CN104933414A - Living body face detection method based on WLD-TOP (Weber Local Descriptor-Three Orthogonal Planes) - Google Patents

Abstract

The invention discloses a living body face detection method based on WLD-TOP (Weber Local Descriptor-Three Orthogonal Planes). The method comprises the following steps: (1) training stage: reading a training set video, performing face region detection on each frame, converting the frames into a gray level facial image frame sequence to construct a three-dimensional image matrix, constructing a filtering template, calculating WLD features, generating a WLD-TOP feature vector, and inputting the feature vector into an SVM classifier for training to establish an SVM model; and (2) testing stage: for an image sequence under test, performing face detection on each frame, converting the frames into a gray level facial image sequence, constructing a three-dimensional image matrix and a filtering template, calculating WLD features, generating a WLD-TO feature vector, and feeding the WLD-TOP feature vector into a trained SVM model to obtain a living body face detection result. The Weber law is adopted on the basis of the LBP-TOP, so that the size relationships between neighborhood pixels and a center pixel are reflected, and the differences between the neighborhood pixels and the center pixel are quantified. Thus, the features of a descriptor are more complete.

Description

A kind of living body faces detection method based on WLD-TOP

Technical field

The present invention relates to the research field of Face datection, particularly a kind of living body faces detection method based on WLD-TOP.

Background technology

Face recognition technology by comparing and analyze the biological characteristic of face, thus differentiates the identity of people.Face recognition technology in the past few decades between achieve significant progress, the product of recognition of face is applied to multiple occasions such as gate inhibition, important place monitoring, entry and exit.An advantage of face recognition technology automatically identifies target, without the need to supervision, but also leave safe hidden danger, if lawless person use the photo of user even video can to out-trick easily face identification system, will work the mischief, social safety and stability has been arrived in serious threat.

Common face spoofing attack comprises photo and attacks and video attack.Photo is attacked with the face characteristic of user, and video is attacked more with the behavioral characteristics of validated user, as the change of nictation and facial expression, has more duplicity, seriously have impact on the accuracy of face identification system differentiation.

Present living body faces detection method mainly contains following several: one is the method analyzed based on texture structure, and the method, by the otherness of analyzing three-dimensional living body faces with face imaging of retaking, is extracted associated texture feature and differentiated; Two is the methods based on facial movement information analysis, and the essential distinction of living body faces and face of retaking is that the former is three-dimensional body, and the latter is two-dimension plane structure, and there is the secondary shooting of face, the movement effects that they produce is diverse; Three is the methods based on living body characteristics information analysis, and the method analyzes the living body characteristics such as the thermal infrared images of face, nictation and lip motion, the checkout equipment support that this method may need some extra, on promoting, therefore there is the restriction of hardware.

The realization of above-mentioned three kinds of methods all will use suitable iamge description, and it can improve the accuracy rate that living body faces detects greatly.Because face spoofing attack means get more and more, especially based on the fraud of video, there is the behavioral characteristics of living body faces, such as can be obtained the change of the facial expressions such as nictation by the dynamic video of validated user thus reach the object of spoofing attack, so we need the descriptor of an energy joining day and spatial information to distinguish.

Summary of the invention

Fundamental purpose of the present invention is that the shortcoming overcoming prior art is with not enough, a kind of living body faces detection method based on WLD-TOP is provided, by extracting WLD descriptor, and add the timeline information of frame of video, thus form WLD-TOP (Weber Local Descriptor-Three Orthogonal Planes) descriptor, it has merged the space characteristics of WLD descriptor and the temporal characteristics of frame of video, improves the accuracy rate that living body faces detects.

