CN105550634A - Facial pose recognition method based on Gabor features and dictionary learning - Google Patents

Abstract

The invention discloses a facial pose recognition method based on Gabor features and dictionary learning. The facial pose recognition method comprises the following steps that firstly a facial pose is discretized into different subspaces, and a child dictionary is trained for each subspace by using K-SVD so that each subspace is enabled to be corresponding to one category; then all the dictionaries are combined into a super complete dictionary; and finally pose classification is performed by adopting a method based on the gabor features and sparse expression. A shielding face dictionary is reconstructed in order to enhance robustness of the algorithm so that the problem of face shielding in facial pose recognition can be solved. The problems of illumination, noise and shielding in facial pose estimation can be solved so that the front face, head raising, nodding, left deflection, the left side face, right deflection and the right side face can be rapidly and robustly recognized. The facial pose recognition method can be greatly applied to the field of safe driving, human-computer and face recognition.

Description

Based on the human face posture recognition methods of Gabor characteristic and dictionary learning

Technical field

The invention belongs to image procossing, pattern-recognition, computer vision and human-computer interaction technique field, relate to a kind of human face posture recognition methods, be specifically related to a kind of human face posture recognition methods based on dictionary learning and rarefaction representation.

Background technology

Human face modeling has huge application prospect in intelligent video monitoring, recognition of face, man-machine interaction and field of virtual reality.Such as, in intelligent video monitoring, human face modeling can be applied to driving supervisory system, identifies that whether driver focuses one's attention on driving, avoids the generation of collision situation by the human face posture change monitoring driver.In addition, the accuracy of human face modeling to recognition of face has a great impact, many face recognition algorithms can reach good discrimination to front face image, but for the non-frontal facial image of multi-pose, its recognition accuracy meeting degradation, and be a kind of important channel solving plurality of human faces gesture recognition by human face posture pre-estimation.

Current existing human face posture detection method can be divided three classes substantially: texture subspace method, 3D method, other class methods.First kind method is by realizing the Detection and estimation of attitude based on the learning method of 2D face outward appearance.Wherein more typically there are principal component analysis (PCA) and linear discriminant (LDA) etc.Because PCA is a kind of linear dimension reduction method, and human face posture 3D to rotate in the degree that alters a great deal be a kind of nonlinearities change.Therefore scholars use kernel principal component analysis (KPCA), and Manifold learning method solves this nonlinearities change problem.But kernel method and Manifold learning method have a defect: along with face training sample increases, it is difficult to isolate identity and attitude.This just means, when face training storehouse is enough large, the accuracy rate of Attitude estimation can change according to the difference of people.The maximum feature of first kind method is that processing speed is fast, easily realizes, but needs the training by great amount of samples, comparatively responsive to the change such as illumination, expression of face, particularly declines obviously to its accuracy rate of video human face image of illumination extreme difference.

Equations of The Second Kind method thinks that human face posture detects an inherently 3D problem, only has the essential characteristic that could be characterized human face posture by 3D information.Therefore these class methods characterize different attitude often through extraction 3D feature, or utilize the multiple image under different visual angles, rebuild the detection of the 3D model realization attitude of face in three dimensions.These class methods often to the size of image and quality requirements very high, and a large amount of operation time can be spent.Equations of The Second Kind method can obtain very high-accuracy by 3D method, but real-time is not high, simultaneously to the ultra-low resolution in video monitoring with to block facial image effect be not fine.

3rd class methods are some non-mainstream methods, can only solve subproblem in human face modeling and maybe can only be applied to some specific occasion.Such as, Rafael the face pose estimation of polyphaser is proposed Deng people.In order to correctly estimate human face posture, the 6 width images needing to utilize 6 cameras all around to take pictures in their method carry out fusion and differentiate.The method that the people such as J.Nuevo propose block cluster carries out human face modeling, achieves good effect, but the attitude that their method is estimated is limited in scope (attitudes vibration that can only identify 45 degree of scopes).The people such as the Chen Zhenxue of Shandong University propose leg-of-mutton face pose estimation, obtain the accuracy rate of about 91%, but their method can only deflect effectively around Y and Z axis face, and invalid for rotating human face posture around X-axis, namely loses efficacy to the upper and lower rotational case of face.

Want computing machine to possess the gesture recognition ability the same with the mankind at present to be also difficult to, main cause is illumination, noise, block, the change of the factor such as resolution, identity, expression all can produce tremendous influence to the accuracy of Attitude estimation, the impact how eliminating these factors is the problem needing solution at present badly.

