CN103984933A - Anti-occlusion face recognition method based on DWT-DCT watermark under big data - Google Patents

Abstract

The invention discloses an anti-occlusion face recognition method based on DWT-DCT watermarks under big data. The anti-occlusion face recognition method mainly includes a watermark embedding process and a watermark extraction process, wherein face recognition is realized in the watermark extraction process. The anti-occlusion face recognition method comprises the following steps that in the watermark embedding process, firstly, wavelet transformation is conducted on all original faces and global DCT is conducted on approximation coefficients of the original faces so that feature vectors can be worked out, and secondly, the watermark of each face is associated with the feature vector of the corresponding face through a hash function in cryptology; in the watermark extraction process, thirdly, the feature vector of the face to be tested is worked out, the correlation coefficient maximum value of the feature vector of the face to be tested and the original face feature vector is obtained, face recognition is completed according to a serial number corresponding to the correlation coefficient maximum value, and the corresponding embedded watermark is obtained, and fourthly, the feature vector of the face to be tested is used for watermark extraction, and the correlation coefficient of the watermark is calculated. The anti-occlusion face recognition method has the advantages of being applicable to the big data because sampling training is not needed, and has high resistance to occlusion, light and other attacks.

Description

Under large data, based on DWT-DCT watermark anti-, block attack recognition of face

Technical field

The present invention relates to a kind of field of multimedia signal processing, be specifically related under large data to block attack recognition of face based on DWT-DCT watermark anti-.

Technical background

Face recognition technology, as a kind of effective biometrics identification technology, obtains the attention of industry member and academia day by day over nearly 40 years.Because face recognition technology has height, can accept, naturality, be difficult for by advantages such as people discover, so it has larger purposes at aspects such as amusement, crime survey, gate control system, military affairs.

The method of recognition of face is at present mainly based on machine learning methods such as PCA, neural network, SVM, owing to will carrying out training study, when sample number larger, under large data environment, the time of study is longer, and current face identification method is for illumination variation, and expression shape change or shade etc. are more responsive, therefore how to solve under large data environment, the face identification method that anti-illumination variation, expression shape change or shade etc. are attacked is significant.

Digital watermark technology is the copyright protection for the Digital Media on internet at first, and its key property is robustness and invisibility; The present invention can be hidden in people's signature or ID etc. in its corresponding facial image as watermark, utilizes the robust property of watermark to realize face recognition algorithms, experimental results show that this algorithm especially to illumination, the attack such as block and there is good robustness.At present for based under large data environment, anti-illumination, block Face Recognition less of attack, therefore study the anti-attack face identification method that blocks based on DWT-DCT watermark under large data, there is larger meaning.

Summary of the invention

Goal of the invention: technical matters to be solved by this invention is for the deficiencies in the prior art, provides at a high speed, the face recognition algorithms of high robust.Specifically disclosing the anti-attack face identification method that blocks based on DWT-DCT watermark under a kind of large data, is a kind of zero watermarking project, and the embedding of watermark does not affect original facial image.

To achieve these goals, the present invention is performed such: first facial image is carried out to wavelet transformation and obtain " approximation coefficient " and " detail coefficients ", and known according to Principle of Wavelet." approximation coefficient " represents the low frequency characteristic of facial image, reflection be the main profile of facial image; " detail coefficients " represents the high frequency characteristics of facial image, reflection be the detailed information of facial image.Because the resist geometric attacks ability of wavelet transformation itself is poor, for this reason, we first carry out wavelet transformation (DWT) to facial image, then " approximation coefficient " of reflection low frequency characteristic carried out to overall cosine transform (DCT), in dct transform coefficient, extract a proper vector, find an anti-proper vector of blocking attack, and watermark sequence is associated and realizes the embedding of watermark with this proper vector; Then for testing image, first calculate its proper vector, then calculate the related coefficient of the proper vector of testing image and original image, utilize related coefficient maximal value, realize the detection of people's face; And realized the extraction of watermark.

Now to method of the present invention, be elaborated as follows:

First select a significant binary sequence will embed in facial image as watermark, be designated as W={w (j) | w (j)=0,1; 1≤i≤L}; Meanwhile, choosing and establishing F is original facial image, represents: F={f (i, j) | f (i, j) ∈ R; 1≤i≤M1,1≤j≤N1}.Wherein, f (i, j) represents the pixel value of primitive man's face.