In order to achieve the above object, the present invention is by the following technical solutions:

Based on a living body faces detection method for WLD-TOP descriptor, comprise the steps:

S1, training stage: read training set video, human face region detection is carried out to each frame, and convert Gray Face sequence of image frames to, structure three-dimensional image matrix, then construct Filtering Template and calculate WLD feature, regeneration WLD-TOP proper vector, finally inputs SVM classifier by proper vector and trains, thus sets up SVM model;

S2, test phase: for the image sequence of test, Face datection is carried out to each frame and is converted to Gray Face image sequence, then three-dimensional image matrix and Filtering Template is constructed, calculate WLD feature, generate WLD-TOP proper vector, finally send into the SVM model trained, draw living body faces testing result.

Preferably, in step S1, described training set video is living body faces video, the photo face of recording, Replay Attack or print picture and attack.

Preferably, in step S1, after reading in frame of video, extract haar feature and carry out human face region detection with adaboosting algorithm, extract colored human face figure wherein and transfer the consistent gray-scale map of size dimension to.

Preferably, in step S1, the method for described structure three-dimensional image matrix is:

The frame of video length R once read is set, chooses the boundary threshold L of T coordinate _t, then the actual frame of video length R as center pixel process _t=R-2L _t, then the gray-scale value of this group frame of video is read in one containing in the three-dimensional matrice I (x, y, t) of X, Y and T coordinate.

Preferably, in step S1, structure Filtering Template the method calculating WLD feature are:

Choose X, the abundant value L in Y-coordinate border respectively _x, L _y, determine the glide filter template length p of WLD descriptor, the Filtering Template forming p*p, to three of I orthogonal planes XY, XT and YT, utilizes the difference excitation ξ and direction gradient Φ of each central pixel point after the abundant value in p*p formwork calculation removing border of WLD method respectively _t, computing method are as follows: first suppose that the central point calculated is x _c, its eight consecutive point are x respectively _i, i=0 ..., p ²-1, $v_0=\underset{i=0}{\overset{p^2-1}{\Σ}}\left({\mathrm{\Δx}}_i\right)=\underset{i=0}{\overset{p^2-1}{\Σ}}(x_i-x_c),$ Definition ${\mathrm{\ξ}}^{\′}\left(x_c\right)=arctan\[\frac{v_0}{x_c}\]=arctan\[\underset{i=0}{\overset{p^2-1}{\Σ}}\left(\frac{x_i-x_c}{x_c}\right)\],$ Make v ₁=x ₅-x ₁, v ₂=x ₇-x ₃, get definition

then wherein θ ' ∈ [0,2 π), S is the dimension of direction gradient feature, then Φ _t=0,1 ..., S-1, by above-mentioned steps obtain WLD descriptor ξ ' (x _c), Φ _t.

Preferably, in step S1,

WLD-TOP computation process is: first to ξ ' (x _c) do following normalization:

therefore ξ (x _c) value is 0 to this N number of round values of N-1; Φ _tnormalize to S the direction represented to S-1 with integer 0, for XY plane, to WLD{ ξ (x _c), Φ _ttwo-dimensional histogram carries out dimensionality reduction, fixing Φ _t, ask corresponding ξ (x _c) sub-histogram, according to the Φ of S dimension _tbe divided into the sub-histogram of S group, according to Φ _torder from small to large connects this S sub-histogram successively, definition

F (x, y)=N × Φ _t+ ξ (x _c), then the histogram of XY plane

H _{i, XY=}Σ _x,ym{f (x, y)=i}, i=0,1 ..., N Φ _t-1, wherein $M\left(A\right)=\left\{\begin{array}{c}0,ifAisfa1se;\\ 1,ifAistrue.\end{array}\right.,$ Thus form N Φ _tthe WLD histogram H of dimension _xY, then obtain three orthogonal planes by this method

The histogram h of (n=0:XY, n=1:XT, n=2:YT) _i,n=Σ _{x, y, t}m{f (x, y, t)=i}, i=0,1 ..., N Φ _t-1, they are transferred to N Φ _tthe row vector H of dimension _n, connect successively, generate 3N Φ _tthe WLD-TOP feature row vector H of dimension _wT=[H ₀h ₁h ₂].