Summary of the invention

The object of the invention is to the deficiency overcoming above method, a kind of human face posture recognition methods based on Gabor characteristic and dictionary learning is proposed, the problem such as solve illumination, the noise in human face posture identification and block, the identifying front, come back, nod of robustness, left avertence turns, left side face, right avertence turn and right side face.

Object of the present invention is achieved through the following technical solutions:

Based on a human face posture recognition methods for Gabor characteristic and dictionary learning, comprise the following steps:

Gabor characteristic extraction is carried out to the human face posture image to be identified of online input, builds Gabor characteristic vector y;

Use the complete dictionary of attitude to carry out linear combination expression to described Gabor characteristic vector y, set up sparse representation model and solve coefficient vector, wherein Gabor characteristic vector m is Gabor characteristic vector dimension;

Coefficient vector according to the above-mentioned linear combination solved carries out human face posture Classification and Identification.

Further, described human face posture is divided into 7 different attitude classifications, is defined as that left avertence turns, left side face, right avertence turn respectively, right side face, front, comes back and nod, the subspace that its each correspondence is different.

Further, the complete dictionary of attitude is used to carry out before linear combination represents to described Gabor characteristic vector y, also comprise the training step of the complete dictionary of described attitude, the complete dictionary of wherein said attitude comprise corresponding unscreened human face posture the first attitude complete dictionary D and to there being the complete dictionary D of human face posture second attitude blocked _e, described first attitude complete dictionary D and the complete dictionary D of described second attitude _etraining independently complete.

Further, the training process of described first attitude complete dictionary D is specific as follows:

Collect the human face posture image pattern of each attitude classification respectively, and the human face posture Gabor characteristic training set that Gabor filtering process and feature extraction dyad form each attitude classification is carried out to described human face posture image pattern;

Use K-SVD to carry out training optimization to the human face posture Gabor characteristic training set of every class attitude classification and draw best sub-dictionary D respectively _i, i=1,2 ..., 7;

By sub-for all kinds of the best dictionary D _iform the complete dictionary D=of the first attitude [D ₁, D ₂..., D ₇].

Further, the Gabor characteristic of described human face posture image is extracted as:

$S=(a_{\mathrm{0,0}}^{\left(\mathrm{\ρ}\right)};a_{\mathrm{1,0}}^{\left(\mathrm{\ρ}\right)};...;a_{\mathrm{7,4}}^{\left(\mathrm{\ρ}\right)})$

Wherein, by the mould to Gabor filter factor carry out ρ sampling and the column vector that obtains, μ, v are direction and the yardstick of Gabor filter, adopt the Gabor filter of direction and Scale invariant to describe the feature of human face posture image, for face pose presentation and Gabor core ψ _{μ, v}convolution, Gabor core is defined as:

${\mathrm{\ψ}}_{\mathrm{\μ},v}\left(z\right)=\frac{{\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2}{{\mathrm{\σ}}^2}{e}^{-{\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2{\left|\right|z\left|\right|}^2/2{\mathrm{\σ}}^2}(e^{i{k}_{\mathrm{\μ},v}z}-e^{-{\mathrm{\σ}}^2/2})$

Wherein, z (x, y) represents pixel; for small echo item, k _v=k _max/ f ^v, φ _μ=π μ/8, σ=1.5 π controls the ratio of Gauss's window width and wavelength.

Further, when the human face posture image to be identified of described online input is unscreened human face posture image, extract the Gabor characteristic vector of described human face posture image to be estimated m is Gabor characteristic vector dimension, y is regarded as the complete dictionary of described first attitude linear combination represent: wherein N is the total number of dictionary atom, and described sparse representation model is

Further, described sparse representation model

Solve by being with the least square method of sparse constraint:

$x=\underset{x}{\min }{\left|\right|y-\mathrm{Dx}\left|\right|}_F^2+\mathrm{\λ}{\left|\right|x\left|\right|}_1$

Wherein, λ is balance factor, plays Equilibrium fitting error and openness effect.