First: the embedding of watermark

1) by each original facial image F (n) is carried out to wavelet transformation, then pairing approximation coefficient carries out overall dct transform, obtains the proper vector set V (n) of original image;

First successively each original facial image F (n) is carried out to L level wavelet decomposition and obtain ll channel coefficient FA _l(i, j).Because wavelet decomposition progression is higher, the resolving time taking just can be longer, so president is sent out in the evacuation of intelligent anti-counterfeiting consuming time.At this, we select L=1.Then to ll channel FA _l(i, j) carries out overall dct transform, obtains DWT-DCT matrix of coefficients FD (i, j),, in DWT-DCT matrix of coefficients FD (i, j), choose front 8 * 8 coefficient FD ₈(i, j), and then to the matrix of coefficients FD selecting ₈(i, j) carries out binary conversion treatment, when coefficient is more than or equal to zero, gets 1, and being less than 0 is to get zero, obtains proper vector V, and main process is described below:

FA _L(i,j)＝DWT2(F(i,j))

FD ₈(i,j)＝DCT2(FA _L(i,j))

V(n)＝BINARY(FD ₈(i,j))

2) utilize cryptography HASH function, generate two-value key sequence Key (n) with watermarked information, realize the embedding of zero watermark;

Key(n)＝V(n)⊕W(n)

Here, 1≤n≤N, N is the number of everyone face; Key (n) is that the Hash function conventional by cryptography generates by the proper vector V (n) of all original images and corresponding digital watermarking W (n); The random series that W (n) is 64bit by length forms; Preserve Key (n), while extracting watermark below, will use; By Key (n) is applied for to third party as key, to obtain the right to use and the entitlement of facial image;

Second portion: the identification of people's face and the extraction of watermark

3) obtain the proper vector V' of people's face F' to be measured;

If people's face to be measured is F'(i, j), through wavelet transformation (DWT), to carry out obtaining DWT-DCT matrix of coefficients after overall dct transform be FD ' (i to pairing approximation coefficient again, j), by the method for above-mentioned Step1, try to achieve the visual feature vector V ' of testing image;

FA _L’(i,j)＝DWT2(F’(i,j))

FD ₈’(i,j)＝DCT2(FA _L’(i,j))

V'＝BINARY(FD ₈’(i,j))

4) calculate the related coefficient NC (n) of the proper vector V' of people's face to be measured and the V (n) of N primitive man's face, carry out the identification of people's face;

Obtain the corresponding n value of related coefficient maximal value of V' and V (n), establish n=k; According to k value, can obtain key K ey (k), obtain original facial image is F (k) and the watermark value W (k) that is embedded in F (k); The normalized correlation coefficient formula of calculated characteristics vector is as follows:

$\mathrm{NC}=\frac{\underset{i}{\mathrm{\Σ}}V\left(i\right)V^{\′}}{\underset{i}{\mathrm{\Σ}}{V}^2\left(i\right)}$

5) utilize the proper vector V ' that is present in third-party two-valued function key sequence Key (k) and people's face to be measured, extract the watermark W' in testing image;

W'＝Key(k)⊕V'

6) calculate the related coefficient between W' and W (k); Obtain the similarity of watermark, and can differentiate according to watermark the content of testing image.

The present invention and existing face recognition technology relatively have following advantage:

First: preferably the identification of the robustness of watermark and invisibility and people's face has been carried out to organic combination; Utilize the Robust Algorithms of watermark, realized the attack such as anti-illumination, shade, face's distortion of people's face algorithm, and the embedding of watermark does not affect the pixel value of original image; Secondly: watermark can also be protected individual privacy, only has the user of mandate just can carry out the identification of people's face; At utmost protect individual privacy; Finally, due to this face recognition technology, do not need the study of sample, so can carry out a large amount of recognitions of face, be suitable for the identification of large data human face.

Below from theoretical foundation and experimental data explanation:

1) wavelet transform (DWT)

The wavelet transformation (DWT) that S.Mallat proposed in 1988, is the new signal analysis theory rising in recent years, its " time one frequently " analytical approach that is a kind of, and its basic thought is with wavelet function ψ _a,b(t) be substrate, signal f (t) is decomposed.

$W{f}_{a,b}={\∫}_{R}f\left(t\right){\stackrel{\‾}{\mathrm{\ψ}}}_{a,b}\left(t\right)\mathrm{dt}$

Its Wavelets ψ _a,b(t) be through translation, flexible and one group of function obtaining by same basis function ψ.