Preferably, in step S1,

Described SVM classifier uses the SVM implementation tool based on LIBSVM; All k the proper vector composing training collection eigenmatrixes that training set is obtained $H_{train}=\left(\begin{array}{c}{H}_{WT,train,1}\\ .\\ .\\ .\\ H_{WT,train,k}\end{array}\right),$ Adopt SVM training, and with the model of training out to the test set eigenmatrix containing j proper vector $H_{test}=\left(\begin{array}{c}{H}_{WT,test,1}\\ .\\ .\\ .\\ H_{WT,test,j}\end{array}\right)$ Whether classification is the differentiation label of living body faces, and with true tag comparison, thus obtain the accuracy rate that In vivo detection differentiates.

Preferably, in step S1,

Described SVM classifier, for the training sample set of input, adopts cross validation method to train, and utilizes trellis search method to find the optimized parameter collection { C, γ } of SVM.

Compared with prior art, tool has the following advantages and beneficial effect in the present invention:

1, the WLD-TOP method of the present invention's proposition, utilize Weber('s)law, on LBP-TOP basis, not only embody the magnitude relationship of neighborhood territory pixel and center pixel, also quantize the difference of neighborhood territory pixel and center pixel, and using this species diversity as a kind of feature, then bonding position Gradient Features, make the feature of descriptor more comprehensive.

2, WLD descriptor is expanded to three dimensions by the present invention, adds timeline information, Time and place information fusion is integrated, and attacks improve detection accuracy to the video with behavioral characteristics.

3, this invention simplifies traditional WLD calculated direction gradient and histogrammic mathematical procedure, computing method of the present invention only change the order of proper vector element in classic method, on the basis not changing traditional WLD proper vector element size, construct these features with more intuitive mathematic(al) representation.

4, the present invention does training and testing on different pieces of information collection, detects data set by constructing SYSU living body faces, and and CASIA data set combine and done the experiment of cross datasets, improve the Generalization Capability of WLD-TOP.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 (a)-Fig. 2 (c) is the effect exemplary plot that WLD-TOP of the present invention describes face;

Fig. 3 is the schematic diagram of WLD difference of the present invention excitation and direction gradient;

Fig. 4 is the histogram of WLD-TOP descriptor of the present invention;

Fig. 5 (a)-Fig. 5 (d) is the CASIA human face region figure of the present invention for training and testing;

Fig. 6 (a)-Fig. 6 (d) is the SYSU human face region figure of the present invention for training and testing.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited thereto.

Embodiment

As shown in Figure 1, the present invention is based on the living body faces detection method of WLD-TOP, comprise the following steps:

(1) training stage: read training set video, human face region detection is carried out to each frame, and convert Gray Face sequence of image frames to, structure three-dimensional image matrix, then construct Filtering Template and calculate WLD feature, regeneration WLD-TOP proper vector, finally inputs SVM classifier by proper vector and trains, thus sets up SVM model;

The particular content in (1) stage is:

(1.1) training set video is read: read training set video, it may be living body faces video, also may be the photo face, Replay Attack, the attack of printing picture etc. recorded, we read in frame of video, then extract haar feature and carry out human face region detection with adaboosting algorithm, extract colored human face figure wherein and transfer the consistent gray-scale map of size dimension to;

(1.2) three-dimensional image matrix is constructed: the frame of video length R once read is set, chooses the boundary threshold L of T coordinate _t, then the actual frame of video length R as center pixel process _t=R-2L _t, then the gray-scale value of this group frame of video is read in one containing in the three-dimensional matrice I (x, y, t) of X, Y and T coordinate.