Further, the human face posture image to be identified of described online input is when having the human face posture image blocked, and extracts the Gabor characteristic vector of described human face posture image to be estimated m is Gabor characteristic vector dimension, is regarded as by y by the described first complete dictionary D and the complete dictionary of the second attitude linear combination represent:

$y=y_0+e_0=\mathrm{Dx}+e_0=[D,D_e]\left[\begin{array}{c}x\\ x_e\end{array}\right]=\mathrm{B\ω},$

Wherein unobstructed image y ₀with block error image e ₀respectively can by the first attitude complete dictionary D and the complete dictionary of the second attitude rarefaction representation, D _efor orthogonal matrices, described sparse representation model is

Further, described sparse representation model following optimization problem is changed into by there being the human face posture problem of image recognition of blocking:

This problem can be solved by the linearity specifications method of standard.

Further, the described coefficient vector according to the above-mentioned linear combination solved carries out human face posture Classification and Identification and adds up specifically by carrying out validity to the coefficient vector of the described linear combination solved, using the maximum foundation as judging classification of aggregate-value, namely

$\underset{i}{\max }{p}_i\left(y\right)=\underset{i}{\mathrm{\Σ}}{\mathrm{\δ}}_i\left(f\left({\hat{x}}_1\right)\right)$

Wherein, f () is special function, and the negative factor for the expression coefficient by sparse representation model sets to 0, selection of Function, for only selecting the expression coefficient of sparse representation model in expression coefficient corresponding to the i-th class subspace matrices, and other is represented that coefficient sets to 0, p _iy () is that the validity of the expression coefficient of subspace that in training sample set, the i-th class attitude classification the is corresponding sparse representation model corresponding with it adds up the factor.

The present invention has following advantage and effect relative to prior art:

1) current most of face gesture recognition method is only classified to face deflection, and the present invention not only can classify to face deflection, with carrying out upper and lower attitude classification.The present invention can identify the roughly left and right deflection state up and down of facial image at about 0.3 second, accuracy of identification, more than 95%, can be applied to the field such as safe driving and man-machine interaction preferably.

2) owing to the present invention is based on dictionary learning and rarefaction representation principle, so illumination, noise in human face posture identification can be solved, express one's feelings and the difficult problem such as to block simultaneously;

3) because Gabor filter can from the local feature of multiple dimensioned multi-direction extraction image.Therefore, the present invention adopts the Gabor characteristic of facial image as the atom vector of super complete dictionary.The method improves the stability of human face posture identification.

Accompanying drawing explanation

In order to be illustrated more clearly in the technical scheme in the embodiment of the present invention, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

The division classification schematic diagram of human face posture image in Fig. 1 the present invention;

The main-process stream schematic diagram of Fig. 2 human face posture recognition methods based on Gabor characteristic and dictionary learning disclosed by the invention;

In Fig. 3 the present invention, human face posture image Gabor characteristic extracts schematic diagram;

Based on the human face posture classification schematic diagram of rarefaction representation in Fig. 4 human face posture recognition methods disclosed by the invention.

Embodiment

The technological means realized for making the present invention, creation characteristic, reach object and effect is easy to understand, developing simultaneously referring to accompanying drawing, the present invention is described in more detail for embodiment.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Term " first ", " second ", " the 3rd " and " the 4th " etc. in instructions of the present invention and claims and above-mentioned accompanying drawing are for distinguishing different object, instead of for describing particular order.In addition, term " comprises " and " having " and their any distortion, and intention is to cover not exclusive comprising.Such as contain the process of series of steps or unit, method, system, product or equipment and be not defined in the step or unit listed, but also comprise the step or unit do not listed alternatively, or also comprise alternatively for other intrinsic step of these processes, method, product or equipment or unit.

To be described in detail respectively according to embodiment below.

Embodiment

Based on the human face posture recognition methods of Gabor characteristic and dictionary learning disclosed in the embodiment of the present invention, mainly carry out human face posture classification according to the thought of dictionary learning and rarefaction representation.

In advance human face posture is divided, turn to different subspaces by discrete for human face posture, the corresponding a kind of human face posture classification of each word space.The definition of the division subspace corresponding to it of 7 kinds of inhomogeneity attitude classifications is adopted in the embodiment of the present invention, by discrete for human face posture be 7 different subspace, i.e. left side 1, left side 2, right side 1 and right side 2, front, new line, first-class 7 the attitude classifications of point, as shown in Figure 1, be defined as that left avertence turns, left side face, right avertence turn respectively, right side face, front, come back and nod.