ψ _a,b(t)＝|a| ^-1/2ψ((t-b)/a) a,b∈R，a≠0

ψ is called base small echo, and a is contraction-expansion factor, and b is shift factor.

Mallat algorithm decomposition formula is:

$c_{j+1,k}=\underset{n\∈z}{\mathrm{\Σ}}{c}_{j,n}{\stackrel{\‾}{h}}_{n-2k},k\∈z$

$d_{j+1,k}=\underset{n\∈z}{\mathrm{\Σ}}{c}_{j,n}{\stackrel{\‾}{g}}_{n-2k},k\∈z$

Mallat algorithm reconstruction formula is:

$c_{j,k}=\underset{n\∈z}{\mathrm{\Σ}}{c}_{j+1,n}{h}_{k-2n}+\underset{n\∈z}{\mathrm{\Σ}}{d}_{j+1,n}{g}_{k-2n},k\∈z$

2D signal image is carried out after one-level wavelet decomposition, former figure is divided into four subgraphs, wherein three high frequency details subgraphs (level, vertical and diagonal) and a low frequency ll channel, high frequency details subgraph has mainly comprised the marginal information of former figure, but be easily subject to the impact (conventional image is processed) of external disturbance, and the essential information that low frequency ll channel comprises image (low frequency part), be subject to external action little, therefore utilize low frequency ll channel to obtain the robustness that proper vector can strengthen this algorithm.

2) discrete cosine transform

DCT is the standard of now widely used JPEG compression and MPEG-1/2 for Image Coding.DCT is in the little suboptimum orthogonal transformation that is only second to Karhunen-Loeve transformation drawing of Minimum Mean Square Error condition, is a kind of harmless chief of a tribe conversion.Its fast operation, precision is high, to extract the ability of characteristic component and the optimum balance between arithmetic speed and famous.

2-D discrete cosine direct transform (DCT) formula is as follows:

$F(u,v)=c\left(u\right)c\left(v\right)\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}f(x,y)\cos \frac{\mathrm{\π}(2x+1)u}{2M}\cos \frac{\mathrm{\π}(2y+1)v}{2N}$

u＝0,1,…,M-1；v＝0,1,…,N-1；

In formula

$c\left(u\right)=\left\{\begin{array}{cccc}\sqrt{1/M}& u=0& & \\ \sqrt{2/M}& u=\mathrm{1,2},...,M-1& & \end{array}c\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N}& v=0\\ \sqrt{2/N}& v=\mathrm{1,2},...,N-1\end{array}\right.\right.$

2-D discrete cosine inverse transformation (IDCT) formula is as follows:

$f(x,y)=\underset{u=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{v=0}{\overset{N-1}{\mathrm{\Σ}}}c\left(u\right)c\left(v\right)F(u,v)\cos \frac{\mathrm{\π}(2x+1)u}{2M}\cos \frac{\mathrm{\π}(2y+1)v}{2N}$

x＝0,1,…,M-1；y＝0,1,…,N-1

X wherein, y is spatial domain sampled value; U, v is frequency field sampled value, digital picture represents by pixel square formation conventionally, i.e. M=N

From formula above, the coefficient symbols of DCT is relevant with the phase place of component.

3) choosing method of the visual feature vector of facial image robust

Current most of face recognition algorithms is that attack to implement be that under spatial domain, larger variation has occurred pixel value, particularly the impact of indoor and outdoor illumination when these to illumination, shade, the poor reason of expression shape change anti-attack ability; Can the proper vector that find an anti-illumination, shade etc. to attack significant.If can find the proper vector of a reflection facial image geometrical feature, when there is little geometric transformation and illumination variation in facial image, can there is not obvious sudden change in this proper vector, so according to this proper vector, utilize the maximal value of related coefficient, we try to achieve corresponding original image, and extract corresponding watermark.

