CN107103266A - The training of two-dimension human face fraud detection grader and face fraud detection method - Google Patents
The training of two-dimension human face fraud detection grader and face fraud detection method Download PDFInfo
- Publication number
- CN107103266A CN107103266A CN201610098933.8A CN201610098933A CN107103266A CN 107103266 A CN107103266 A CN 107103266A CN 201610098933 A CN201610098933 A CN 201610098933A CN 107103266 A CN107103266 A CN 107103266A
- Authority
- CN
- China
- Prior art keywords
- image
- fraud detection
- characteristic vector
- dimension
- mrow
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/40—Spoof detection, e.g. liveness detection
Abstract
The invention provides the generation method of two-dimension human face fraud detection model, methods described includes：First, all people's face picture in training set is pre-processed, obtains normalizing facial image；Second, extract LBP characteristic vectors, Gabor wavelet feature vector sum one-dimensional pixel characteristic vector from each normalization facial image；3rd, these three characteristic vectors are carried out to be spliced to form final characteristic vector；4th, using SVMs to being trained based on being spliced to form final characteristic vector, obtain two-dimension human face fraud detection grader；This method is extracted the characteristic information of the difference of face and photo；Feature extraction is simply efficient, it is not necessary to which user purposely coordinates, and good effect can be obtained in the case of low resolution.In addition, the two-dimension human face fraud detection grader obtained based on the above method, the invention also provides face fraud detection method, this method has the advantages that detection accuracy is high, can effectively prevent face fraud.
Description
Technical field
The present invention relates to computer vision and graph and image processing field, the more particularly to training of two-dimension human face fraud detection grader and face fraud detection method.
Background technology
At present, two-dimentional biological identification technology (identification i.e. based on two-dimension human face biological characteristic) is a critically important research field.View transformation, block, complicated outdoor light be always recognition of face difficult point.These problems are solved although doing a lot of work, the fraud attack leak of face identification system is but ignored by big multisystem.Face identification system relies on planar graph and carries out identity detection, and system is highly susceptible to the fraud attack of photograph print or electronic photo.For example, association, the Windows XP and Vista of Asus and Toshiba notebook computer all built-in IP Camera and biological recognition system, by the face certification user for scanning user.In the black cap conference of 2009 (world-leading technical security meeting), the security breaches research group of Ha Noi university illustrates face identification system (the Veriface III of association for notebook of how easily being out-tricked using the human face photo of validated user, the SmartLogon V1.0.0005 of Asus, the Face Recognition 2.0.2.32 of Toshiba) enter notebook, and the use of these systems is all highest safe class.This leak is included into national vulnerability scan by National Institute of Standards and Technology now.Therefore solves the problems, such as that cheating attack improves the security and robustness of face identification system and be applied to it in practice, be a very urgent demand.
Fraud attack refers to a people and attempts to disguise oneself as other people by data falsification to obtain illegal access.Such as, a people cheats face identification system before camera by the photo of a validated user, video, mask, 3D models.Although a people can also use other fraud modes such as cosmetic, lift face, photo attack is most common fraud attack, because human face photo can easily pass through acquisition of downloading, take on the sly on the net.
At present, fraud attack leak has attracted the attention of more and more people, has also specially held within 2011 one " contest of IJCB2011 two-dimension human faces fraud attack defending ".Although the research databases that are more and more and having issued some public openings successively in this field, but the standard database that relatively objective exploitation test can be provided for fraud detection algorithm is not a lot, this field or not mature enough, up to the present ununified on optimal fraud detection algorithm common recognition.Or current existing face fraud detection research method is too complicated, there is no practicality (processing real-time is required in actual use), using some unconventional imaging systems (multiple spectra imaging) and high-resolution camera, these do not possess the condition of practical application.
The content of the invention
It is an object of the invention to overcome the drawbacks described above that current two-dimension human face fraud detection method is present, by finding the trickle difference of face and corresponding photo, and design a feature space to protrude the difference.In fact, human face photo all contains a certain degree of printing quality defect, can well it be detected with texture.Inspired by picture quality, printed matter feature, light reflection differences, the present invention proposes a kind of training method of two-dimension human face fraud detection grader.This method from picture by extracting corresponding characteristic vector in LBP, Gabor wavelet, three kinds of feature spaces of pixel characteristic, these three characteristic vectors are synthesized into a final characteristic vector again, this characteristic vector finally is sent into a nonlinear SVM classifier progress classification based training obtains a two-dimension human face fraud detection grader.Based on the two-dimension human face fraud detection grader, present invention also offers a kind of face fraud detection method, it can be determined that it is face or fraud image to go out input picture.
To achieve these goals, the invention provides the training method of two-dimension human face fraud detection grader, methods described includes：First, all people's face picture in training set is pre-processed, obtains normalizing facial image；Second, extract LBP characteristic vectors, Gabor wavelet feature vector sum one-dimensional pixel characteristic vector from each normalization facial image；3rd, these three characteristic vectors are carried out to be spliced to form final characteristic vector；4th, it is trained using the SVMs characteristic vector final to being spliced to form, obtains two-dimension human face fraud detection grader.
