CN103530634A - Face characteristic extraction method - Google Patents

Abstract

The invention discloses a face characteristic extraction method, which comprises the following steps of 11, inputting a face photo; 12, performing log transformation on the face photo to obtain a log-domain image; 13, decomposing the log-domain image in an EMD (empirical mode decomposition) way to obtain an EMD image; 14, extracting a high-frequency component from the EMD image as the estimate of a reflecting component to obtain a lighting-invariable face expression and an estimated lighting component. According to the face characteristic extraction method, the reflecting component and the lighting component of the log-domain image in a log space form a linear superposition relationship by performing log space transformation on the face photo, and the log-domain image is decomposed in the EMD way to extract the lighting-invariable face expression from the log-domain image, so that the lighting component can be effectively distinguished from the lighting component, and a better effect can be achieved.

Description

Face feature extraction method

Technical field

The present invention relates to face recognition technology field, relate in particular to the face feature extraction method based on empirical mode decomposition.

Background technology

Face recognition technology based on human face photo is all widely used in fields such as public safety, digital art, game.Illumination variation problem is the key factor that affects recognition of face performance always.In order to overcome the impact of illumination variation on face recognition technology, need to from the human face photo of input, obtain the illumination invariant of people's face and express.There are some researches show: illumination variation major effect is to the low-frequency component of photo, and less on radio-frequency component impact.Therefore, existing main stream approach is to carry out spectrum analysis by comparison film, extracts the radio-frequency component of photo and expresses as illumination invariant.In existing algorithm, the main spectral analysis algorithm adopting comprises: wavelet transformation is (as document T.Zhang, B.Fang, Y.Yuan etal., Multiscale facial structure representation for face recognition undervarying illumination, Pattern Recognition, 42 (2009) 251-258. and document C.Garcia, G.Zikos, and G.Tziritas, A wavelet-based framework for facerecognition, in Proc.Workshop on Advances in Facial Image Analysis andRecognition Technology, ECCV, Freiburg, 1998, pp.84-92.), Gabor conversion is (as document K.Okada, J.Steffens, T.Maurer et al., The Bochum/USC facerecognition system, Face Recognition:From Theory to Applications, Springer, Berlin, 1998, pp.186-205.), Weighted Gauss filtering is (as document H.Wang, S.Z.Li and Y.Wang, Face Recognition under Varying Lighting Conditions usingSelf Quotient Image, in Proc.Conf.Automatic Face and GestureRecognition, Seoul, 2004, pp.819-824.), discrete cosine transform is (as document Z.Hafed, and M.Levine, Face recognition using the discrete cosine transform, International Journal of Computer Vision, 43 (2001) 167-188.), total variation model is (as document T.Chen, W.Yin, X.Zhou, D.Comaniciu, T.S.Huang, Totalvariation models for variable lighting face recognition, IEEE Trans.PatternAnal.Mach.Intel., 28 (2006) 1519 – 1524.), profile wave convert is (as document X.Xie, J.Lai, W.-S.Zheng, Extraction of Illumination Invariant Facial Features froma Single Image Using Nonsubsampled Contourlet Transform, PatternRecognition, 43 (2010) 4177-4189.) and Fourier transform (as document J.Lai, P.C.Yuen, and G.Feng, Face recognition using holistic Fourier invariantfeatures, Pattern Recognition, 34 (2001) 95-109.).

The main thought of above-mentioned spectral analysis algorithm is to be the linear combination of (calling " base ") of one group of baseband signal picture signal decomposition.But these methods " base " used are manually and design in advance, and want decomposed signal irrelevant.

For addressing this problem, there is the method for researching and proposing field experience Mode Decomposition to decompose (as document D.Zhang and Y.Y.Tang facial image, Extraction ofIllumination-Invariant Features in Face Recognition by Empirical ModeDecomposition, International Conference on Biometrics (ICB), 2009., document D.Zhang, J.Pan, Y.Y.Tang, C.Wang, Illumination invariant facerecognition based on the new phase features, IEEE InternationalConference on Systems Man and Cybernetics, 2010, pp.3909-3914. with document Wang Ke, Dang Deyu, Sun Bin. a kind of face identification method based on Bidimensional Empirical Mode Decomposition. software guide, 2010, 09 (2)).Empirical mode decomposition proposed (N.E.Huang in 1998 by people such as N.E.Huang, Z.Shen, S.R.Long, M.C.Wu, H.H.Shih, Q.Zheng, N.C.Yen, C.C.Tung, and H.H.Liu, The empirical mode Decomposition and theHilbert Spectrum for Nonlinear and nonstationary time series analysis, inProceedings of the Royal Society London A, 1998, pp.903-1005.), English Empirical Mode Decomposition by name, is called for short EMD.EMD is a kind of adaptive Algorithm of Signal Decomposition, can according to signal adaptive to be decomposed calculate base and then carry out signal decomposition, on signal analysis and signal are processed, there is larger application potential quality.But the method for existing field experience Mode Decomposition is all directly EMD to be acted on to original human face photo, original human face photo be decomposed into a plurality of subsignals and, and then select parton signal and identify as face characteristic.Because illumination imaging model is not simple signal stack, therefore, the effect of the method for existing field experience Mode Decomposition is also more limited.

