CN1975780A - Robust digital watermark inserting and detecting method based on supporting vector - Google Patents

Abstract

The invention is a robust digital watermark embedding and detecting method based on the supported vector machine, which belongs to multi-media information security area. The invention consists of watermark embedding and watermark detecting. In the process of watermark embedding, a randomly generated referencing watermark is embedded to host media data together with real watermark; in the process of watermark detecting, the well trained supported vector machine can detect watermark with embedded watermark key. The supported vector machine is trained to study the attack with natural image characteristics, referencing watermark data and embedded watermark position data.

Description

Robust digital watermark based on support vector machine embeds and detection method

Technical field

The present invention relates to a kind of method of multi-media information security technical field, specifically is that a kind of robust digital watermark based on support vector machine embeds and detection method.

Background technology

Along with popularizing and development of infotech and internet, increasing digital information is propagated by network, numerical information duplicate and propagation becomes more and more simpler.Digital watermark technology effectively to be protected and is grown up specially the copyright of digital product in order to solve.Digital watermark technology is compared with traditional encipherment protection; an important advantage is that it provides a seamless connection between user and digital product; make the user not only can visit easily and be subjected to the Digital Media that watermark is protected; simultaneously, still can provide rights and interests to a certain degree to protect the owner of copyright.

Current, digital watermark technology can be divided into two big classes simply: based on the technology of spatial domain with based on the technology of transform domain.The spatial domain method reaches the purpose of embed watermark by the pixel value of direct modification original image, but the watermarking algorithm of spatial domain is easy to be subjected to the puzzlement of daily Flame Image Process such as compression of images, conversion, and robustness is not very strong.Current robust digital watermark technical research mainly concentrates in the transform domain.According to the conversion difference that adopts, the transform domain digital watermark can be divided into following a few class: (i) based on the digital watermark technology of dct transform.At present, JPEG and mpeg standard have all adopted dct transform as the basis, so much all be based on dct transform domain based on the transform domain digital watermark; (ii) based on the robustness digital watermark of DFT conversion.The image/video watermark is easy to be subjected to the puzzlement of various geometric attacks, utilizes some unchangeability of this conversion based on the digital watermark of Fourier-Mellin conversion, to a certain extent geometric attack is had certain robustness.(iii) based on the digital watermark technology of wavelet transformation.Because wavelet transformation has good spatio-temporal frequency characteristic, can realize the decomposition and the transmission of multiresolution, (2000, MPEG-7 has adopted wavelet transformation to follow-on compressed encoding standard JPE, so proposed a series of wavelet field digital watermark technology.For digital watermark technology can well be used in actual environment, robustness is a very necessary condition.At present, there is robustness watermarking algorithm to be suggested successively based on genetic algorithm and nerual network technique.And support vector machine is as a sorter based on Statistical Learning Theory, and its adopts the optimal classification lineoid implementation structure principle of minimization risk of class interval maximum, therefore has results of learning preferably.Especially under the condition of small sample, have the incomparable superiority of method such as neural network.Simultaneously, the non-linear discriminating analysis method has also obtained scholar's concern in recent years, and is widely used in research fields such as recognition of face and data mining, has showed extremely strong classification generalization ability.

Find through literature search prior art, P.T.Yu etc. have delivered " Digital watermarking based onneural networks for color images " (based on the digital watermark technology of neural network) on 663 pages of the 3rd phases of volume " Signal Processing " (signal Processing) calendar year 2001 the 81st, propose in this article, with Application of Neural Network in the embedding and testing process of digital watermarking, concrete grammar is: watermark signal is modulated in the blue component of image, when watermark detection, make full use of the learning ability of neural network, the network weight that utilization trains carries out the detection that embed watermark has or not.Its deficiency is, the introducing of neural network can solve the robustness problem of watermark detection to a certain extent, but because the training process of neural network is slow, simultaneously, in the embed watermark off-capacity, the result of training can not effectively guarantee the robustness of watermarking algorithm.

Summary of the invention

The objective of the invention is to make full use of the strong study generalization ability of support vector machine, provide a kind of robust digital watermark to embed and detection method based on support vector machine, the testing process of watermark is combined with the sorting technique in the data mining, can embed under the capacity prerequisite in less watermark, realize quick, robust detection effectively embed watermark.