In above-mentioned technical proposal, methods described is specifically included：
Step S1) i-th of face picture in training set is pre-processed, wherein, 1≤i≤L；Obtain the normalization facial image z of 64 × 64 pixel sizesi；
Step S2) from normalization facial image ziMiddle extraction LBP characteristic vector L (zi)；
Step S3) from normalization facial image ziMiddle extraction Gabor wavelet feature vector G (zi)；
Step S4) will normalization facial image ziZoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional pixel characteristic vector P (zi)；
Step S5) by step S2), step S3), step S4) three kinds of textural characteristics extracting are spliced into final characteristic vector D (zi)=(L (zi),G(zi),P(zi))；
Step S6) based on Support vector regression algorithm to all characteristic vector D (zi) be trained, two-dimension human face fraud detection grader is obtained, wherein, 1≤i≤L.
In above-mentioned technical proposal, the step S1) specifically include：
Step S1-1) image gray processing processing is carried out to face picture：
The facial picture of face is traveled through, each pixel is handled, the rgb value of each pixel is obtained, red, blue, green value is extracted respectively by computing, the gray value after each pixel conversion is calculated：
Grey=(9798R+19235G+3735B)/32768
Wherein, Grey represents red component, green component and the blue component of each pixel in the gray value after conversion, R, G, B difference representative image；
Step S1-2) gray level image size is adjusted to 64 × 64 using bilinear interpolation；
Step S1-3) to being sized after image strengthen：
The modification of image histogram is carried out using histogrammic statistics, image each grey level pixel probability of occurrence is equal to change the pixel value of each grey level in image by adjusting, so as to realize image enhaucament；
Step S1-4) to enhanced image zooming-out image pixel matrix, obtain normalization facial image zi。
In above-mentioned technical proposal, the step S2) specifically include：
Step S2-1) to normalization facial image ziWithOperator obtains LBP images, then LBP images are divided into 3 × 3 overlapping regions, and the statistic histogram for extracting 59 dimensions from each region respectively synthesizes the statistic histogram characteristic vector of one 531 dimension；
Step S2-2) to normalization facial image ziWithOperator, extracts the statistic histogram characteristic vector for obtaining 59 dimensions；
Step S2-3) to normalization facial image ziWithOperator, extracts the statistic histogram characteristic vector for obtaining 243 dimensions；
Step S2-4) by step S2-1), step S2-2), step S2-3) obtained characteristic vector synthesizes a characteristic vector L (zi), the dimension of this feature vector is 59*9+59+243=833 dimensions.
In above-mentioned technical proposal, the step S3) specifically include：
Step S3-1) will normalization facial image zi32 × 32 sizes are zoomed to, Gabor wavelet conversion is carried out to the image after scaling：
The Gabor filter of p different directions and q different scale is taken to handle image after scaling, each pixel t0P × q Gabor amplitude Characteristics can be obtained, p × q Gabor amplitude Characteristics, which cascade up, is referred to as a Jet, is abbreviated as J, then pixel t in image0Jet be：
J(t0)=(M0,0(t0),...,M0,7(t0),...,M4,0(t0),...,M4,7(t0))
The Gabor amplitude Characteristics of all pixels point are cascaded up and obtain the characteristic vector F (z of facial imagei)：
F(zi)={ J (t0):t0∈zi}
Step S3-2) determine step S3-1) obtained characteristic vector F (zi) dimensionality reduction dimension, based on PCA to characteristic vector F (zi) dimensionality reduction is carried out, obtain the vector G of the Gabor wavelet feature after dimensionality reduction (zi)。
In above-mentioned technical proposal, the step S3-2) specifically include：
Step S3-2-1) by step S3-1) obtained characteristic vector F (zi) dimension d be divided into n deciles, it is determined that new dimension d' span；
Original feature vector F (zi) dimension be d, d value is divided into n equal portions, value set is as follows：
Wherein, "" represent integer rounding operation；
If obtaining after dimensionality reduction is characterized as G (zi), then its dimension d' takes this n value respectively：
Step S3-2-2) d' takes each value in set successively, calculates the corresponding face fraud detection mean absolute error set { MAE of all pictures of training setm}；
To the L pictures in training set, calculate when d' takes each value in set, the corresponding face fraud detection mean absolute error of all pictures of training set is as follows：
Wherein, j represents jth pictures in training set, and k represents that d' takes k-th of value in set, i.e.,ljRepresent the corresponding class label of jth pictures in training set：0 represents fraud image, and 1 represents real human face image,Represent the classification estimate of jth pictures in training set；Different MAE value sets { MAE may finally be obtainedmWherein m ∈ 1,2 ..., n；
Step S3-2-3) take set { MAEmIn minimum M AEmin, with MAEminCorresponding d' is used as final dimensionality reduction dimension；
Step S3-2-4) it is based on step S3-2-3) obtained d', using PCA to characteristic vector F (zi) dimensionality reduction is carried out, obtain the vector G of the Gabor wavelet feature after dimensionality reduction (zi)。
In above-mentioned technical proposal, the step S6) specifically include：
Step S6-1) optimization problem is built based on Support vector regression algorithm；
Hypothesized model training set sample is { x(i),y(i)(i=1,2 ..., L), x(i)Represent normalization facial image ziCharacteristic vector D (zi), y(i)Represent the corresponding classification of the image：Facial image or fraud image；Assuming that sample dimension is N, thenThe target of Support vector regression algorithm is to solve for two-dimension human face fraud detection grader f (x), makes f (x(i)) and y(i)Between difference be not more than ε, ε is threshold value, controls the worst error between actual label value and predicted estimate value；So, f (x) is defined as follows：
Wherein, " " represents inner product of vectors；W and b is the parameter solved；
Step S6-2) optimization problem (1) is converted using method of Lagrange multipliers, be converted to and its dual problem is solved, obtain two-dimension human face fraud detection grader f (x) expression formula.