(1.3) WLD feature is calculated: choose X, the abundant value L in Y-coordinate border respectively _x, L _y, determine the glide filter template length p of WLD descriptor, form Filtering Template (p=3, the L in experiment of p*p _x=L _y=L _t=1); To three orthogonal planes XY, XT and YT of I, utilize the difference excitation ξ (x of each central pixel point after the abundant value in p*p formwork calculation removing border of WLD method respectively _c) and direction gradient Φ _t, computing method are as follows: first suppose that the central point calculated is x _c, its eight consecutive point are x respectively _i, i=0 ..., p ²-1, $v_0=\underset{i=0}{\overset{p^2-1}{\Σ}}\left({\mathrm{\Δx}}_i\right)=\underset{i=0}{\overset{p^2-1}{\Σ}}(x_i-x_c),$ Definition ${\mathrm{\ξ}}^{\′}\left(x_c\right)=arctan\[\frac{v_0}{x_c}\]=arctan\[\underset{i=0}{\overset{p^2-1}{\Σ}}\left(\frac{x_i-x_c}{x_c}\right)\],$ Make v ₁=x ₅-x ₁, v ₂=x ₇-x ₃, get $\mathrm{\θ}\left(x_c\right)=arctan\left(\frac{v_2}{v_1}\right),$ Definition then wherein θ ' ∈ [0,2 π), S is the dimension of direction gradient feature, then Φ _t=0,1 ..., S-1, by above-mentioned steps obtain WLD descriptor ξ ' (x _c), Φ _t, as shown in Figure 3,

(1.4) WLD-TOP proper vector is generated: first to ξ ' (x _c) do following normalization:

therefore ξ (x _c) value is 0 to this N number of round values of N-1; Φ _tnormalize to S the direction represented to S-1 with integer 0, for XY plane, to WLD{ ξ (x _c), Φ _ttwo-dimensional histogram carries out dimensionality reduction, fixing Φ _t, ask corresponding ξ (x _c) sub-histogram, according to the Φ of S dimension _tbe divided into the sub-histogram of S group, according to Φ _torder from small to large connects this S sub-histogram successively, definition

F (x, y)=N × Φ _t+ ξ (x _c), then the histogram of XY plane

H _{i, XY}=Σ _x,ym{f (x, y)=i}, i=0,1 ..., N Φ _t-1, wherein $M\left(A\right)=\left\{\begin{array}{c}0,ifAisfa1se;\\ 1,ifAistrue.\end{array}\right.,$ Thus form N Φ _tthe WLD histogram H of dimension _xY, then obtain three orthogonal planes by this method

The histogram h of (n=0:XY, n=1:XT, n=2:YT) _i,n=Σ _{x, y, t}m{f (x, y, t)=i}, i=0,1 ..., N Φ _t-1, they are transferred to N Φ _tthe row vector H of dimension _n, connect successively, generate 3N Φ _tthe WLD-TOP feature row vector H of dimension _wT=[H ₀h ₁h ₂], as shown in Figure 4.

(1.5) proper vector is inputted SVM classifier to train: SVM classifier uses the SVM implementation tool based on LIBSVM; All k the proper vector composing training collection eigenmatrixes that training set is obtained $H_{train}=\left(\begin{array}{c}{H}_{WT,train,1}\\ .\\ .\\ .\\ H_{WT,train,k}\end{array}\right),$ Adopt SVM training, and with the model of training out to the test set eigenmatrix containing j proper vector $H_{test}=\left(\begin{array}{c}{H}_{WT,test,1}\\ .\\ .\\ .\\ H_{WT,test,j}\end{array}\right)$ Whether classification is the differentiation label of living body faces, and with true tag comparison, thus obtain the accuracy rate that In vivo detection differentiates.

(1.6) SVM model is set up: SVM classifier, for the training sample set of input, adopts cross validation method to train, and utilizes trellis search method to find the optimized parameter collection { C, γ } of SVM.

(2) test phase: for the image sequence of test, Face datection is carried out to each frame and is converted to Gray Face image sequence, then three-dimensional image matrix and Filtering Template is constructed, calculate WLD feature, generate WLD-TOP proper vector, finally send into the SVM model trained, draw living body faces testing result.

Structure three-dimensional image matrix in above-mentioned steps (2), calculate WLD feature, generate WLD-TOP proper vector and proper vector inputted Methods and steps (1) that SVM classifier carries out training in method identical.