Summary of the invention

For the problems referred to above, the object of this invention is to provide a kind of face feature extraction method solving the problems of the technologies described above.

, it comprises the steps:

S11, input human face photo;

S12, described human face photo is carried out to log-transformation, obtain log-domain image;

S13, field experience Mode Decomposition decompose described log-domain image, obtain empirical mode decomposition image;

S14, extract the radio-frequency component in described empirical mode decomposition image, as the estimation of reflex components, can obtain the illumination composition that people's face of illumination invariant is expressed and estimated to obtain.

In a preferred embodiment of the present invention, in described step S11, according to Lambertian reflection model, obtain the expression formula of described human face photo:

I(x,y)=R(x,y)L(x,y)，

Wherein, R is reflex components, is that a kind of people's face of illumination invariant is expressed, and L is illumination composition.

In a preferred embodiment of the present invention, in described step S12, the expression formula of described human face photo is carried out to log-transformation, obtain the expression formula of described log-domain image: f=v+u, wherein, v and u are respectively R and L in the value of log-domain.

In a preferred embodiment of the present invention, in described step S13, field experience Mode Decomposition decomposes the expression formula of described log-domain image, obtains the expression formula of described empirical mode decomposition image:

wherein, d _kto decompose the subsignal obtaining, along with the increase of k, d _kfrom high frequency to low frequency variations, r decomposes residual error gradually.

In a preferred embodiment of the present invention, in described step S14, the expression formula of described empirical mode decomposition image is carried out to the extraction of radio-frequency component, obtains:

People's face of illumination invariant is expressed:

Estimate the illumination composition obtaining:

Wherein, $\hat{v}\≈{\mathrm{\Σ}}_{k=1}^{K_0}{d}_k,\hat{u}=f-\widehat{v,}$ K ₀=1 or K ₀=2.

Compared to prior art, described face feature extraction method provided by the invention first carries out log space conversion to human face photo, obtain log-domain image, and make interior in log space of log-domain image form linear superposition relation at composition (reflex components and illumination composition), then use EMD to decompose log-domain image, therefrom extract people's face of illumination invariant and express.Because EMD is decomposed into input signal (i.e. the human face photo of input) linear superposition of some subsignals (being reflex components and illumination composition) just, so described face feature extraction method can be distinguished illumination composition and illumination invariant composition effectively, obtains preferably effect.

Above-mentioned explanation is only the general introduction of technical solution of the present invention, in order to better understand technological means of the present invention, and can be implemented according to the content of instructions, and for above and other objects of the present invention, feature and advantage can be become apparent, below especially exemplified by embodiment, and coordinate accompanying drawing, be described in detail as follows.

Accompanying drawing explanation

The process flow diagram of the face feature extraction method that Fig. 1 provides for a preferred embodiment of the present invention;

Fig. 2 is for to carry out the corresponding experimenter's performance curve of recognition of face figure based on method not of the same race;

Fig. 3 utilizes face feature extraction method shown in Fig. 1 to carry out the effect schematic diagram of face characteristic extraction.

Embodiment

Below in conjunction with drawings and the specific embodiments, the present invention is further detailed explanation.

Refer to Fig. 1, a preferred embodiment of the present invention provides a kind of face feature extraction method, and it comprises the following steps:

S11, input human face photo I.

In the present embodiment, according to Lambertian reflection model, obtain the expression formula of described human face photo:

I(x,y)=R(x,y)L(x,y) (1)

Wherein, R is reflex components, and L is illumination composition.Usually, R is a kind of people's face expression of illumination invariant.In the present embodiment, need to obtain the estimated value of R and L.

S12, described human face photo I is carried out to log-transformation, obtain log-domain image.

Expression formula (1) is carried out to log-transformation, thus, described human face photo I can be converted into log space (log-domain), obtain the expression formula of log-domain image:

f=logI

=logR+logL

Δ

=v+u，(2)

Wherein, v and u are respectively R and L in the value of log-domain, i.e. v=logR, u=logL.

Be understandable that, in log-domain image f=v+u, its inherent composition v and u form linear superposition relation, and reflex components and illumination composition form linear superposition relation.

S13, field experience Mode Decomposition (EMD) decompose described log-domain image f, obtain empirical mode decomposition image.

In the present embodiment, expression formula (2) is carried out to EMD decomposition, the expression formula that obtains EMD image is:

$f={\mathrm{\Σ}}_{k=1}^k{d}_k+r---\left(3\right)$

Wherein, d _kbe to decompose the subsignal obtaining, be also called intrinsic mode function, along with the increase of k, d _kgradually from high frequency to low frequency variations; R decomposes residual error.

In the present embodiment, concrete EMD algorithm can list of references N.E.Huang, Z.Shen, S.R.Long, M.C.Wu, H.H.Shih, Q.Zheng, N.C.Yen, C.C.Tung, and H.H.Liu, The empirical mode Decomposition and the Hilbert Spectrum forNonlinear and nonstationary time series analysis, in Proceedings of theRoyal Society London A, 1998, pp.903-1005., repeat no more herein.