For this reason, choose some illumination and attack, the experimental data of the conventional attacks such as shade attack and filtering is shown in Table 1.The original image that is used as test in table 1 is Fig. 1, is the first width people face of ORL face database, by the AT & T of Cambridge University establishment of laboratory; What in table 1, " the 1st row " showed is the attack type that face recognition algorithms is subject to, and this facial image after the illumination that is subject to varying strength is attacked is shown in Fig. 2 to Fig. 4, and intensity of illumination is ascending; Block to attack and see Fig. 5 to Fig. 7; Comprise, study that common glasses block, mouth mask blocks blocks with cap; Fig. 8 is that extruding is attacked, and Figure 10 is that sphere is attacked, and such attack is similar to expression shape change and attacks; The conventional attacks such as Gaussian noise, JPEG compression, medium filtering are shown in Figure 10 to Figure 12." the 2nd row " of table 1 represents is the Y-PSNR (PSNR) of facial image after under attack; " the 10th row " arrive in " the 3rd row " of table 1, are the low frequency coefficients after DWT-DCT converts, choose here " D (1,1), D (1,2) ... D (1,8) " etc. eight low frequency variations domain coefficients." the 11st row " of table 1 are the symbol sebolic addressings for feature extraction.We find data by this table, and for attacks such as illumination, shade bird cagings, the low frequency value D of these dct transform domains (1,1), may there are some conversion in D (1,8) etc., but its coefficient symbols is still constant, we,, by the coefficient that is more than or equal to 0, are designated as 1; Be less than 0 be designated as 0, so for primitive man's face data, coefficient value D (1,1), D (1,2) ... the symbol binary sequence of the correspondences such as D (1,8) is: " 11000000 ", the 11st row in Table 1, observing these row can find, no matter illumination is attacked, and shade is attacked, distortion attack or conventional attack, the maintenance of its symbol sebolic addressing and raw data is similar, and all larger with the symbol sebolic addressing normalized correlation coefficient of raw data, in Table the 12nd row.The related coefficient that result shows is all greater than 0.5, and this is the facies relationship numerical value that 64 bit sign sequences are tried to achieve before low frequency real part coefficient is got.

Table 1 facial image DWT-DCT coefficient and part coefficient are subject to the changing value after different attack

In order to further illustrate the proper vector that symbol sebolic addressing is facial image, we obtain its visual feature vector by different facial images, then calculate the related coefficient between them; Here getting people's face of front 10 normal expressions of ORL face database tests; Figure 13-Figure 22, from angle of statistics, has got front 8 * 8 64 DWT-DCT coefficients altogether of DWT-DCT territory low frequency part here.And obtain the related coefficient between each proper vector, result of calculation is as shown in table 2.

As can be seen from Table 2, first, the related coefficient between facial image self is maximum, is 1.00; Secondly, the related coefficient of the characteristics of image of different people face is all not more than 0.5; This with our human eye actual observation to be consistent, the proper vector that this explanation is extracted by the method for this invention, reflected main resemblance and the main profile of people's face, and people's face profile is more similar, the similarity degree of proper vector is higher.

Related coefficient (vector length 64bit) between 10 different people face image feature vectors of table 2.

	V1	V2	V3	V4	V5	V6	V7	V8	V9	V10
											V1	1.00	0.22	0.44	0.30	0.43	0.25	0.01	0.36	0.06	0.13
V2	0.22	1.00	0.09	0.38	0.30	-0.03	0.21	0.28	0.02	0.02
											V3	0.44	0.09	1.00	0.25	0.38	0.25	0.05	0.19	0.24	0.05
V4	0.30	0.38	0.25	1.00	0.30	0.25	0.00	0.43	0.19	0.31
											V5	0.43	0.30	0.38	0.30	1.00	0.18	0.12	0.18	0.06	0.19
V6	0.25	-0.03	0.25	0.25	0.18	1.00	0.37	0.12	0.49	0.18
											V7	0.01	0.21	0.05	0.00	0.12	0.37	1.00	0.00	0.30	0.05
V8	0.36	0.28	0.19	0.43	0.18	0.12	0.00	1.00	0.00	0.12
											V9	0.06	0.02	0.24	0.19	0.06	0.49	0.30	0.00	1.00	0.24
V10	0.13	0.02	0.05	0.31	0.19	0.18	0.05	0.12	0.24	1.00

4) relation of the robustness of the length of proper vector and people's face algorithm

According to human visual system (HVS), Low Medium Frequency signal is larger to people's visual impact, for two dimensional image, is image outline.Therefore, we choose the Low Medium Frequency coefficient of facial image when facial image is chosen to proper transformation coefficient, the number of Low Medium Frequency coefficient is selected and the original facial image size of carrying out DWT-DCT conversion, and the quantity of information of the watermark of disposable embedding and require relevant, the length L of the proper vector of choosing is less, the quantity of information of disposable embedding is fewer, but the robustness of watermark is higher.Consider experiment below, the length that we choose L when specific experiment is 64.