The present invention is achieved by the following technical solutions, the present invention includes two processes of detection of the embedding and the watermark of watermark, in the telescopiny of watermark, a reference watermark that produces at random and real watermark data are embedded in host's media data as watermark together, take into full account human vision property; In the testing process of watermark, make full use of the feature of natural image own, be that pass between each neighbor ties up to after the multiple attack of experience, this major part that concerns can both effectively remain, utilize the data of reference watermark and embedding reference watermark position, can learn its suffered attack by the training support vector machine, utilize the support vector machine and the watermark that train to embed secret key, can realize robust detection watermark signal.

In natural image, has very strong correlativity between each neighbor, after watermark signal is embedded in corresponding spatial domain or the transform domain, learning process by machine learning method, this relation can be remembered by machine learning method, utilize strong study of support vector machine and generalization ability, can realize robust detection watermark.In order to guarantee the robustness of watermarking algorithm, the telescopiny of digital watermark signal is considered the visual characteristic of human eye fully, embeds the watermark signal of varying strength for the zones of different of image; In the leaching process of digital watermarking, use support vector machine is carried out watermark as the instrument of watermark detection detection.Test findings shows that the present invention can experience under the condition of attacks such as multiple signal Processing and geometric warping at image, successfully carries out the extraction of embed watermark signal.

The embedding of described watermark is specially: utilize the above-mentioned correlativity between the neighbor in the natural image, reference watermark and real watermark signal are embedded in the blue component of image by the position of selecting at random.The effect of the reference watermark here is, utilizes the relation between the support vector machine study image pixel in the testing process of watermark, thereby can remember the sort of relation that image was kept before living through various attack.What will note here a bit is, in the telescopiny of watermark, should avoid reference watermark and real watermark is embedded into marginal point or two reference point are overlapped, otherwise, can cause the erroneous judgement in the watermark detection process.

The detection of described watermark is specially: at first be used in key identical in the telescopiny and generate the reference watermark signal, and generate the embedded location of watermark according to the another one key.According to selected watermark embedded location, determine the relation data between each reference watermark embedded location neighbor pixel, constitute the training vector of support vector machine, utilize this data acquisition and reference watermark, just can train corresponding watermark and extract support vector machine.According to selected real watermark embedded location, can determine the relation data of each effective watermark data neighbor pixel then, thereby constitute the vector data set that watermark is extracted.The support vector machine that utilization trains just can realize the detection to watermark signal.

The effect that the present invention is useful: the present invention is based on the strong study generalization ability of support vector machine, can under condition of small sample, train support vector machine quickly and efficiently, carry out real-time watermark embedding and robust detection in colour/gray level image thereby be implemented in.With comparing based on neural network method that P.T.Yu proposes, the present invention has better watermark detection efficient and robustness.From comparative test result as can be seen, the present invention is not attacking, under the multiple attack conditions such as fuzzy, mosaic, brightness and contrast's adjustment, filtering, change of scale, has the robustness and the efficient (specifically quantize comparative result and see Table 1) that obviously are better than based on neural network algorithm.

Description of drawings

Fig. 1 is a watermark embed process block diagram of the present invention.

In Fig. 1, utilize two keys to generate the position that reference watermark and watermark embed respectively, guarantee the security of watermaking system and guarantee training precision in the training process of support vector machine.

Fig. 2 is a watermark extraction process block diagram of the present invention.

In Fig. 2, utilize given reference watermark and train support vector machine from the data that the reference watermark embedded location extracts, realize the robustness of embed watermark signal is extracted.

The result that Fig. 3 embeds for the part watermark

Fig. 3 (a) is the Lena figure of original embed watermark; Fig. 3 (b) is the Lena figure behind the embed watermark; Fig. 3 (c) is original watermarking images; The watermark of Fig. 3 (d) for from the image of embed watermark, extracting

Fig. 4 is that partial results is extracted in the watermark of image after under attack

Fig. 4 (a) is the result of Mosaic after handling; Fig. 4 (b) is the result after the Fuzzy processing; Fig. 4 (c) is the result behind distortion 3 degree; Fig. 4 (d) is the result behind the interpolation noise; The adjusted result of Fig. 4 (e) brightness and contrast; Fig. 4 (f) image cut falls the result after 25%.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: present embodiment has provided detailed embodiment and process being to implement under the prerequisite with the technical solution of the present invention, but protection scope of the present invention is not limited to following embodiment.