The training of attitude dictionary and attitude Classification and Identification two large divisions is comprised based on the human face posture recognition methods of Gabor characteristic and dictionary learning disclosed in the embodiment of the present invention.Wherein attitude dictionary training part comprises again Gabor characteristic extraction, and K-SVD dictionary is optimized and built three steps such as the complete dictionary of attitude; Attitude Classification and Identification part comprises Gabor characteristic and extracts, and sparse representation model coupling solves and 3 steps such as attitude Classification and Identification.One by one each step is described in detail below, based on the human face posture recognition methods of Gabor characteristic and dictionary learning flow process as shown in Figure 2:

Step S1: attitude dictionary training

This part is that off-line carries out.First, collect the human face posture image pattern of each attitude classification respectively, and the human face posture Gabor characteristic training set that Gabor filtering process and feature extraction dyad form each attitude classification is carried out to described human face posture image pattern, human face posture is divided into 7 different subspace in the present embodiment, i.e. left side 1, left side 2, right side 1 and right side 2, front, new line, first-class 7 the attitude classifications of point, therefore respectively to quantizing the different human face posture Gabor characteristic training set of generation 7.Then, adopt K-SVD method to carry out training optimization to Different categories of samples collection and obtain all kinds of sub-dictionary.Finally, all kinds of sub-dictionary is merged the complete attitude dictionary of composition.

Wherein the complete dictionary of attitude comprises the first attitude complete dictionary D and the complete dictionary D of the second attitude _e, wherein the first attitude complete dictionary D corresponding unscreened human face posture, wherein complete dictionary D of the second attitude _ethe human face posture that correspondence is blocked, has another name called and blocks corrosion dictionary D _e.Described first attitude complete dictionary D and the complete dictionary D of described second attitude _etraining independently complete.

Concrete steps and embodiment comprise:

Step S1a:Gabor feature extraction

The Gabor characteristic of human face posture image is extracted as:

$S=(a_{\mathrm{0,0}}^{\left(\mathrm{\ρ}\right)};a_{\mathrm{1,0}}^{\left(\mathrm{\ρ}\right)};...;a_{\mathrm{7,4}}^{\left(\mathrm{\ρ}\right)})$

Wherein, by the mould to Gabor filter factor carry out ρ sampling and the column vector that obtains, μ, v are direction and the yardstick of Gabor filter, adopt the Gabor filter of direction and Scale invariant to describe the feature of human face posture image. for face pose presentation and Gabor core ψ _{μ, v}convolution.Gabor core is defined as:

${\mathrm{\ψ}}_{\mathrm{\μ},v}\left(z\right)=\frac{{\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2}{{\mathrm{\σ}}^2}{e}^{-{\left|\right|k_{\mathrm{\μ},v}\left|\right|}^2{\left|\right|z\left|\right|}^2/2{\mathrm{\σ}}^2}(e^{i{k}_{\mathrm{\μ},v}z}-e^{-{\mathrm{\σ}}^2/2})$

Wherein, z (x, y) represents pixel; for small echo item, k _v=k _max/ f ^v, φ _μ=π μ/8, σ=1.5 π controls the ratio of Gauss's window width and wavelength.The value of each parameter of the present invention is k _max=pi/2, ρ ≈ 40, μ=0 ..., 7}, v={0 ..., 4}.The example that Gabor characteristic is extracted as shown in Figure 3.

Step S1b:K-SVD optimizes training

First, in employing, one step process carries out Gabor characteristic extraction composition training set to all training samples, then, adopts K-SVD method to be optimized training obtain the complete dictionary D of attitude to training set.K-SVD is a kind of dictionary training algorithm of classics, according to error minimum principle, carries out SVD decomposition to error term, selects the decomposition item making error minimum as the dictionary atom upgraded and corresponding atom coefficient, through continuous iteration thus the solution be optimized.

Detailed process is: first, uses K-SVD to carry out training optimization draw best sub-dictionary D to every class attitude sample set _i; Then, by all kinds of sub-dictionary D _iform the complete dictionary D=of the first attitude [D ₁, D ₂..., D ₇].

The present invention adopts K-SVD to carry out training optimization and draws best sub-dictionary D _i.K-SVD is the process of a kind of cross-iteration rarefaction representation and dictionary updating.For having n _iindividual sample training collection, the objective function of K-SVD is:

$\begin{array}{ccc}\underset{D,X}{\min }{\left|\right|Y-D_{i}X\left|\right|}_F^2& s.t.& \∀i,{\left|\right|x_i\left|\right|}_0\≤T_0\end{array}$

Wherein, for the sparse coefficient collection of sample, be i-th training sample y _irarefaction representation coefficient vector, k is dictionary D _iatom number.K-SVD algorithm comprises two processes.In first process, fixing dictionary D _i, therefore above formula objective function becomes the optimization problem solving rarefaction representation coefficient, has many kinds of method for solving at present.Second process is that the sparse coefficient utilizing the first process to solve upgrades dictionary D _i.This process is by progressive updating D _ioften row d _kwith i-th row of X realize.And d _k, solve, be realized by svd (SVD).