Accompanying drawing explanation

Fig. 1 is original facial image.

Fig. 2 is the facial image after illumination is attacked, and intensity of illumination is S.

Fig. 3 is the facial image after illumination is attacked, and intensity of illumination is M.

Fig. 4 is the facial image after illumination is attacked, and intensity of illumination is L.

Fig. 5 is the facial image blocking through glasses.

Fig. 6 is the facial image blocking through mouth mask.

Fig. 7 is the facial image blocking through cap.

Fig. 8 is the facial image after extruding distortion is attacked, and distortion quantity is-40%.

Fig. 9 is the facial image after sphere distortion is attacked, and distortion quantity is 20%.

Figure 10 is the facial image disturbing through Gauss, and gaussian intensity is 5%.

Figure 11 is that compression quality is 5% through the facial image of JPEG compression.

Figure 12 is the facial image through medium filtering, and filtering parameter is [3x3], and filter times is 10.

The facial image of the 1st people in Figure 13 ORL face database.

The facial image of the 2nd people in Figure 14 ORL face database.

The facial image of the 3rd people in Figure 15 ORL face database.

The facial image of the 4th people in Figure 16 ORL face database.

The facial image of the 5th people in Figure 17 ORL face database.

The facial image of the 6th people in Figure 18 ORL face database.

The facial image of the 7th people in Figure 19 ORL face database.

The facial image of the 8th people in Figure 20 ORL face database.

The facial image of the 9th people in Figure 21 ORL face database.

The facial image of the 10th people in Figure 22 ORL face database.

The facial image that Figure 23 is to be measured, noiseless attack.

Figure 24 when attacking, the original image detecting.

Figure 25 when attacking, the watermark extracting.

Figure 26 is that intensity of illumination is-100% facial image.

When Figure 27 intensity of illumination is-100%, the original facial image detecting.

Figure 28 intensity of illumination is-100% watermark extracting.

The facial image that Figure 29 glasses block, blocks size for M.

When Figure 30 glasses block, the facial image detecting.

When Figure 31 glasses block, the watermark of extraction.

The facial image that Figure 32 mouth mask blocks, blocks size for M.

Figure 33 mouth mask blocks and is, the facial image detecting.

When Figure 34 mouth mask blocks, the digital watermarking of extraction.

The facial image that Figure 35 cap blocks, blocks size for M.

When Figure 36 cap blocks, the facial image detecting.

When Figure 37 cap blocks, the digital watermarking of extraction.

The facial image of Figure 38 face extruding, crushing strength is 40%.

Figure 39 when crushing strength is 40%, the facial image of identification.

Figure 40 when crushing strength is 40%, the digital watermarking extracting.

The facial image of Figure 41 sphere distortion, distortion quantity is 20%.

Figure 42 when sphere distortion quantity is 20%, the facial image detecting.

Figure 43 when sphere distortion quantity is 20%, the watermark of extraction.

The facial image that Figure 44 Gaussian noise is attacked, Gaussian noise intensity is 5%.

Figure 45 is 5% in Gaussian noise intensity, the facial image recognizing.

Figure 46 is in Gaussian noise intensity 5%, the watermark of extraction.

The facial image of Figure 47 JPEG compression, compression quality is 5%.

Figure 48 at JPEG boil down to 5% is, the original image of identification.

Figure 49 when JPEG boil down to 5%, the watermark information of extraction.

The facial image of Figure 50 medium filtering, filtering parameter is: [3x3], filter times is 10.

Figure 51 is [3x3] at medium filtering, the original image detecting.

Figure 52 is in medium filtering parameter [3x3], the watermark of extraction.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described, emulation platform is Matlab2010a, produces 1000 groups of independently binary pseudo-random (value+1 or 0), and every group of sequence length is 64bit, corresponding 1000 digital watermarkings of these 1000 groups of data, W (n).Each watermark correspondence an original image, establishes the piece image of the corresponding ORL of the 500th group of watermark here, sees Fig. 1, and its size be 92x112, and this figure is in this as original image be subject to the test pattern of zero attack.