If the image I of watermark to be embedded is 24 true color pictures of M * N, dividing other red, green, blue colouring component of remembering it is R, G and B.The telescopiny block diagram of watermark as shown in Figure 1.

1. the telescopiny of watermark:

A. watermark signal generates.According to given key K ₁Generating length randomly is the reference watermark signal Rf=r of K ₁r ₂... .r _KThe real valid data of watermark are that width of cloth size is the two-value marking pattern of m * n.Then the reference watermark signal is connected together with real watermark signal, forms watermark sequence to be embedded:

W＝r ₁r ₂....r _Ks ₁s ₂....s _L＝w ₁w ₂....w _K+L

And further it is modulated into bipolar signal sequence, i.e. w _i∈ 1 ,+1}.

B. watermark signal embeds.The key K that provides according to the copyright owner ₂Produce a mutually different binary sequence p at random _t=(i _t, j _t), t=1,2 ...., K+L, wherein 1＜i _t＜M, 1＜j _t＜N, the position sequence that this sequence will embed as watermark.

Watermark signal is embedded in the spatial domain, several selections are arranged altogether, promptly watermark signal can be embedded in certain component of RGB individually, also can be embedded in the combination of some component.Experimental result shows, is easier to form bright spot or dim spot for the modification ratio of red and green component, and the watermark strength of embedding is very limited, and robustness is poor.Because people's vision will be hanged down a bit relatively to the blue component sensitivity, so under the certain intensity of being embedded in of blue component, human eye is difficult for discovering the existence of watermark signal noise, so select here watermark signal is embedded in the blue component.

The blue component of coloured image is released separately from image, be designated as B.Respectively to each selected embedded location p _tWith an embedding position information w _t, operate as follows:

With each position of choosing is cruciform window of central configuration, makes the average of each pixel on this window:

$A_{p_t}=(\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}B(i_t+l,j_t)+\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}B(i_t,j_t+l)2B(i_t,j_t))/8,$

Press the method embed watermark data of following formula then:

$\tilde{B}(i_t,j_t)=A_{p_t}+{\mathrm{\α}}_t(1-2w_t),$ α wherein _tIntensity factor for embed watermark.

In digital watermarking system early, watermark signal directly is embedded in the host signal with certain constant intensity, in order to make the watermark signal that embeds have robustness preferably, watermark signal must be embedded in the host signal with maximum intensity.When embedment strength was big, be easy to can cause visual distortion in some part, and when intensity hour, the robustness of watermark can be poor, and therefore for piece image, the embedded mode of using constant intensity can not make the embed watermark maximization effectively.By finding for the human vision Study of model, people's vision is different for the susceptibility of different luminance areas, such as in darker zone, being easy to and being discovered as minor modifications, but for the higher zone of those brightness, such modification human eye is difficult to perceive, that is to say that human vision is low for the low zone of the area sensitive specific luminance of high brightness.Amid all these factors,, make the watermark embed strength maximum,, select the embedment strength of watermark by following mode here in order to make full use of human visual theory $\mathrm{\α}=\mathrm{\β}{L}_{p_t}$ , wherein β is the constant of a control watermark strength, $L_{p_t}=0.299R_{p_t}+0.587G_{p_t}+0.114B_{p_t}$ Be that coloured image is at select location p _tBrightness value.

C. last, with the blue component behind the embed watermark

Be recombined into together with red and green component, obtain the image behind the final embed watermark.

2. watermark detection process:

The given image of waiting to extract watermark, the extraction process block diagram of watermark as shown in Figure 2.

Identical key K when a. utilizing with embed watermark ₂Calculate two tuple sequence { ρ of watermark embedded location _t=(i _t, j _t) _T=1...K+L, simultaneously, utilize key K ₁Generate reference watermark signal Rf=r randomly ₁r ₂... r _K

B. given coloured image to be detected extracts blue component wherein, is designated as B.According to given embedded location sequence p _t=(i _t, j _t), t=1,2 ...., K+L is to each embedded location p _t, the training of structure support vector machine and detection sample set, and training support vector machine.