Step S1c: build the complete dictionary of attitude

The sub-dictionary D after all kinds of optimization is obtained by step 1 (b) _i, all sub-dictionaries are merged the super complete dictionary of composition.Because human face posture is divided into 7 kinds by the present invention, therefore the complete dictionary of human face posture is: wherein, m is sample characteristics dimension, and n is the total number of dictionary atom.

Step 2: attitude Classification and Identification

This part is at line process.First, to online its Gabor characteristic vector of input human face posture image zooming-out.Then, use the complete dictionary of attitude built to carry out linear combination expression to the Gabor characteristic vector of the human face posture image of online input, set up sparse representation model, and adopt Lars-Lasso algorithm to carry out Model Matching to solve.Finally, utilize and solve coefficient and carry out attitude classification.Concrete steps and embodiment comprise:

Step S2a:Gabor feature extraction

This part is identical with step S1a, extracts the Gabor characteristic vector y of the human face posture image of online input.

Step S2b: sparse representation model mates and solves

After extracting the Gabor characteristic vector y of the human face posture image of online input, the complete dictionary of attitude built is used to carry out linear combination expression, vectorization composition human face posture Gabor characteristic training set.

Wherein the complete dictionary of attitude comprises the first attitude complete dictionary D and the complete dictionary D of the second attitude _e, wherein the first attitude complete dictionary D corresponding unscreened human face posture, wherein complete dictionary D of the second attitude _eto the human face posture blocked should be had, have another name called and block corrosion dictionary D _e.Extract the Gabor characteristic vector of human face posture image to be estimated m is Gabor characteristic vector dimension, y is regarded as the complete dictionary of the first attitude the linear combination of (N is the total number of dictionary atom) represents:

Wherein it is a very sparse linear combination coefficient vector.Ideally, if test sample y belongs to the i-th class attitude, so x except i-th class of its correspondence be not except 0, remaining is all 0.The attitude classification of test pattern y can be obtained according to coefficient vector x.Therefore, human face posture identification problem is converted into the problem of solve linear equations.If m > is N, system of equations is overdetermination, and x has unique solution or without solution.If m < is N, x has multiple solution.In human face posture identification application, often m < N, namely needs to solve underdetermined system of equations problem, follows the trail of, the achievement in research of compressed sensing and sparse representation method shows according to atom, if the solution of the above-mentioned underdetermined system of equations is enough sparse, and just can be by the minimization problem of norm regularization is obtained:

This problem can be solved by the linearity specifications method of standard.

Namely sparse representation model coupling solves following system of linear equations:

(unobstructed) or (blocking)

Only having the coefficient vector x (for unobstructed) of linear combination or ω (to block) in above-mentioned system of equations for unknown, therefore can solve by being with the least square method of sparse constraint:

$x=\underset{x}{\min }{\left|\right|y-\mathrm{Dx}\left|\right|}_F^2+\mathrm{\&lambda;}{\left|\right|x\left|\right|}_1$

Wherein, λ is balance factor, and play Equilibrium fitting error and openness effect, the present invention gets λ=0.01.Above-mentioned minimum Optimized model the present invention adopts minimum angles to return (LeastAngleRegression, LAR) Lars-Lasso algorithm carries out solving (B.Efron, T.Hastie, I.JohnstoneandR.Tibshirani, " LeastAngleRegression ", AnnalsofStatistics, 32,407-499,2004).Similar for there being circumstance of occlusion to solve.

Step S2c: human face posture identification

Coefficient vector x according to the linear combination solved carries out human face posture Classification and Identification.In theory should only with the related intimate of the test sample book of a certain class attitude in training sample, the sign coefficient non-zero of its correspondence.Therefore, can clearly classify to this attitude to be measured.But, because noise and modeling error can cause the part of sample to be tested uncorrelated categorized representation coefficient to occur the nonzero element that numerical value is very little, bring impact to correct classification.Considering that sample to be tested closes rarefaction representation based on entirety training set at training sample set, considering that sample to be tested is carried out validity at the expression coefficient that training sample set closes to be added up accordingly, using the maximum foundation as judging classification of aggregate-value.