Obtained two-dimension human face fraud detection grader is trained based on the above method, present invention also offers face fraud detection method, this method includes：
Step T1) face picture to be detected is pre-processed, obtain the normalization facial image z of 64 × 64 pixel sizes0；
Step T2) from normalization facial image z0Middle extraction LBP characteristic vector L (z0)；
Step T3) from normalization facial image z0Middle extraction Gabor wavelet feature vector G (z0)；
Step T4) will normalization facial image z0Zoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional pixel characteristic vector P (z0)；
Step T5) by step T2), step T3), step T4) three kinds of textural characteristics extracting are spliced into final characteristic vector D (z0)=(L (z0),G(z0),P(z0))；
Step T6) by step T5) obtained characteristic vector D (z0)) be input in two-dimension human face fraud detection grader, obtain testing result：Facial image or fraud image.
The advantage of the invention is that：
1st, the training method of two-dimension human face fraud detection grader proposed by the present invention；The trickle difference of face and corresponding photo is captured by merging a variety of textural characteristics；Absorb LBP, the complementary attribute of the powerful texture operator of Gabor wavelet two, wherein, what LBP was included is microtexture information, what Gabor wavelet was included is the information of macroscopic view, and pixel characteristic provides the information of the overall situation, therefore the characteristic vector trained fully is extracted the information of the difference of face and photo；
2nd, in the training method of two-dimension human face fraud detection grader proposed by the present invention, feature extraction is simply efficient, it is not necessary to which user purposely coordinates, and good effect can be obtained in the case of low resolution；Two-dimension human face characteristic vector pickup combination SVR regression algorithms realize available lower error during face fraud detection, meet practical application scene needs；
3rd, the textural characteristics that the present invention is used may also be used for carrying out recognition of face, be that face fraud detection and recognition of face provide a unique feature space；
4th, face fraud detection method of the invention has the advantages that detection accuracy is high, can effectively prevent face fraud；
5th, method of the invention can be applicable to many occasions such as face verification system, security monitoring.
Brief description of the drawings
Fig. 1 is the flow chart of the training method of the two-dimension human face fraud detection grader of the present invention.
Embodiment
Technological concept of the present invention is briefly described below.
LBP (Local Binary Pattern, local binary model) is characterized in that face Related Research Domain more commonly uses textural characteristics.Most basic LBP Image Codings are defined as in 3 × 3 window, using window center pixel as threshold value, and the gray value of 8 adjacent pixels is compared with it, if surrounding pixel values are more than center pixel value, the position of the pixel is marked as 1, is otherwise 0.So, 8 points in 3 × 3 neighborhoods can produce 8 bits (being typically converted into decimal number i.e. LBP codes, totally 256 kinds) through comparing, that is, obtain the LBP values of the window center pixel, and reflect with this value the texture information in the region.In order to adapt to the textural characteristics of different scale, by 3 × 3 neighborhood extendings to any neighborhood, and square neighborhood is instead of with circle shaped neighborhood region, the gray value for the point for not entirely falling within location of pixels is calculated using bilinear interpolation algorithm.In addition, having any number of pixels in the circle neighbour that it is R in radius that the LBP operators after improving, which allow, domain.LBP equivalent formulations (uniform patterns) can carry out dimensionality reduction to LBP characteristic vectors.When the circulation binary number corresponding to some local binary pattern is from 0 to 1 or from 1 to 0 be up to saltus step twice, binary system corresponding to the local binary pattern is known as an equivalent formulations class, such as 00000000,11111111,10001111 be all equivalent formulations class, pattern in addition to equivalent formulations class is all classified as another kind of, referred to as mixed mode class.Improved by such, the species of local binary patterns greatly reduces.The overwhelming majority that equivalent formulations class is accounted in assemble mode, using the histogram of these equivalent formulations classes and mixed mode class, can extract the feature of Geng Neng representative images intrinsic propesties.SymbolRepresent the equivalent formulations operator for P point in R circular field in radius.
Gabor wavelet is widely used in image recognition, image processing field, in area of pattern recognition, and Gabor wavelet conversion is also a kind of very effective Feature Descriptor.Gabor wavelet is very sensitive for the edge of image, using the teaching of the invention it is possible to provide good set direction and scale selection characteristic, and insensitive for illumination variation, using the teaching of the invention it is possible to provide the adaptability good to illumination variation.Meanwhile, two-dimensional Gabor function can strengthen the information of facial critical component (eyes, nose, face etc.), so that enhancing local characteristicses are possibly realized while overall face information is retained.In spatial domain, the Gabor filter of one 2 dimension is the product of sinusoidal a plane wave and gaussian kernel function, with the characteristic for obtaining optimal partial simultaneously in spatial domain and frequency domain, it is much like with human biological's visual characteristic, therefore, it is possible to describe the partial structurtes information corresponding to spatial frequency (yardstick), locus and set direction well.Substantially, Gabor wavelet conversion be in order to extract signal Fourier conversion local message, use a Gauss function as window function, because the Fourier of a Gauss function is converted or a Gauss function, Fourier inverse transformations are also local.By the selection of frequency parameter and Gaussian function parameter, Gabor transformation can choose the characteristic information at many positions.Facial image is inputted after Gabor wavelet is converted, one group of Gabor wavelet response image cluster is can obtain.