Each select location in blue component, calculate on this position the poor of other pixel averages on the pixel value and cross window:

$d_{\mathrm{ij}}=\tilde{B}(i,j)-A^{\′}(i,j),$

Wherein $A^{\′}(i,j)=(\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}\tilde{B}(i+l,j)+\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}\tilde{B}(i,j+l)-2\tilde{B}(i,j))/8,$ Be so that (i, j) position is an average of removing all pixels behind the center on the cross window at center.

Telescopiny by watermark can see that the pixel value of each watermark embedded location all can be higher or low than other pixel averages on its corresponding cross window, can utilize this corresponding relation to extract the watermark of embedding.Usually, image to be detected has all experienced Flame Image Process or attack more or less, detects error thereby can produce some inevitably, in order to reduce the detection error that attack brings, selects to use support vector machine and assists the detection of carrying out watermark.Because support vector machine is a good sorter of setting up according to the theory of statistical learning, theoretical and experiment shows that it has good extensive popularization ability, only needs training sample seldom usually, just can obtain training result preferably.The detection method of using support vector machine is different from other detection methods, after the various attack of image experience, training by support vector machine, it can remember the relativeness of the position of original embed watermark with point around him, thereby can under the condition of experience various attack, correctly carry out watermark extracting.

The data set that is defined as follows:

$D={\{d_{i_t-2,j_t},d_{i_t-1,j_t},d_{i_t,j_t},d_{i_t+1,j_t},d_{i_t+2,j_t},d_{i_t,j_t-2},d_{i_t,j_t-1},d_{i_t,j_t+1},d_{i_t,j_t+2},r_t\}}_{t=1...K}={\{D_t,r_t\}}_{t=1...K}$

R wherein _tBe the reference watermark position that adds in order to carry out the support vector machine training, perhaps also can regard the target label value of training as, D _tConveniently write as the data of vector form in order to represent.Use given training dataset D, just can train and supported vector machine as follows:

$f\left(x\right)=\mathrm{sign}[\underset{t=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\λ}}_t{r}_t\mathrm{Ker}(x,D_t)+b],$

Wherein Ker is the kernel function of selecting, λ _tBe the coefficient that obtains after the training, b is the deviation of training.

C. last, structure extracts the detection data set of watermark:

$E={\{d_{i_t-2,j_t},d_{i_t-1,j_t},d_{i_t,j_t},d_{i_t+1,j_t},d_{i_t+2,j_t},d_{i_t,j_t-2},d_{i_t,j_t-1},d_{i_t,j_t+1},d_{i_t,j_t+2}\}}_{t=K+1...K+L}={\left\{D_t\right\}}_{t=K+1...K+L}$

Support vector machine that application training is good and detection data set E, just the digital watermark signal that can obtain extracting is expressed as follows:

$\stackrel{\‾}{S_i}=f\left(D_{K+i}\right),i=\mathrm{1,2},...,L$

The watermark signal that is drawn into is compared with the watermark signal that provides, and the existence that can verify watermark signal whether.For the existence that determines watermark signal more objectively whether, can also pass judgment on the existence of watermark objectively, be defined as follows by the index--bit rate of mistake (BER, Bit Error Rate)--that quantizes:

$\mathrm{BER}=\frac{\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\stackrel{\‾}{s_i}\⊕s_i}{L},$ Wherein _ be xor operator.

Experimental result:

Fig. 3 has shown original Lena figure and had added the Lena figure of watermark that the embedding of watermark has good invisibility as can be seen from this figure.Do not having under the condition of attacking, watermark can be extracted (Fig. 3 (d)) without any error ground.

For the robustness of illustration method, Fig. 4 (a-f) is the watermark figure that the image of embed watermark extracts after a series of Flame Image Process of experience, and provides corresponding error rate BER (Bit Error Rate).Wherein Fig. 4 (a) is the result of image after handling through Mosaic, the watermark BER=0.036 of extraction; Fig. 4 (b) is the result of image after by Fuzzy processing, the watermark BER=0.005 of extraction; Fig. 4 (c) is the result behind distortion 3 degree, the watermark BER=0.108 of extraction; Fig. 4 (d) is the result behind the interpolation noise, the watermark BER=0.046 of extraction; The adjusted result of Fig. 4 (e) brightness and contrast, the watermark BER=0.001 of extraction; Fig. 4 (f) image cut falls the result after 25%, the watermark BER=0.0055 of extraction.