Convenient in order to test the experiment that anti-illumination, shade attacks, the other 1000 groups of two-value random seriess of regeneration (value is+1, or 0), every group of sequence length is 64bit, these 1000 groups of proper vectors as 1000 original images, V (n); The 500th unit deposited to the original feature vector of the first width people face figure in ORL face database;

Its basic ideas are, establish F ' and be a facial image to be measured while not being interfered, and see Figure 23, obtain its proper vector V', then, by asking the related coefficient of the V (n) of V' and all N original image, obtain its maximal value; Then determine corresponding sequence number k, make when n=k, the related coefficient of V' and V (k) is maximum, according to sequence number k value, try to achieve the original facial image F (k) of identification, watermark W (k) and corresponding key K ey (k), wherein Key (k)=W (k) ⊕ V (k) with the embedding of corresponding original image F (k); According to facial image F' characteristic of correspondence vector V' to be measured and Key (k), obtain the watermark information W' in testing image F', W'=V' ⊕ Key (k), and obtain the related coefficient of W and W ', according to NC, be worth size to determine the degree of correlation of watermark and the watermark content of testing image.

Figure 23 is the facial image to be measured not adding while disturbing;

Figure 24 is the original facial image detecting;

Figure 25 is the watermark of extracting, and can see NC=1.00, can accurately extract watermark.

Below we are judged the anti-illumination of this face recognition algorithms based on robust watermarking, are blocked and the abilities such as conventional attack by specific experiment.

First test the anti-illumination of this face recognition algorithms, block with distortion and attack.

(1) experiment is attacked in illumination

Table 3 is experimental datas that anti-illumination is attacked.Therefrom can see, when testing image reduces intensity of illumination, when intensity of illumination is-100%, the signal to noise ratio (S/N ratio) of testing image is 7.21dB, by asking the related coefficient maximal value of proper vector, corresponding original image correctly detected, and can accurately extract digital watermarking, related coefficient NC=0.82; And adopt google to scheme to search figure function, when testing image is Figure 23, when it is subject to attack that intensity of illumination is-100%, cannot search correct original image.As can be seen from Table 3, when intensity of illumination is 100%, at this moment the signal to noise ratio (S/N ratio) of testing image is 9.34dB, correct original image still can be detected and extract watermark, and watermark related coefficient is 0.95; Illustrate that this face recognition algorithms based on watermark has good anti-illumination attacking ability.

Figure 26 is that intensity of illumination is-80% testing image; Dark images visually, signal to noise ratio (S/N ratio) is lower, and PSNR is 8.43dB;

Figure 27 is the facial image recognizing, and the proper vector related coefficient of the proper vector of this figure and testing image is maximal value;

Figure 28 is the watermark of extracting, and NC=0.97, is easy to extract watermark.

Experimental data is attacked in the anti-illumination of table 3

(2) block and attack experiment

Blocking is in face recognition process, and insoluble problem, but face identification method based on robust watermarking are utilized the watermarking algorithm of robust, preferably resolve this problem.Table 4 is anti-experimental datas of blocking attack.Comprise that glasses block, mouth mask blocks with cap and blocks; In table, S, M, L represents respectively the relative size of the area that blocks; From table 4, can see, when glasses block when more, block size for L, the PSNR=14.73dB of shielded image, signal to noise ratio (S/N ratio) is lower, but still can correctly identify original image, and extracts watermark, NC=0.84; When mouth mask blocks morely, block size when the L, the PSNR=17.09dB of shielded image, still can extract watermark, NC=0.74; When cap blocks, shielded area is larger, and during for L, at this moment signal to noise ratio (S/N ratio) is PSNR=12.20dB, still can extract watermark, NC=0.68

Figure 29 is that glasses block, and blocking size is the testing image of M; At this moment have certain occlusion effect, signal to noise ratio (S/N ratio) PSNR is 16.28dB;

Figure 30 is the original facial image detecting; Can correctly identify corresponding original image;

Figure 31 is the watermark of extracting, NC=0.82, can be correct extract watermark.

Figure 32 is that mouth mask blocks, and blocks size for M, and at this moment occlusion effect is obvious, and signal to noise ratio (S/N ratio) is lower, and PSNR is 17.39dB;

Figure 33 is the original image of identification, can find out and can correctly identify original image;

Figure 34 is the watermark of extracting, and NC=0.74, can correctly extract the watermark of testing image.