The principle of principal component analysis (Principal Component Analysis, PCA) dimension reduction method removes redundancy to remove the dependency relation in initial data between each data component, retains topmost composition.PCA calculates the characteristic vector in initial data covariance matrix corresponding to maximum several characteristic values and obtains corresponding subspace, by eigenvector projection to the subspace to reach the purpose that sample space is described with small number of feature.SVMs (being typically two classification problems) in solution classification problem, is that the criterion based on structural risk minimization finds an optimal separating hyper plane, sample is divided into two, and reaches that different classes of has interval between maximum class.The classification that discrete integer value represents sample is commonly used in classification problem, different from classification problem, the label of each sample in regression problem is continuous real number.Therefore Support vector regression (Support Vector Regression, SVR) is intended to find a hyperplane, is capable of the distribution of Accurate Prediction sample, approximate sample data.
In conjunction with the drawings and specific embodiments, the present invention will be further described.
As shown in figure 1, a kind of training method of two-dimension human face fraud detection grader, methods described includes：
Step S1) in training set i-th (1≤i≤L) individual face picture is pre-processed successively, obtain the normalization facial image z of 64 × 64 pixel sizesi(1≤i≤L)；Specifically include：
Step S1-1) image gray processing processing is carried out to the individual face pictures of i-th (1≤i≤L)：
The face picture includes：True collection face picture and secondary imaging (forgery) face picture.
Face-image is traveled through, each pixel is handled, the rgb value of each pixel is obtained, red, blue, green value is extracted respectively by computing, different to the sensitivity of red bluish-green three kinds of colors according to human eye, optimum gradation conversion formula is：
Grey=(9798R+19235G+3735B)/32768
Wherein Grey represents red component, green component and the blue component of each pixel in the gray value after conversion, R, G, B difference representative image；
Step S1-2) gray level image size is adjusted to 64 × 64 using bilinear interpolation；
Step S1-3) to being sized after image strengthen：
The modification of image histogram is carried out using histogrammic statistics, image each grey level pixel probability of occurrence is equal to change the pixel value of each grey level in image by adjusting, so as to realize image enhaucament；
Step S1-4) to enhanced image zooming-out image pixel matrix, obtain normalization facial image zi(1≤i≤L)。
Step S2) from normalization facial image ziMiddle extraction LBP characteristic vector L (zi)；
The step S2) specifically include：
Step S2-1) to normalization facial image ziWithOperator obtains LBP images, then LBP images are divided into 3 × 3 overlapping regions, and the statistic histogram for extracting 59 dimensions from each region respectively is finally synthesizing the statistic histogram characteristic vector of a 531 final dimensions；
Wherein,Represent the equivalent formulations operator for 8 points in 1 circular field in radius.
Step S2-2) to normalization facial image ziWithOperator, extracts the statistic histogram characteristic vector for obtaining 59 dimensions；
Wherein,Represent the equivalent formulations operator for 8 points in 2 circular field in radius.
Step S2-3) to normalization facial image ziWithOperator, extracts the statistic histogram characteristic vector for obtaining 243 dimensions；
Wherein,Represent the equivalent formulations operator for 16 points in 2 circular field in radius.
Step S2-4) by step S2-1), step S2-2), step S2-3) obtained characteristic vector synthesizes a characteristic vector L (zi), characteristic vector L (zi) dimension be 59*9+59+243=833 dimension.
Step S3) extract Gabor wavelet feature vector G (z from normalization facial imagei)；
The step S3) specifically include：
Step S3-1) will normalization facial image zi32 × 32 sizes are zoomed to, Gabor wavelet conversion is carried out to the image after scaling：
The Gabor filter of p different directions and q different scale is taken to handle image after scaling, each pixel t0P × q Gabor amplitude Characteristics can be obtained, p × q Gabor amplitude Characteristics cascade up commonly known as one Jet, are abbreviated as J, then pixel t in image0Jet be：
J(t0)=(M0,0(t0),...,M0,7(t0),...,M4,0(t0),...,M4,7(t0))
The Gabor amplitude Characteristics of all pixels point are cascaded up and obtain the characteristic vector F (z of facial imagei)：
F(zi)={ J (t0):t0∈zi}
Step S3-2) determine step S3-1) obtained characteristic vector F (zi) dimensionality reduction dimension, based on PCA to characteristic vector F (zi) dimensionality reduction is carried out, obtain the vector G of the Gabor wavelet feature after dimensionality reduction (zi)；Specifically include：
Step S3-2-1) by step S3-1) obtained characteristic vector F (zi) dimension d be divided into n deciles, it is determined that new dimension d' span；
Original feature vector F (zi) dimension be d, d value is divided into n equal portions, value set is as follows：
Wherein, "" represent integer rounding operation；
If obtaining after dimensionality reduction is characterized as G (zi), then its dimension d' takes this n value respectively：
Step S3-2-2) d' takes each value in set successively, calculates the corresponding face fraud detection mean absolute error set { MAE of all pictures of training setm}；
To the L pictures in training set, calculate when d' takes each value in set, the corresponding face fraud detection mean absolute error of all pictures of training set is as follows：
Wherein, j represents jth pictures in training set, and k represents that d' takes k-th of value in set, i.e.,ljThe corresponding class label of jth pictures in training set (0 represents fraud image, and 1 represents real human face image) is represented,Represent the classification estimate of jth pictures in training set；Different MAE value sets { MAE may finally be obtainedmWherein m ∈ 1,2 ..., n；