Under the various attack condition, carried out some relatively with the method for proposition and based on the method for neural network method and Kutter, comparative result is as shown in table 1.From the listed result of form as can be seen, when under same experimental conditions, present embodiment has better watermark and extracts performance.Present embodiment especially to adding attacks such as noise, image adjustment, scalloping, has better extraction effect.

In the method for the invention, make full use of support vector machine better learning generalization ability, remember the various attack process that image experiences, thereby can realize the robustness of watermark signal is extracted, simultaneously in the extraction process of watermark, do not need original view data, have good convenience.

Experimental result under the different attack conditions of table 1.

Attack type	PSNR (dB)	Mistake bit rate (BER)
					The method that proposes	The method of Kutter	The method of Yu
		Do not attack	41.47	0				0.04	0.00375
Fuzzy	38.10				0.005	0.04125	0.01
		Mosaic	30.83	0.03625				0.08125	0.05375
The brightness and contrast improves 70%	7.53				0.01	0.58875	0.03
		Shear 25%	11.95	0.055				0.20625	0.0925
Jpeg compresses (90%)	34.03				0.1412	0.1525	0.16875
		Jpeg compresses (80%)	33.00	0.2475				0.315	0.32625
Jpeg compresses (70%)	31.55				0 0.46625	0.30875	0.44625
		Filtering	25.85	0.0102				0.0775	0.0325
Distortion (15 °)	20.71				0.1962	0.4351	0.2343
		Distortion (50 °)	17.48	0.2825				0.5187	0.3125
Scaling (50%)	28.19				0.0275	0.155	0.0325
		Rotation (15 °)	8.7533	0.1863				0.2866	0.1872
Add noise	18.144				0.04625	0.07375	0.0525

Claims (4)

1, a kind of robust digital watermark based on support vector machine embeds and detection method, it is characterized in that: two processes of detection that comprise the embedding and the watermark of watermark, the embedding of described watermark, at first generate a reference watermark signal at random according to given user key, then reference watermark and real watermark signal are merged into the watermark sequence to be embedded of an integral body, according to given user key, generate a watermark embedded location sequence at random, at last on selected embedded location, this watermark sequence to be embedded is embedded in the blue component of image, in the telescopiny of watermark, utilize the vision system characteristic of human eye, reconcile intensity in each position watermark embedding of image;

The detection of described watermark, at first be used in user key used in the telescopiny and generate the reference watermark signal, and according to another one key generation watermark embedded location sequence, then on selected watermark embedded location, determine the relation data between each reference watermark embedded location neighbor pixel, constitute the training vector of support vector machine; Secondly, the data acquisition and the reference watermark that utilize this training vector to constitute train the support vector machine that watermark is extracted; Once more,, determine the relation data of each effective watermark embedded location neighbor pixel, form the vector set that watermark is extracted at each selected real watermark embedded location; Utilize the support vector machine and the watermark that train to extract the vector set at last, realize detection watermark signal.

2, the robust digital watermark based on support vector machine according to claim 1 embeds and detection method, it is characterized in that, and the embedding of described watermark, the specific implementation step is as follows:

A. watermark signal generates: according to given key K ₁Generating length randomly is the reference watermark signal Rf=r of K ₁r ₂... .r _K, the real valid data of watermark are that width of cloth size is the two-value marking pattern of m * n, the stretching one-tenth one-dimensional data of marking pattern data need be connected together the reference watermark signal then with real watermark signal, form watermark sequence to be embedded:

W＝r ₁r ₂....r _Ks ₁s ₂....s _L＝w ₁w ₂.....w _K+L

And further it is modulated into bipolar signal sequence, i.e. w _i∈ 1 ,+1};

B. watermark signal embeds: the key K that provides according to the copyright owner ₂Produce a mutually different binary sequence p at random _t=(i _t, j _t), t=1,2 ...., K+L, wherein 1＜i _t＜M, 1＜j _t＜N, the position sequence that this sequence will embed as watermark is embedded into watermark signal in the blue component of image, and the blue component of coloured image is released separately from image, is designated as B, respectively to each selected embedded location p _tWith an embedding position information w _i, operate as follows:

With each position of choosing is cruciform window of central configuration, makes the average of each pixel on this window:

