attribute, digital simulation and test and appraisal are carried out to various face character simultaneously.Secondly, Image semantic classification is carried out to input face and simulated data, and extracts " bottom " feature.Again, according to mixed Gauss model, the various visualization feature of input face in attribute space are also obtained to model attributes.Finally carry out recognition of face and checking.

To input human face data A and B, the concrete steps of this method are as follows:

Step 1: regard face visualization feature as face character, the prediction face overall situation and local attribute; Namely global property observes the characteristic that whole face draws, local attribute is then taken from each sub regions of face, and in practice, the present invention have chosen 64 kinds of face characters.

The face global property chosen is as follows:

The face local attribute chosen is as follows:

Step 2: carry out digital simulation to face character, to data prediction, extracts " bottom " feature;

2.1: when carrying out face character simulation, the people in evaluation and test group first carries out data primary election to this attribute, and such as attribute A1 is D by the data selected ₁={ D ₁₁, D ₁₂..., D _1n, D ₁in data from D ₁₁start to issue other members of group successively and carry out subjectivity evaluation and test, evaluating standard is 0,1, and only have when evaluation result is all 1, these data are passed through, and are added in this type of attribute data, otherwise, then abandon.

2.2: for face global property, the advanced row Face datection of attribute data, its three key point coordinates returned are followed successively by: left eye center, right eye center, mouth center.Set a front face as standard faces, select above-mentioned three key points to be that reference point carries out affined transformation.

Two dimensional affine transformation for mula is as follows:

$\left\{\begin{array}{c}x={\mathrm{ax}}^{\′}+{\mathrm{by}}^{\′}+c\\ y={\mathrm{dx}}^{\′}+{\mathrm{ey}}^{\′}+f\end{array}\right.$

Wherein x ＇, y ＇ are the coordinate figures of conversion preceding pixel, and x, y are the coordinate figures of pixel after conversion, and a, b, c, d, e, f are affined transformation coefficients.According to above-mentioned coordinate information, obtain affined transformation coefficient and carry out face alignment.

2.3: for local attribute, as opened one's mouth, eye is closed etc., and Face datection is identical with step 2.2 with face alignment procedure, but attribute data will be divided into 8 pieces of face subregions.That is: forehead, eyebrow, eyes, nose, cheek, beard, mouth and chin.

2.4: choose " bottom " feature that sift is data.Global property data obtain feature after processing according to step 2.2: fea_global, and local attribute's data obtain feature after processing according to step 2.3: fea_local={fea_1 ..., fea_n}, in formula, n represents face zone number.Input data A or B will obtain fea_global, fea_local simultaneously.

Step 3: evaluated and tested attribute data by support vector machine (SVM), obtains the optimization simulation of data and attribute;

3.1: after attribute data is determined, form 300 pairs of positive examples and 300 pairs of counter-examples with attribute data and some non-attribute datas, by support vector machine (SVM), the test data formed is evaluated and tested.When the evaluation result of attribute is lower than benchmark, digital simulation need be re-started to this attribute, in reality, use the SVM that kernel function is RBF.Attribute evaluation and test partial results as shown in Figure 2.

Step 4: the visualization feature utilizing mixed Gauss model to obtain input face, as shown in Figure 3.

4.1: utilize mixed Gauss model (GMM) analog nature Data distribution8.As face character class C comprises round face, square face etc., in same type (C class), all properties mixed Gaussian degree K is identical.X represents the given data of Attribute class C, and mixed Gauss model parameter is estimated by X;

Step 4.2: calculate the probability of input human face data in various property distribution, form attribute space.Human face data in the posterior probability of Attribute class C be:

$p\left(x\right|\mathrm{\Θ})=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\ω}}_k^{C}N(x;{\mathrm{\μ}}_k^C,{\mathrm{\Σ}}_k^C)$

In formula, Attribute class C selects round face, square face etc. successively, represent the weights of K Gauss in attribute, average, variance;

Step 4.3: the corresponding face character of posterior probability is counted as the visualization feature of face.The proper vector F (I) of input face, according to obtaining attribute V _{i=1 ... n}, visualization feature can be expressed as: V (I)={ V ₁(F (I)) ..., V _n(F (I)) }.

Step 5: adopt Z-Score method to carry out standardization to the visualization feature be quantized.Standardisation process is as follows:

$V_i^{\′}\left(F\left(I\right)\right)=\frac{V_i\left(F\left(I\right)\right)-E\left(V\left(I\right)\right)}{\mathrm{std}\left(V\left(I\right)\right)}$

In formula, average, standard deviation that E (V (I)), std (V (I)) they are data.After standardization completes, the COS distance of vector is utilized to ask the similarity of input A, B.Formula is as follows:

$\mathrm{Sim}(A,B)=\frac{\mathrm{\ΣV}\_A\left(I\right)*V\_B\left(I\right)}{\sqrt{(\mathrm{\ΣV}\_A{\left(I\right)}^2)(\mathrm{\ΣV}\_B{\left(I\right)}^2)}}$

Sim (A, B) is finally utilized to carry out recognition of face and checking.

Claims (2)

1. based on a face visual feature representation method for attribute space, it is characterized in that, specifically comprise the following steps:

Step 1: prediction is used for the face overall situation and the local attribute of recognition of face;

Step 2: carry out digital simulation to the face character that step 1 is chosen, to data prediction, extracts low-level image feature; Digital simulation carries out in the following ways, and the people in evaluation and test group first carries out primary election to the data of every attribute, data is issued other members of group afterwards, only has group member all to agree to, these data just can be added in this type of attribute data;

Step 3: evaluated and tested attribute data by support vector machine, obtains the optimization simulation of attribute and data;

Step 4: utilize mixed Gauss model analog nature Data distribution8, forms attribute space, calculates the posterior probability of input human face data various attribute data distribution in attribute space, and the corresponding face character of posterior probability is considered as the visualization feature of face; For the face feature vector F (I) of input, according to obtaining face character V _{i=1 ... n}, calculating corresponding visualization feature is:

V(I)＝{V ₁(F(I)),…,V _n(F(I))}；

Step 5: adopt Z-Score method to carry out standardization to the visualization feature be quantized, standardization formula is as follows:

$V_i^{\′}\left(F\right(I\left)\right)=\frac{V_i\left(F\right(I\left)\right)-E\left(V\right(I\left)\right)}{std\left(V\right(I\left)\right)}$

In formula, average, standard deviation that E (V (I)), std (V (I)) they are data;

After standardization completes, the COS distance of proper vector is utilized to obtain inputting the similarity of data A, B:

$Sim(A,B)=\frac{\mathrm{\Σ}V\_A\left(I\right)*V\_B\left(I\right)}{\sqrt{(\mathrm{\Σ}V\_A{\left(I\right)}^2)(\mathrm{\Σ}V\_B{\left(I\right)}^2)}}$

Finally use Sim (A, B) to carry out recognition of face and checking, complete recognition of face and checking.

2. a kind of face visual feature representation method based on attribute space according to claim 1, is characterized in that, the sub-attribute mixed Gaussian degree that described step 4 comprises same Attribute class is identical.