Step S3-2-3) take set { MAEmIn minimum M AEmin, with MAEminCorresponding d' is used as final dimensionality reduction dimension；
Step S3-2-4) it is based on step S3-2-3) obtained d', using PCA to characteristic vector F (zi) dimensionality reduction is carried out, obtain the vector G of the Gabor wavelet feature after dimensionality reduction (zi)。
Step S4) will normalization facial image ziZoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional pixel characteristic vector P (zi)；
Step S5) by step S2), step S3), step S4) three kinds of textural characteristics extracting are spliced into final characteristic vector D (zi)=(L (zi),G(zi),P(zi))；
Step S6) based on Support vector regression algorithm to all characteristic vector D (zi), (1≤i≤L) is trained, and obtains two-dimension human face fraud detection grader；Specifically include following steps：
Step S6-1) optimization problem is built based on Support vector regression algorithm；
Hypothesized model training set sample is { x(i),y(i)(i=1,2 ..., L), x(i)Represent normalization facial image ziCharacteristic vector D (zi), y(i)Represent the corresponding classification of the image：Facial image or fraud image.Assuming that sample dimension is N, thenSVR target is to be to solve for two-dimension human face fraud detection grader f (x), makes f (x(i)) and y(i)Between difference be not more than ε, ε is a minimum number, controls the worst error between actual label value and predicted estimate value.So, f (x) is defined as follows：
Wherein, " " represents inner product of vectors；W and b is the parameter solved；The w of solution should make | | w | |2It is minimum；Commonly referred to as the hyperplane model is ε-SVR；So ε-SVR optimization problem is represented by equation below：
s.t.|w·x(i)+b-y(i)|≤ε,i∈(1,2,...,m)
SVR introduces penalty coefficient and slack variable to be adjusted：
s.t.w·x(i)+b-y(i)≤ε+ξi (3)
y(i)-w·x(i)-b≤ε+ξi *
ξi,ξi *>=0, i=1,2 ..., m
Wherein,
Represent loss function；
In the present embodiment, penalty coefficient C=128, the learning parameter g=0.1 in setting SVR parameters, are trained using RBF (radial direction gaussian basis) kernel function；
Step S6-2) optimization problem is converted using method of Lagrange multipliers, be converted to and its dual problem is solved, obtain two-dimension human face fraud detection grader f (x) expression formula；
Introduce following Lagrangian：
WhereinRepresent αiWithRepresent ηiWithAnd αi、ηiAndIt is Lagrange multiplier.The solution of formula (9) belongs to convex quadratic programming problem category, by first solving L (w, α, η, b) function pair w, b, ξ minimum value conversion, then solve and meet L (w, α, η, b) function pair w, b, ξ local derviation are respectively 0 " saddle point "：
It is availableTaking back in formula (9) to obtain：
So, the Optimal solution problem of solution formula (9) is converted into the solution of following dual problem：
So, the optimization problem of an alpha parameter is only included with regard to problem is converted into, obtains that corresponding w just can be obtained after α value, final f (x) is：
According to KKT conditions, when SVR dual problem only meets following condition, the solution of dual problem is just equivalent to the solution of former problem：
αi(ε+ξi-y(i)+ wx+b)=0
(C-αi)ξi=0 (10)
It can be seen that and work as αiDuring=C, ξiJust it is not equal to 0, now sample point can fall in ε (i.e. outlier), then：
So b value is met：
It is also seen that by KKT conditions for | f (x)-y(i)|=ε+ξi (*)Sample point, its is correspondingJust it is not equal to 0；These sample points are supporting vector.
When kernel function is K (x(i), when x), detection function f (x) is changed into：
Training is obtained after SVR models, and only the corresponding point of supporting vector decides the predicted value of recurrence.
Obtained two-dimension human face fraud detection grader is trained based on the above method, present invention also offers two-dimension human face fraud detection method, methods described includes：
Step T1) face picture to be detected that is collected to camera pre-processes, and obtains the normalization facial image z of 64 × 64 pixel sizes0；
Step T2) from normalization facial image z0Middle extraction LBP characteristic vector L (z0)；
Step T3) from normalization facial image z0Middle extraction Gabor wavelet feature vector G (z0)；
Step T4) will normalization facial image z0Zoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional pixel characteristic vector P (z0)；
Step T5) by step T2), step T3), step T4) three kinds of textural characteristics extracting are spliced into final characteristic vector D (z0)=(L (z0),G(z0),P(z0))；
Step T6) by step T5) obtained characteristic vector D (z0)) be input to step S6) obtained two-dimension human face fraud detection grader f (x), obtain testing result：Facial image or fraud image.
It should be noted last that, the above embodiments are merely illustrative of the technical solutions of the present invention and it is unrestricted.Although the present invention is described in detail with reference to embodiment, it will be understood by those within the art that, technical scheme is modified or equivalent substitution, without departure from the spirit and scope of technical solution of the present invention, it all should cover among scope of the presently claimed invention.
Claims (8)
1. the training method of two-dimension human face fraud detection grader, methods described includes：First, to institute in training set
Somebody's face picture is pre-processed, and obtains normalizing facial image；Second, from each normalization facial image
Extract LBP characteristic vectors, Gabor wavelet feature vector sum one-dimensional pixel characteristic vector；3rd, these three are special
Vector is levied to carry out being spliced to form final characteristic vector；4th, utilize the spy that SVMs is final to being spliced to form
Levy vector to be trained, obtain two-dimension human face fraud detection grader.