The coefficient vector x solved according to step S2b can obtain the attitude classification of human face posture image y to be estimated.As shown in Figure 4, test human face posture belongs to the 4th class, and therefore its non-zero coefficient mainly concentrates on x ₄in, the coefficient overwhelming majority of other classification is 0.

In order to reduce the impact of noise and modeling error; Considering that sample to be tested closes rarefaction representation based on entirety training set at training sample set simultaneously, considering that the expression coefficient validity by sample to be tested closes at training sample set adds up accordingly, using the maximum foundation as judging classification of aggregate-value.Therefore the attitude based on rarefaction representation is categorized as:

$\underset{i}{\max }{p}_i\left(y\right)=\underset{i}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_i\left(f\left({\hat{x}}_1\right)\right)$

Wherein, f () is a special function, represents and the negative factor of rarefaction representation coefficient is set to 0; be a Selection of Function, only select rarefaction representation coefficient in expression coefficient corresponding to the i-th class subspace matrices, and other to be set to 0; p _iy () is that the validity of the rarefaction representation coefficient that in training sample set, the i-th class subspace is corresponding with it adds up the factor.

Correspondence blocks human face posture process, solves human face posture image occlusion issue by increasing the method for blocking attitude face dictionary.Due to, above-mentioned human face posture recognition methods make use of rarefaction representation sorting technique principle dexterously.Therefore, to human face light, noise, expression and change resolution, there is robustness.In order to solve human face posture occlusion issue: first, setting up one again according to step S1 and blocking corrosion dictionary D _e.Then, the Gabor characteristic vector of human face posture image to be estimated is extracted according to step S2a and step S2b m is Gabor characteristic vector dimension, is regarded as by y by unobstructed dictionary D and blocks corrosion dictionary collinearity combination represents that (the dictionary learning model in corresponding step S2b) is:

$y=y_0+e_0=\mathrm{Dx}+e_0=[D,D_e]\left[\begin{array}{c}x\\ x_e\end{array}\right]=\mathrm{B\&omega;}$

Wherein, unobstructed image y ₀with block error image e ₀by dictionary D and dictionary can be blocked respectively rarefaction representation.D _efor orthogonal matrices.Finally, block human face posture identification problem and change into following optimization problem:

This problem can be solved by the linearity specifications method of standard.Namely sparse representation model coupling solves following system of linear equations:

(blocking)

The coefficient vector ω (to block) of linear combination is only had for unknown in above-mentioned system of equations.

Arriving this, blocking dictionary D by increasing _emethod successfully solve face occlusion issue.

Above-described embodiment is the present invention's preferably embodiment; but embodiments of the present invention are not restricted to the described embodiments; change, the modification done under other any does not deviate from Spirit Essence of the present invention and principle, substitute, combine, simplify; all should be the substitute mode of equivalence, be included within protection scope of the present invention.

Claims (10)

1., based on a human face posture recognition methods for Gabor characteristic and dictionary learning, it is characterized in that, comprise the following steps:

Gabor characteristic extraction is carried out to the human face posture image to be identified of online input, builds Gabor characteristic vector y;

Use the complete dictionary of attitude to carry out linear combination expression to described Gabor characteristic vector y, set up sparse representation model and solve coefficient vector, wherein Gabor characteristic vector m is Gabor characteristic vector dimension;

Coefficient vector according to the above-mentioned linear combination solved carries out human face posture Classification and Identification.

2. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 1, is characterized in that,

Described human face posture is divided into 7 different attitude classifications, is defined as that left avertence turns, left side face, right avertence turn respectively, right side face, front, comes back and nod, the subspace that its each correspondence is different.

3. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 1, is characterized in that,

The complete dictionary of attitude is used to carry out before linear combination represents to described Gabor characteristic vector y, also comprise the training step of the complete dictionary of described attitude, the complete dictionary of wherein said attitude comprise corresponding unscreened human face posture the first attitude complete dictionary D and to there being the complete dictionary D of human face posture second attitude blocked _e, described first attitude complete dictionary D and the complete dictionary D of described second attitude _etraining independently complete.

4. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 3, it is characterized in that, the training process of described first attitude complete dictionary D is specific as follows:

Collect the human face posture image pattern of each attitude classification respectively, and the human face posture Gabor characteristic training set that Gabor filtering process and feature extraction dyad form each attitude classification is carried out to described human face posture image pattern;

Use K-SVD to carry out training optimization to the human face posture Gabor characteristic training set of every class attitude classification and draw best sub-dictionary D respectively _i, i=1,2 ..., 7;

By sub-for all kinds of the best dictionary D _iform the complete dictionary D=of the first attitude [D ₁, D ₂..., D ₇].

5. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 4, it is characterized in that, the Gabor characteristic of described human face posture image is:

$S=(a_{0,0}^{\left(\mathrm{\&rho;}\right)},a_{1,0}^{\left(\mathrm{\&rho;}\right)},...,a_{7,4}^{\left(\mathrm{\&rho;}\right)})$

Wherein, by the mould to Gabor filter factor carry out ρ sampling and the column vector that obtains, μ, v are direction and the yardstick of Gabor filter, for face pose presentation and Gabor core ψ _{μ, v}convolution, Gabor core is defined as:

${\mathrm{\&psi;}}_{\mathrm{\&mu;},v}\left(z\right)=\frac{\left|\right|k_{\mathrm{\&mu;},v}|{|}^2}{{\mathrm{\&sigma;}}^2}{e}^{-\left|\right|k_{\mathrm{\&mu;},v}\left|{|}^2\right|\left|z\right|{|}^2/2{\mathrm{\&sigma;}}^2}(e^{{\mathrm{ik}}_{\mathrm{\&mu;},v}z}-e^{-{\mathrm{\&sigma;}}^2/2})$

Wherein, z (x, y) represents pixel; for small echo item, k _v=k _max/ f ^v, φ _μ=π μ/8, σ=1.5 π controls the ratio of Gauss's window width and wavelength.

6. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 3, it is characterized in that, when the human face posture image to be identified of described online input is unscreened human face posture image, extract the Gabor characteristic vector of described human face posture image to be estimated m is Gabor characteristic vector dimension, y is regarded as the complete dictionary of described first attitude linear combination represent: wherein N is the total number of dictionary atom, and described sparse representation model is

7. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 6, is characterized in that, described sparse representation model

Solve by being with the least square method of sparse constraint:

$x=\underset{x}{min}\left|\right|y-Dx|{|}_F^2+\mathrm{\&lambda;}|\left|x\right|{|}_1$

Wherein, λ is balance factor, plays Equilibrium fitting error and openness effect.

8. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 3, it is characterized in that, the human face posture image to be identified of described online input is when having the human face posture image blocked, and extracts the Gabor characteristic vector of described human face posture image to be estimated m is Gabor characteristic vector dimension, is regarded as by y by the first described attitude complete dictionary D and the complete dictionary of the second attitude collinearity combination represents:

$y=y_0+e_0=Dx+e_0=\&lsqb;D,D_e\&rsqb;\left[\begin{array}{c}x\\ x_e\end{array}\right]=B\mathrm{\&omega;},$

Wherein unobstructed image y ₀with block error image e ₀respectively can by the first attitude complete dictionary D and the complete dictionary of the second attitude rarefaction representation, D _efor orthogonal matrices, described sparse representation model is

9. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 8, is characterized in that, described sparse representation model following optimization problem is changed into by there being the human face posture problem of image recognition of blocking:

(l ¹):ω＝argmin||ω|| ₁s.t.Bω＝y，

This problem can be solved by the linearity specifications method of standard.

10. the human face posture recognition methods based on Gabor characteristic and dictionary learning according to claim 1, it is characterized in that, the described coefficient vector according to the above-mentioned linear combination solved carries out human face posture Classification and Identification and adds up specifically by carrying out validity to the coefficient vector of the described linear combination solved, using the maximum foundation as judging classification of aggregate-value, namely

$\underset{i}{max}{p}_i\left(y\right)=\underset{i}{\&Sigma;}{\mathrm{\&delta;}}_i\left(f\right({\hat{x}}_1\left)\right)$

Wherein, f () is special function, and the negative factor for the expression coefficient by sparse representation model sets to 0, selection of Function, for only selecting the expression coefficient of sparse representation model in expression coefficient corresponding to the i-th class subspace matrices, and other is represented that coefficient sets to 0, p _iy () is that the validity of the expression coefficient of subspace that in training sample set, the i-th class attitude classification the is corresponding sparse representation model corresponding with it adds up the factor.