The present invention is described effect of the present invention by following experiment: (1) experimental data collection selects CASIA living body faces to detect data set (as shown in Fig. 5 (a)-Fig. 5 (d), Fig. 5 (a), Fig. 5 (b) is respectively live body and the non-living body picture of training set, Fig. 5 (c), Fig. 5 (d) is respectively live body and the non-living body picture of test set) and the homemade SYSU living body faces of applicant detect data set (as shown in Fig. 6 (a)-Fig. 6 (d), Fig. 6 (a), Fig. 6 (b) is respectively live body and the non-living body picture of training set, Fig. 6 (c), Fig. 6 (d) is respectively live body and the non-living body picture of test set), the CASIA dataset acquisition face video of 50 users, wherein 20 are used as training set, 30 are used as test set, each user recorded live body, shearing photo is attacked, manual bending photo is attacked, video attacks the video of Four types, every type has 3 sections of videos, be divided into low, in, high-resolution human face video, namely unique user has 3 sections of living body faces videos, 9 sections of personation face video, the SYSU dataset acquisition face video of 29 users, wherein 20 are used as training, 9 are used as test, attack pattern is the secondary face video that iphone mobile phone and ipad flat board are recorded, each user has 3 photographed scenes, corresponding each scene, unique user always has 3 sections of living body faces videos and 6 sections of personation face video.

CASIA data set and SYSU data set Internal Experiment: the video of experimental data collection is extracted human face region by every frame, and the face gray scale picture of (62 × 62 pixel) in the same size is become as normalized, the glide filter template length p=3 of WLD descriptor, boundary threshold is set to L _x=L _y=L _t=1; Reading each user's length in training set is R _tthe face video frame of=3, as a three-dimensional image matrix, the design sketch of image as shown in Fig. 2 (a)-Fig. 2 (c), then calculates WLD{ ξ (x to each orthogonal plane XY, XT, YT according to method shown in Fig. 3 successively _c), Φ _t, the dimension N=256 of difference excitation, the number S=8 of direction gradient, adopt process flow diagram as shown in Figure 4 to calculate WLD two-dimensional histogram; Different Φ is pressed to the WLD two-dimensional histogram of each orthogonal plane _treduce to 8 sub-histograms of one dimension, they are pressed Φ _tsize order is stitched together and obtains the WLD one dimensional histograms of this pixel, this histogram contains 2048 dimensional feature row vectors, then splices according to the order of XY, XT, YT plane the histogram that these one dimensional histograms obtain 6144 dimension WLD-TOP feature row vectors of this point successively; By that analogy, the face frame under the different video of each user of training set is trained, obtain the eigenmatrix H of a training set _train, the different user of test set is also obtained eigenmatrix H by (1)-(2) process _test; To H _traincarry out with LIBSVM software package the training of svmtrain function, obtain training pattern model, utilize model to carry out SVM prediction and draw differentiation accuracy rate,-the t of adjustment svmtrain function,-c, the parameters such as-g, find the optimized parameter making accuracy rate the highest, draw the final differentiation accuracy rate accuracy of WLD-TOP on CASIA (or SYSU) data set.