2. the training method of two-dimension human face fraud detection grader according to claim 1, it is characterised in that
Methods described is specifically included：
Step S1) i-th of face picture in training set is pre-processed, wherein, 1≤i≤L；Obtain 64 × 64
The normalization facial image z of pixel sizei；
Step S2) from normalization facial image ziMiddle extraction LBP characteristic vector L (zi)；
Step S3) from normalization facial image ziMiddle extraction Gabor wavelet feature vector G (zi)；
Step S4) will normalization facial image ziZoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional
Pixel characteristic vector P (zi)；
Step S5) by step S2), step S3), step S4) three kinds of textural characteristics extracting be spliced into it is final
Characteristic vector D (zi)=(L (zi),G(zi),P(zi))；
Step S6) based on Support vector regression algorithm to all characteristic vector D (zi) be trained, obtain two dimension
Face fraud detection grader, wherein, 1≤i≤L.
3. the training method of two-dimension human face fraud detection grader according to claim 2, it is characterised in that
The step S1) specifically include：
Step S1-1) image gray processing processing is carried out to face picture：
The facial picture of face is traveled through, each pixel is handled, the rgb value of each pixel is obtained, passes through
Computing extracts red, blue, green value respectively, calculates the gray value after each pixel conversion：
Grey=(9798R+19235G+3735B)/32768
Wherein, each pixel is red in the gray value after Grey representatives conversion, R, G, B difference representative image
Colouring component, green component and blue component；
Step S1-2) gray level image size is adjusted to 64 × 64 using bilinear interpolation；
Step S1-3) to being sized after image strengthen：
The modification of image histogram is carried out using histogrammic statistics, is gone out by adjusting each grey level pixel of image
Existing probability is equal to change the pixel value of each grey level in image, so as to realize image enhaucament；
Step S1-4) to enhanced image zooming-out image pixel matrix, obtain normalization facial image zi。
4. the training method of two-dimension human face fraud detection grader according to claim 3, it is characterised in that
The step S2) specifically include：
Step S2-1) to normalization facial image ziWithOperator obtains LBP images, then by LBP images
3 × 3 overlapping regions are divided into, the statistic histogram for extracting 59 dimensions from each region respectively synthesizes one 531
The statistic histogram characteristic vector of dimension；
Step S2-2) to normalization facial image ziWithOperator, extracts the statistic histogram for obtaining 59 dimensions
Characteristic vector；
Step S2-3) to normalization facial image ziWithOperator, extracts the statistics Nogata for obtaining 243 dimensions
Figure characteristic vector；
Step S2-4) by step S2-1), step S2-2), step S2-3) obtained characteristic vector synthesizes a spy
Levy vectorial L (zi), the dimension of this feature vector is 59*9+59+243=833 dimensions.
5. the training method of two-dimension human face fraud detection grader according to claim 3, it is characterised in that
The step S3) specifically include：
Step S3-1) will normalization facial image zi32 × 32 sizes are zoomed to, Gabor is carried out to the image after scaling
Wavelet transformation：
The Gabor filter of p different directions and q different scale is taken to handle image after scaling, often
Individual pixel t0P × q Gabor amplitude Characteristics can be obtained, p × q Gabor amplitude Characteristics cascade up and are referred to as one
Individual Jet, is abbreviated as J, then pixel t in image0Jet be：
J(t0)=(M0,0(t0),...,M0,7(t0),...,M4,0(t0),...,M4,7(t0))
The Gabor amplitude Characteristics of all pixels point are cascaded up and obtain the characteristic vector F (z of facial imagei)：
F(zi)={ J (t0):t0∈zi}
Step S3-2) determine step S3-1) obtained characteristic vector F (zi) dimensionality reduction dimension, based on principal component analysis
Method is to characteristic vector F (zi) dimensionality reduction is carried out, obtain the vector G of the Gabor wavelet feature after dimensionality reduction (zi)。
6. the training method of two-dimension human face fraud detection grader according to claim 5, it is characterised in that
The step S3-2) specifically include：
Step S3-2-1) by step S3-1) obtained characteristic vector F (zi) dimension d be divided into n deciles, it is determined that
New dimension d' span；
Original feature vector F (zi) dimension be d, d value is divided into n equal portions, value set is as follows：
Wherein,Represent integer rounding operation；
If obtaining after dimensionality reduction is characterized as G (zi), then its dimension d' takes this n value respectively：
Step S3-2-2) d' takes each value in set successively, calculates the fraud inspection of all pictures of training set corresponding face
Survey mean absolute error set { MAEm}；
To the L pictures in training set, calculate when d' takes each value in set, all picture correspondences of training set
Face fraud detection mean absolute error it is as follows：
<mrow>
<msub>
<mi>MAE</mi>
<mi>k</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>L</mi>
</mfrac>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>L</mi>
</munderover>
<mo>|</mo>
<msub>
<mi>l</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mover>
<mi>l</mi>
<mo>^</mo>
</mover>
<mi>j</mi>
</msub>
<mo>|</mo>
</mrow>
Wherein, j represents jth pictures in training set, and k represents that d' takes k-th of value in set, i.e.,