Figure 35 is that cap blocks, and blocking size is the testing image of M; At this moment occlusion effect is obvious, and signal to noise ratio (S/N ratio) PSNR is 13.10dB;

Figure 36 is the facial image of identification; Can correctly identify corresponding original image;

Figure 37 is the watermark of extracting, and NC=0.65, is easy to detect watermark.

Table 4 is anti-blocks attack

(3) experiment is attacked in face's extruding

The experimental data that the extruding of table 5 Shi Kang face is attacked, the variation of expression sometimes can be reflected in facial extruding.Therefrom can see, when testing image face extruding degree is larger, when extruding quantity is 100%, at this moment the signal to noise ratio (S/N ratio) of image is lower is 16.60dB, but still original image can be detected and extract watermark, NC=0.59;

Figure 38 is that crushing strength is 40% testing image; Facial expression changes more obvious, and signal to noise ratio (S/N ratio) is 19.03dB;

Figure 39 is the facial image recognizing, and can correctly recognize facial image;

Figure 40 is the watermark of extracting, and NC=0.90, is easy to detect watermark.

The extruding of table 5 face is attacked

(4) sphere distortion is attacked

Table 6 is experimental datas that the sphere distortion of the anti-face of watermarking algorithm is attacked.From table, can see, when the distortion of testing image sphere is larger, when distortion quantity is 60%, at this moment the signal to noise ratio (S/N ratio) of testing image is lower is 17.62dB, but still original image can be detected, and can extract watermark, NC=0.54;

Figure 41 is that sphere distortion is attacked, the testing image that distortion quantity is 20%; The distortion of face is more obvious, and signal to noise ratio (S/N ratio) is 20.45dB;

Figure 42 is the facial image recognizing, and primitive man's face can be detected;

Figure 43 is the watermark of extracting, and NC=0.84, can correctly extract watermark.

The distortion of table 6 sphere is attacked

Above-mentioned mainly to illumination, the attack such as block and test, below the conventional attacks such as filtering are tested;

(1) attacked by noise experiment

Use imnoise () function in testing image, to add Gaussian noise.

Table 7 is experimental datas of anti-Gauusian noise jammer.Therefrom can see, when testing image Gaussian noise intensity is up to 30% time, the PSNR of testing image is down to 7.98dB, and at this moment testing image is second-rate; Still original image can be correctly detected, watermark can be extracted, NC=0.70.And use google to scheme to search figure function, when the Gaussian noise intensity of testing image is only 5%, just cannot carry out the identification of normal original image.The anti-Gaussian noise ability that this explanation adopts this invention to have.

Figure 44 is the testing image of Gaussian noise intensity 5%, visually very fuzzy, PSNR=13.07dB;

Figure 45 is the people's face original image recognizing, and can correctly identify;

Figure 46 is the watermark of extracting, and can accurately must extract watermark, NC=0.91.

Table 7 is anti-Gauss disturb

Noise intensity (%)	1	2	3	5	10	20	30
								PSNR	19.25	16.42	14.94	13.02	10.57	8.83	7.98
NC	1.00	0.97	0.97	0.90	0.88	0.82	0.70

(2) JPEG compression attack experiment

Adopt image compression quality percentage, as parameter, testing image is carried out to certificate and carry out JPEG compression; Table 8 is testing image JPEG compression experiment data.When compression quality is only 2%, at this moment compression quality is lower, and PSNR=22.21dB still can extract watermark, NC=0.68.

Figure 47 is that compression quality is 5% testing image, and blocking artifact, PSNR=25.12dB have appearred in this figure;

Figure 48 is the original image of identification, can correctly identify;

Figure 49 is the watermark of extracting in testing image, and NC=0.82 can accurately extract watermark.

Table 8 JPEG attacks experiment

Compression quality	2	5	10	20	30	40
							PSNR	22.21	25.12	27.7	29.86	31.15	32.08
NC	0.68	0.91	0.87	1.00	0.97	1.00

(3) medium filtering is attacked experiment

Table 9 is that anti-medium filtering is attacked experiment, and it can be seen from the table, when medium filtering parameter is [7x7], filtering multiplicity is 10 o'clock, the lower PSNR=20.92dB of signal to noise ratio (S/N ratio) of image, but still can record the existence of watermark, NC=0.87.