S14, extract the radio-frequency component in described empirical mode decomposition image, as the estimation of reflex components, can obtain the illumination composition that people's face of illumination invariant is expressed and estimated to obtain.

In the present embodiment, the expression formula of described empirical mode decomposition image is carried out to the extraction of radio-frequency component, obtains:

People's face of illumination invariant is expressed:

$\hat{R}=\exp \left(\hat{v}\right)---\left(4\right)$

Estimate the illumination composition obtaining:

$\hat{L}=\exp \left(\hat{u}\right)---\left(5\right)$

Wherein, $\hat{v}\≈{\mathrm{\Σ}}_{k=1}^{K_0}{d}_k,\hat{u}=f-\hat{u}$ K ₀=1 or K ₀=2.

Be understandable that,

people's face of the illumination invariant that obtained is expressed,

estimate the illumination composition obtaining.

Described face feature extraction method first carries out log space conversion to human face photo, obtain log-domain image, and make interior in log space of log-domain image form linear superposition relation at composition (reflex components and illumination composition), then use EMD to decompose log-domain image, therefrom extract people's face of illumination invariant and express.Because EMD is decomposed into input signal (i.e. the human face photo of input) linear superposition of some subsignals (being reflex components and illumination composition) just, so described face feature extraction method can be distinguished illumination composition and illumination invariant composition effectively, obtains preferably effect.

The face feature extraction method proposing for the present invention, the present inventor has carried out recognition of face test on people's face test library Extended Yale B of Yale University's issue, refer to Fig. 2, Fig. 2 is for to carry out the corresponding experimenter's performance curve of recognition of face figure (receiver operating characteristic curve based on method not of the same race, be called for short ROC curve), as shown in Figure 2, the inventive method can obtain preferably ROC curve (obtaining higher checking rate under identical false acceptance rate), utilize described face feature extraction method provided by the invention to carry out face characteristic extraction, the wavelet transformation of mentioning in obtained recognition of face Performance Ratio background technology, Gabor conversion, Weighted Gauss filtering, discrete cosine transform, total variation model, the obtained performance of profile wave convert and original EMD method is all obviously better.

Referring to Fig. 3, is the effect schematic diagram that the present invention utilizes described face feature extraction method to carry out face characteristic extraction, and wherein, Fig. 3 a is input photo, the illumination composition of Fig. 3 b for estimating to obtain

the illumination invariant expression of Fig. 3 c for estimating to obtain

known, utilize described face feature extraction method to carry out feature extraction to human face photo and can obtain preferably effect.

The above, only embodiments of the invention, not the present invention is done to any pro forma restriction, although the present invention discloses as above with embodiment, yet not in order to limit the present invention, any those skilled in the art, do not departing within the scope of technical solution of the present invention, when can utilizing the technology contents of above-mentioned announcement to make a little change or being modified to the equivalent embodiment of equivalent variations, in every case be not depart from technical solution of the present invention content, any simple modification of above embodiment being done according to technical spirit of the present invention, equivalent variations and modification, all still belong in the scope of technical solution of the present invention.

Claims (5)

1. a face feature extraction method, is characterized in that, described face feature extraction method comprises the steps:

S11, input human face photo;

S12, described human face photo is carried out to log-transformation, obtain log-domain image;

S13, field experience Mode Decomposition decompose described log-domain image, obtain empirical mode decomposition image;

S14, extract the radio-frequency component in described empirical mode decomposition image, as the estimation of reflex components, can obtain the illumination composition that people's face of illumination invariant is expressed and estimated to obtain.

2. face feature extraction method as claimed in claim 1, is characterized in that, in described step S11, obtains the expression formula of described human face photo according to Lambertian reflection model:

I(x,y)=R(x,y)L(x,y)，

Wherein, R is reflex components, is that a kind of people's face of illumination invariant is expressed, and L is illumination composition.

3. face feature extraction method as claimed in claim 2, is characterized in that, in described step S12, the expression formula of described human face photo is carried out to log-transformation, obtain the expression formula of described log-domain image: f=v+u, wherein, v and u are respectively R and L in the value of log-domain.

4. face feature extraction method as claimed in claim 3, is characterized in that, in described step S13, field experience Mode Decomposition decomposes the expression formula of described log-domain image, obtains the expression formula of described empirical mode decomposition image: wherein, d _kto decompose the subsignal obtaining, along with the increase of k, d _kfrom high frequency to low frequency variations, r decomposes residual error gradually.

5. face feature extraction method as claimed in claim 4, is characterized in that, in described step S14, the expression formula of described empirical mode decomposition image is carried out to the extraction of radio-frequency component, obtains:

People's face of illumination invariant is expressed:

Estimate the illumination composition obtaining:

Wherein, $\hat{v}\≈{\mathrm{\Σ}}_{k=1}^{K_0}{d}_k,\hat{u}=f-\widehat{v,}$ K ₀=1 or K ₀=2.

Images