ljRepresent the corresponding class label of jth pictures in training set：0 represents fraud image, and 1 represents real human face image,
Represent the classification estimate of jth pictures in training set；Different MAE value sets { MAE may finally be obtainedm}
Wherein m ∈ 1,2 ..., n；
Step S3-2-3) take set { MAEmIn minimum M AEmin, with MAEminCorresponding d' is as final
Dimensionality reduction dimension；
Step S3-2-4) it is based on step S3-2-3) obtained d', using PCA to characteristic vector F (zi) enter
Row dimensionality reduction, obtains the vector G of the Gabor wavelet feature after dimensionality reduction (zi)。
7. the training method of two-dimension human face fraud detection grader according to claim 3, it is characterised in that
The step S6) specifically include：
Step S6-1) optimization problem is built based on Support vector regression algorithm；
Hypothesized model training set sample is { x(i),y(i)(i=1,2 ..., L), x(i)Represent normalization facial image ziFeature
Vectorial D (zi), y(i)Represent the corresponding classification of the image：Facial image or fraud image；Assuming that sample dimension is N,
ThenThe target of Support vector regression algorithm is to solve for two-dimension human face fraud detection grader f (x), makes
f(x(i)) and y(i)Between difference be not more than ε, ε is threshold value, is controlled between actual label value and predicted estimate value
Worst error；So, f (x) is defined as follows：
F (x)=wx+b
<mrow>
<mtable>
<mtr>
<mtd>
<mrow>
<mi>s</mi>
<mo>.</mo>
<mi>t</mi>
<mo>.</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mo>&ForAll;</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>,</mo>
<mo>|</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>|</mo>
<mo>&le;</mo>
<mi>&epsiv;</mi>
</mrow>
</mtd>
</mtr>
</mtable>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, " " represents inner product of vectors；W and b is the parameter solved；
Step S6-2) optimization problem (1) is converted using method of Lagrange multipliers, be converted to it
Dual problem is solved, and obtains two-dimension human face fraud detection grader f (x) expression formula.
8. face fraud detection method, the two-dimension human face fraud that the method training based on one of claim 1-7 is obtained
Detect that grader is realized, this method includes：
Step T1) face picture to be detected is pre-processed, obtain the normalization face figure of 64 × 64 pixel sizes
As z0；
Step T2) from normalization facial image z0Middle extraction LBP characteristic vector L (z0)；
Step T3) from normalization facial image z0Middle extraction Gabor wavelet feature vector G (z0)；
Step T4) will normalization facial image z0Zoom to 8 × 8 sizes, then by two dimensional image thaumatropy be one-dimensional
Pixel characteristic vector P (z0)；
Step T5) by step T2), step T3), step T4) three kinds of textural characteristics extracting be spliced into it is final
Characteristic vector D (z0)=(L (z0),G(z0),P(z0))；
Step T6) by step T5) obtained characteristic vector D (z0)) be input in two-dimension human face fraud detection grader,
Obtain testing result：Facial image or fraud image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610098933.8A CN107103266B (en) | 2016-02-23 | 2016-02-23 | The training of two-dimension human face fraud detection classifier and face fraud detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610098933.8A CN107103266B (en) | 2016-02-23 | 2016-02-23 | The training of two-dimension human face fraud detection classifier and face fraud detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107103266A true CN107103266A (en) | 2017-08-29 |
| CN107103266B CN107103266B (en) | 2019-08-20 |
Family
ID=59658429
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610098933.8A Active CN107103266B (en) | 2016-02-23 | 2016-02-23 | The training of two-dimension human face fraud detection classifier and face fraud detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107103266B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107704834A (en) * | 2017-10-13 | 2018-02-16 | 上海壹账通金融科技有限公司 | Householder method, device and storage medium are examined in micro- expression face |
| CN107886070A (en) * | 2017-11-10 | 2018-04-06 | 北京小米移动软件有限公司 | Verification method, device and the equipment of facial image |
| CN108009531A (en) * | 2017-12-28 | 2018-05-08 | 北京工业大学 | A kind of face identification method of more tactful antifraud |
| CN108038413A (en) * | 2017-11-02 | 2018-05-15 | 平安科技（深圳）有限公司 | Cheat probability analysis method, apparatus and storage medium |
| CN109558794A (en) * | 2018-10-17 | 2019-04-02 | 平安科技（深圳）有限公司 | Image-recognizing method, device, equipment and storage medium based on moire fringes |
| WO2020097834A1 (en) * | 2018-11-14 | 2020-05-22 | 北京比特大陆科技有限公司 | Feature processing method and apparatus, storage medium and program product |
| CN111428666A (en) * | 2020-03-31 | 2020-07-17 | 齐鲁工业大学 | Intelligent family accompanying robot system and method based on rapid face detection |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101021900A (en) * | 2007-03-15 | 2007-08-22 | 上海交通大学 | Method for making human face posture estimation utilizing dimension reduction method |
| CN105095833A (en) * | 2014-05-08 | 2015-11-25 | 中国科学院声学研究所 | Network constructing method for human face identification, identification method and system |
| CN105117688A (en) * | 2015-07-29 | 2015-12-02 | 重庆电子工程职业学院 | Face identification method based on texture feature fusion and SVM |
-
2016
- 2016-02-23 CN CN201610098933.8A patent/CN107103266B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101021900A (en) * | 2007-03-15 | 2007-08-22 | 上海交通大学 | Method for making human face posture estimation utilizing dimension reduction method |