$A_{\mathrm{pt}}=(\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}B(i_t+l,j_t)+\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}B(i_t,j_t+l)-2B(i_t,j_t))/8,$

Press the method embed watermark data of following formula then:

$\tilde{B}(i_t,j_t)=A_{\mathrm{pt}}+{\mathrm{\α}}_i(1-{2w}_i),$ α wherein _iIntensity factor for embed watermark;

Select the embedment strength of watermark $\mathrm{\α}={\mathrm{\βL}}_{\mathrm{pt}},$ Wherein β is the constant of a control watermark strength,

$L_{\mathrm{pt}}=0.299R_{\mathrm{pt}}+0.587G_{\mathrm{pt}}+0.114B_{\mathrm{pt}}$ Be that coloured image is at select location p _tBrightness value;

C. last, with the blue component behind the embed watermark

Be recombined into together with red and green component, obtain the image behind the final embed watermark.

3, the robust digital watermark based on support vector machine according to claim 1 embeds and detection method, it is characterized in that, and the detection of described watermark, the specific implementation step is as follows:

Key K when a. utilizing embed watermark ₂Calculate watermark embedded location sequence { ρ _t=(i _t, j _t) _T=1, _K+L, simultaneously, utilize key K ₁Generate reference watermark signal Rf=r randomly ₁r ₂... .r _K

B. given coloured image to be detected extracts blue component wherein, is designated as B, according to given embedded location sequence p _t=(i _t, j _t), t=1,2 ...., K+L is to each embedded location p _t, the training of structure support vector machine and detection sample set, and training support vector machine;

Each select location in blue component, calculate on this position the poor of other pixel averages on the pixel value and cross window:

$d_{\mathrm{ij}}=\tilde{B}(i,j)-A^{\′}(i,j),$

Wherein $A^{\′}(i,j)=(\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}\tilde{B}(i+l,j)+\underset{l=-2}{\overset{2}{\mathrm{\Σ}}}\tilde{B}(i,j+l)-2\tilde{B}(i,j))/8,$ Be with (i, j) position is an average of removing all pixels behind the center on the cross window at center;

The data set that is defined as follows:

$D=\{d_{i_t}-{2j}_t,d_{i_t}-1j_t,d_{i_{t,+j_t}},d_{i_t+{1j}_t},d_{i_t}+{2j}_t,d_{i_t,j_t-2},d_{i_t,j_t-1,}{d}_{i_t,j_t+1},d_{i_t,j_t+2},r_t{\}}_{t=1\·\·\·K}=\{D_t,r_r{\}}_{t=1\·\·\·K}$

R wherein _tBe the reference watermark position that embeds in order to carry out the support vector machine training, as the target label value of training, D _tConveniently write as the data of vector form in order to represent; Use given training dataset D, training and supported vector machine are as follows:

$f\left(x\right)=\mathrm{sign}[\underset{t=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\λ}}_t{r}_t\mathrm{Ker}(x,D_t)+b],$

Wherein Ker is the kernel function of selecting, λ _tBe the coefficient that obtains after the training, b is the deviation of training.

C. last, structure extracts the detection data set of watermark:

$E=\{d_{i_t-{2j}_t},d_{i_t-{1j}_t},d_{i_t,j_t},d_{i_t+{1,j}_t},d_{i_t+{2j}_t},d_{i_t,j_t-2},d_{i_t,j_t-1},d_{i_t,j_t+1},d_{i_t,j_t+2}{\}}_{t=K+1...K+L}=\{D_t{\}}_{t=K+1,...K+L}$

Support vector machine that application training is good and detection data set E, just the digital watermark signal that can obtain extracting is expressed as follows:

s _i＝f(D _K+i)，i＝1，2，....，L

The watermark signal that is drawn into is compared with the watermark signal that provides, and whether the existence of checking watermark signal.

4, the robust digital watermark based on support vector machine according to claim 3 embeds and detection method, it is characterized in that whether the existence of described checking watermark signal to be, is specially:

Pass judgment on the existence of watermark objectively, be defined as follows by the index--mistake bit rate--that quantizes:

$\mathrm{BER}=\frac{\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\stackrel{\‾}{s_i}\⊕s_i}{L},$

Wherein _ be xor operator.

Images