CASIA data set and SYSU data set cross-over experiment: train with CASIA data set in the application, test with SUSU, and then train with SYSU data set, test with CASIA.The video of two data sets is extracted human face region by every frame, and becomes the face gray scale picture of (62 × 62 pixel) in the same size as normalized, the glide filter template length p=3 of WLD descriptor, boundary threshold is set to L _x=L _y=L _t=1; Reading each user's length in training set is R _tthe face video frame of=3, as a three-dimensional image matrix, then calculates WLD{ ξ (x to each orthogonal plane XY, XT, YT according to method shown in Fig. 3 successively _c), Φ _t, the dimension N=256 of difference excitation, the number S=8 of direction gradient, adopt process flow diagram as shown in Figure 4 to calculate WLD two-dimensional histogram; Different Φ is pressed to the WLD two-dimensional histogram of each orthogonal plane _treduce to 8 sub-histograms of one dimension, they are pressed Φ _tsize order is stitched together and obtains the WLD one dimensional histograms of this pixel, then splices according to the order of XY, XT, YT plane the histogram that these one dimensional histograms obtain 6144 dimension WLD-TOP feature row vectors of this point successively; By that analogy, the face frame under the different video of each user of training set is trained, obtain the eigenmatrix H of a training set _train, the different user of test set is also obtained eigenmatrix H by (1)-(2) process _test; To H _traincarry out with LIBSVM software package the training of svmtrain function, obtain training pattern model, utilize model to carry out SVM prediction and draw differentiation accuracy rate,-the t of adjustment svmtrain function,-c, the parameters such as-g, find the optimized parameter making accuracy rate the highest, draw the final differentiation accuracy rate accuracy of WLD-TOP on cross datasets.

In experiment, the application also compares with WLD, LBPTOP and LBP tri-kinds of methods.Concrete comparative result is as shown in table 1-table 4:

The experiment of table 1:CASIA data set

Method	Accuracy rate %
		WLD-top	90.78
LBP-top	79.06
		WLD	87.08
LBP	78.35

The experiment of table 2:SYSU data set

Method	Accuracy rate %
		WLD-top	93.44
LBP-top	81.97
		WLD	85.27
LBP	81.42

Table 3 CASIA data set and SYSU data set cross-over experiment

Method	Accuracy rate %
		WLD-top	74.62
LBP-top	53.96
		WLD	64.81
LBP	53.85

The each descriptor of table 4 differentiates (R averaging time of a video _t=3)

Method	Time
		WLD-top	0.29s
LBP-top	0.20s
		WLD	0.22s
LBP	0.16s

From the result of experiment, the effect of WLD-TOP descriptor In vivo detection of the present invention is adopted to be better than existing LBP-TOP and WLD descriptor, the present invention not only has a certain upgrade in accuracy rate, has improve significantly in actual, complicated video human face Attack Scenarios than the accuracy rate based on the descriptor of LBP-TOP and WLD.Because WLD-TOP calculates WLD feature in three planes, so speed can be slightly slower than other two kinds of methods.

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 (8)

1., based on a living body faces detection method for WLD-TOP descriptor, it is characterized in that, comprise the steps:

S1, training stage: read training set video, human face region detection is carried out to each frame, and convert Gray Face sequence of image frames to, structure three-dimensional image matrix, then construct Filtering Template and calculate WLD feature, regeneration WLD-TOP proper vector, finally inputs SVM classifier by proper vector and trains, thus sets up SVM model;

S2, test phase: for the image sequence of test, Face datection is carried out to each frame and is converted to Gray Face image sequence, then three-dimensional image matrix and Filtering Template is constructed, calculate WLD feature, generate WLD-TOP proper vector, finally send into the SVM model trained, draw living body faces testing result.

2. the living body faces detection method based on WLD-TOP descriptor according to claim 1, is characterized in that, in step S1, described training set video is living body faces video, the photo face of recording, Replay Attack or print picture and attack.

3. the living body faces detection method based on WLD-TOP descriptor according to claim 1, it is characterized in that, in step S1, after reading in frame of video, extract haar feature and carry out human face region detection with adaboosting algorithm, extract colored human face figure wherein and transfer the consistent gray-scale map of size dimension to.

4. the living body faces detection method based on WLD-TOP descriptor according to claim 1, is characterized in that, in step S1, the method for described structure three-dimensional image matrix is:

The frame of video length R once read is set, chooses the boundary threshold L of T coordinate _t, then the actual frame of video length R as center pixel process _t=R-2L _t, then the gray-scale value of this group frame of video is read in one containing in the three-dimensional matrice I (x, y, t) of X, Y and T coordinate.