| CN100492399C (en) * | 2007-03-15 | 2009-05-27 | 上海交通大学 | Method for making human face posture estimation utilizing dimension reduction method |
| CN105095833A (en) * | 2014-05-08 | 2015-11-25 | 中国科学院声学研究所 | Network constructing method for human face identification, identification method and system |
| CN105117688A (en) * | 2015-07-29 | 2015-12-02 | 重庆电子工程职业学院 | Face identification method based on texture feature fusion and SVM |
Non-Patent Citations (4)
| Title |
|---|
| J. MAATTA 等: ""Face spoofing detection from single images using texture and local shape analysis"", 《IET BIOMETRICS》 * |
| J.MAATTA等: ""Face Spoofing Detection From Single Images Using Micro-Texture Analysis"", 《2011 INTERNATIONAL JOINT CONFERENCE ON BIOMETRICS》 * |
| 吴仰波: ""基于人脸图像特征表达的年龄估计模型算法研究与实现"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 陆丽: ""基于人脸图像的性别识别与年龄估计研究"", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107704834A (en) * | 2017-10-13 | 2018-02-16 | 上海壹账通金融科技有限公司 | Householder method, device and storage medium are examined in micro- expression face |
| CN107704834B (en) * | 2017-10-13 | 2021-03-30 | 深圳壹账通智能科技有限公司 | Micro-surface examination assisting method, device and storage medium |
| CN108038413A (en) * | 2017-11-02 | 2018-05-15 | 平安科技（深圳）有限公司 | Cheat probability analysis method, apparatus and storage medium |
| WO2019085331A1 (en) * | 2017-11-02 | 2019-05-09 | 平安科技（深圳）有限公司 | Fraud possibility analysis method, device, and storage medium |
| CN107886070A (en) * | 2017-11-10 | 2018-04-06 | 北京小米移动软件有限公司 | Verification method, device and the equipment of facial image |
| CN108009531A (en) * | 2017-12-28 | 2018-05-08 | 北京工业大学 | A kind of face identification method of more tactful antifraud |
| CN108009531B (en) * | 2017-12-28 | 2022-01-07 | 北京工业大学 | Multi-strategy anti-fraud face recognition method |
| CN109558794A (en) * | 2018-10-17 | 2019-04-02 | 平安科技（深圳）有限公司 | Image-recognizing method, device, equipment and storage medium based on moire fringes |
| WO2020097834A1 (en) * | 2018-11-14 | 2020-05-22 | 北京比特大陆科技有限公司 | Feature processing method and apparatus, storage medium and program product |
| CN112868019A (en) * | 2018-11-14 | 2021-05-28 | 北京比特大陆科技有限公司 | Feature processing method and device, storage medium and program product |
| CN111428666A (en) * | 2020-03-31 | 2020-07-17 | 齐鲁工业大学 | Intelligent family accompanying robot system and method based on rapid face detection |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107103266B (en) | 2019-08-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jourabloo et al. | Face de-spoofing: Anti-spoofing via noise modeling | |
| CN107103266B (en) | The training of two-dimension human face fraud detection classifier and face fraud detection method | |
| Huijuan et al. | Fast image matching based-on improved SURF algorithm | |
| CN109858439A (en) | A kind of biopsy method and device based on face | |
| CN109948566B (en) | Double-flow face anti-fraud detection method based on weight fusion and feature selection | |
| CN101142584A (en) | Method for facial features detection | |
| CN110287787B (en) | Image recognition method, image recognition device and computer-readable storage medium | |
| CN110929593A (en) | Real-time significance pedestrian detection method based on detail distinguishing and distinguishing | |
| Björklund et al. | Automatic license plate recognition with convolutional neural networks trained on synthetic data | |
| CN109740572A (en) | A kind of human face in-vivo detection method based on partial color textural characteristics | |
| CN106650606A (en) | Matching and processing method for face image and face image model construction system | |
| CN104021567B (en) | Based on the fuzzy altering detecting method of image Gauss of first numeral law | |
| Wong et al. | Thermal condition monitoring system using log-polar mapping, quaternion correlation and max-product fuzzy neural network classification | |
| Zarkasi et al. | Face movement detection using template matching | |
| Barni et al. | Iris deidentification with high visual realism for privacy protection on websites and social networks | |
| Lian | Pedestrian detection using quaternion histograms of oriented gradients | |
| CN104268509A (en) | Method and system for detecting license plate of dump truck | |
| CN109241932A (en) | A kind of thermal infrared human motion recognition method based on movement variogram phase property | |
| Singh et al. | Template matching for detection & recognition of frontal view of human face through Matlab | |
| Hao et al. | Active cues collection and integration for building extraction with high-resolution color remote sensing imagery | |
| Wu et al. | Face detection based on YCbCr Gaussian model and KL transform | |
| Deb et al. | Statistical characteristics in HSI color model and position histogram based vehicle license plate detection | |
| AKSAKALLI et al. | A novel approach for copy-move forgery detection using bilateral filtering | |
| Mitra et al. | Machine learning approach for signature recognition by harris and surf features detector | |
| Hu et al. | Pedestrian traffic light detection in complex scene using adaboost with multi-layer features |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220729 Address after: 100190, No. 21 West Fourth Ring Road, Beijing, Haidian District Patentee after: INSTITUTE OF ACOUSTICS, CHINESE ACADEMY OF SCIENCES Patentee after: NANHAI RESEARCH STATION, INSTITUTE OF ACOUSTICS, CHINESE ACADEMY OF SCIENCES Address before: 100190, No. 21 West Fourth Ring Road, Beijing, Haidian District Patentee before: INSTITUTE OF ACOUSTICS, CHINESE ACADEMY OF SCIENCES |