Figure 50 is that medium filtering parameter is [3x3], the testing image that filtering multiplicity is 10, and it is fuzzy that image has occurred, PSNR=29.18dB;

Figure 51 is the in the situation that of above-mentioned filtering, the original image detecting, at this moment can be correct original image detected;

Figure 52 is the watermark of extracting in testing image, and NC=0.97 can accurately extract watermark.

Table 9 medium filtering is attacked experiment

Claims (1)

1. the anti-attack face identification method that blocks based on DWT-DCT watermark under large data, it is characterized in that: first primitive man's face is carried out to wavelet transformation, pairing approximation coefficient carries out dct transform, before choosing in DWT-DCT matrix of coefficients, 8x8 coefficient is as proper vector, then watermark information and proper vector are associated, realize the embedding of watermark; Then for people's face to be measured, obtain its proper vector, in the related coefficient that calculates people's face to be measured and everyone face proper vector, obtain related coefficient maximal value, according to this value, obtain the facial image of identification, the watermark of embedding; And then according to the proper vector of testing image, extract watermark, obtain watermark related coefficient; Realize the embedding of digital watermarking and the identification of extraction and people's face; This invention comprises that watermark embeds and extracts two large divisions, wherein during the extraction of watermark, realized recognition of face, amounts to six steps:

First: the embedding of watermark

1) by each original facial image F (n) is carried out to wavelet transformation, then pairing approximation coefficient carries out overall dct transform, obtains the proper vector set V (n) of original image;

First successively each original facial image F (n) is carried out to L level wavelet decomposition and obtain ll channel coefficient FA _l(i, j); Because wavelet decomposition progression is higher, the resolving time taking just can be longer, and at this, we select L=1, then to ll channel FA _l(i, j) carries out overall dct transform, obtains DWT-DCT matrix of coefficients FD (i, j),, in DWT-DCT matrix of coefficients FD (i, j), choose front 8 * 8 coefficient FD ₈(i, j), and then to the matrix of coefficients FD selecting ₈(i, j) carries out binary conversion treatment, when coefficient is more than or equal to zero, gets 1, and being less than 0 is to get zero, obtains proper vector V, and main process is described below:

FA _L(i,j)＝DWT2(F(i,j))

FD ₈(i,j)＝DCT2(FA _L(i,j))

V(n)＝BINARY(FD ₈(i,j))

2) utilize cryptography HASH function, generate two-value key sequence Key (n) with watermarked information, realize the embedding of zero watermark;

Key(n)＝V(n)⊕W(n)

Key (n) is that the Hash function conventional by cryptography generates by the proper vector V (n) of all original images and corresponding n digital watermarking W (n); Here the two-value random series that W (n) is 64bit by length forms; Preserve Key (n), while extracting watermark below, will use; By Key (n) is applied for to third party as key, to obtain the right to use and the entitlement of facial image;

Second portion: the identification of people's face and the extraction of watermark

3) obtain the proper vector V' of people's face F' to be measured;

If people's face to be measured is F', through wavelet transformation (DWT), then pairing approximation coefficient to carry out obtaining DWT-DCT matrix of coefficients after overall dct transform be FD ' (i, j), by above-mentioned 1) method, try to achieve the visual feature vector V ' of testing image;

FA _L’(i,j)＝DWT2(F’(i,j))

FD ₈’(i,j)＝DCT2(FA _L’(i,j))

V'＝BINARY(FD ₈’(i,j))

4) calculate the related coefficient NC (n) of the proper vector V' of people's face to be measured and the proper vector V (n) of primitive man's face, carry out the identification of people's face;

Calculate the corresponding n value of related coefficient maximal value of V' and V (n), establish n=k; According to k value, can obtain key K ey (k), identify original facial image is F (k) and the watermark value W (k) that is embedded in F (k), and the normalized correlation coefficient formula of calculated characteristics vector is as follows:

$\mathrm{NC}=\frac{\underset{i}{\mathrm{\Σ}}V\left(i\right)V^{\′}}{\underset{i}{\mathrm{\Σ}}{V}^2\left(i\right)}$

5) utilize the proper vector V ' that is present in third-party two-valued function key sequence Key (k) and people's face to be measured, extract the watermark W' in testing image;

W'＝Key(k)⊕V'

6) calculate the related coefficient between W' and W (k); Calculate the related coefficient of the watermark of extraction and the watermark of embedding, and according to watermark, differentiate the content of testing image.