5. the living body faces detection method based on WLD-TOP descriptor according to right 1, is characterized in that, in step S1, structure Filtering Template the method calculating WLD feature are:

Choose X, the abundant value L in Y-coordinate border respectively _x, L _y, determine the glide filter template length p of WLD descriptor, the Filtering Template forming p*p, to three of I orthogonal planes XY, XT and YT, utilizes the difference excitation ξ and direction gradient Φ of each central pixel point after the abundant value in p*p formwork calculation removing border of WLD method respectively _t, computing method are as follows: first suppose that the central point calculated is x _c, its eight consecutive point are x respectively _i, i=0 ..., p ²-1, $v_0=\underset{i=0}{\overset{p^2-1}{\Σ}}\left({\mathrm{\Δx}}_i\right)=\underset{i=0}{\overset{p^2-1}{\Σ}}(x_i-x_c),$ Definition ${\mathrm{\ξ}}^{\′}\left(x_c\right)=\arctan \[\frac{v_0}{x_c}\]=arctan\[\underset{i=0}{\overset{p^2-1}{\Σ}}\left(\frac{x_i-x_c}{x_c}\right)\],$ Make v ₁=x ₅-x ₁, v ₂=x ₇-x ₃, get definition order wherein θ ' ∈ [0,2 π), S is the dimension of direction gradient feature, then Φ _t=0,1 ..., S-1, by above-mentioned steps obtain WLD descriptor ξ ' (x _c), Φ _t.

6. the living body faces detection method based on WLD-TOP descriptor according to right 5, is characterized in that, in step S1,

WLD-TOP computation process is: first to ξ ' (x _c) do following normalization: therefore ξ (x _c) value is 0 to this N number of round values of N-1; Φ _tnormalize to S the direction represented to S-1 with integer 0, for XY plane, to WLD{ ξ (x _c), Φ _ttwo-dimensional histogram carries out dimensionality reduction, fixing Φ _t, ask corresponding ξ (x _c) sub-histogram, according to the Φ of S dimension _tbe divided into the sub-histogram of S group, according to Φ _torder from small to large connects this S sub-histogram successively, definition f (x, y)=N × Φ _t+ ξ (x _c), then the histogram h of XY plane _{i, XY}=Σ _x,ym{f (x, y)=i}, i=0,1 ..., N Φ _t-1, wherein $M\left(A\right)=\left\{\begin{array}{ccc}0,& if& Aisfalse;\\ 1,& if& Aistrue.\end{array}\right.,$ Thus form N Φ _tthe WLD histogram H of dimension _xY, then the histogram h of three orthogonal planes (n=0:XY, n=1:XT, n=2:YT) is obtained by this method _i,n=Σ _{x, y, t}m{f (x, y, t)=i}, i=0,1 ..., N Φ _t-1, they are transferred to N Φ _tthe row vector H of dimension _n, connect successively, generate 3N Φ _tthe WLD-TOP feature row vector H of dimension _wT=[H ₀h ₁h ₂].

7. the living body faces detection method based on WLD-TOP descriptor according to right 1, is characterized in that, in step S1,

Described SVM classifier uses the SVM implementation tool based on LIBSVM; All k the proper vector composing training collection eigenmatrixes that training set is obtained $H_{train}=\left(\begin{array}{c}{H}_{WT,train,1}\\ .\\ .\\ .\\ H_{WT,train,k}\end{array}\right),$ Adopt SVM training, and with the model of training out to the test set eigenmatrix containing j proper vector $H_{test}=\left(\begin{array}{c}{H}_{WT,test,1}\\ .\\ .\\ .\\ H_{WT,test,j}\end{array}\right)$ Whether classification is the differentiation label of living body faces, and with true tag comparison, thus obtain the accuracy rate that In vivo detection differentiates.

8. the living body faces detection method based on WLD-TOP descriptor according to right 1, is characterized in that, in step S1,

Described SVM classifier, for the training sample set of input, adopts cross validation method to train, and utilizes trellis search method to find the optimized parameter collection { C, γ } of SVM.