CN100421123C - Robust two-value text image multi-watermark realizing method - Google Patents

Abstract

Multi-watermark technology based on two-valued text image character belongs to multimedia signal dispose field. The embedding of watermark has the follow steps: (1) carry the DFT transform to the two-valued text image to get a feature vector resisting local geometry attack; (2) use some logic tables formed by this vector and some watermarks to be embedded and take those tables as secret key to prevent the ownership of personally original through registering to third party and pick up the watermark; (3) carry the DFT transform to testing image and pick up the feature vector of it; (4) use the key to pick up the corresponding watermark of the image. This enhances the picking-up speed with stronger ability to resist geometry attack.

Description

A kind of two-value text image multi-watermark implementation method of robust

Technical field

The present invention relates to a kind of multiple digital digital watermark based on the two-value text image feature, is a kind of multimedia data protection method, belongs to field of multimedia signal processing.

Technical background

At internet and multimedia era; digital watermark technology is a kind of effective Digital Media copy-right protection method; but people's research at present is more is embed watermark in gray scale, coloured image and video, audio frequency how, and to how in two-value text image the research of embed watermark less.But in actual applications, the two-value text image ubiquity, as the written material of public document, commercial contract, invoice, educational background, case, fax and scanning etc., they more need copyright protection and prevent to distort sometimes.Compare with embed watermark in general Digital Media, mainly contain following difficult point with conventional method embed digital watermark in two-value text image: the one, suppressible quantity of information is few in the two-value text image.The 2nd, to compare with common gray level image, text image has abundant striped and texture, and each striped and texture have definite implication, like this if come embed watermark, the just very possible implication that changes text itself by these textures and striped are changed.

Present domestic research text digital water mark paper is less.Main method is to come embed watermark by adjusting line space, word space or character feature in spatial domain, and is specific as follows:

The people such as Brassil of Bell Laboratory have proposed to be undertaken by the line space of adjusting text the embedding of watermark abroad.Line space coding is exactly in each page or leaf of text, and the embed watermark information that every interval delegation takes turns utilizes the human eye can't identification for the variation of 1/300 inch of vertical moving, by the change line space, carries out the embedding of watermark.And up and down two line positions adjacent with this row are motionless as the decoding reference, in each row that moves, and the information bit of encoding.

Similar line space coding, people such as Huang propose the word space coding, and in this method, the embedding of watermark is by with a word move left and right in certain row of text, carries out the embedding of watermark.Adjacent two word positions, this speech left and right sides are motionless, and as the decoding reference, when the horizontal range of move left and right was in 1/150 inch, human eye can't be found out, with this embed watermark.

Somebody's use characteristic coding is exactly by changing the feature (as word height, font etc.) of certain word in the document, come embed watermark, and other words not changing, as the decoding reference.

But above in two-value text image the method for embed watermark, all be based on spatial domain, the Information hiding amount is few, poor robustness, can't resist geometric attack, particularly can't resist the local geometric attacking ability, because when two-value text image is subjected to the local geometric attack, the size of local line space, word space or word can change, thereby causes the watermark extracting mistake.But in real life, the printing of text image and scanning produce the local geometric distortion through regular meeting.

In addition, the making of a media product may be that a plurality of links or a plurality of people make, and returns many people to own jointly for the tracking or the proof copyright of carrying out product, therefore usually will embed a plurality of watermark informations in Digital Media.

In the digital watermarking research field, the embedding of two-value text watermark up to now and the embedding of many watermarks are more insoluble problems.Do not find efficient ways at present as yet.

Summary of the invention

This paper has proposed a kind of method that embeds many watermarks in two-value text image, and this method has preferably robustness particularly can resist local geometric to attack, and is blind watermark, does not need the urtext image when extracting watermark.Main thought is; obtain the proper vector that an anti-local geometric is attacked by text image being carried out the DFT conversion; the watermark that utilizes this vector sum to embed generates a corresponding logical table; with this logical table as key (Secret key); by to third-party registration, protect own entitlement to original works.And by watermark corresponding with it in this key-extraction two-value text image to prove own entitlement to Digital Media.

Now as follows to the detailed description of the invention:

First: the embedding of watermark and the generation of corresponding key;

With one group of watermark can representing the binary pseudo-random of copyright information as embedding.If g the watermark that embeds is designated as:

W ^g={ w ^g(j) | w ^g(j)=0,1; 1≤j≤L} is as digital watermarking, and original two-value text image is text512, this image be designated as F={f (i, j) | f (i, j) ∈ R; 1≤i≤N1,1≤j≤N2) }, w ^g(j) and f (i j) represents the grey scale pixel value of g watermark sequence and original image respectively, establishes N1=N2=N.

1), in phase space, obtains a proper vector V (j) that anti-local geometric is attacked of text image by the urtext image is carried out the DFT conversion.

Earlier to former figure F (i j) carries out the DFT conversion, obtain DFT matrix of coefficients FF (i, j), again from DFT matrix of coefficients FF (i, j) in, get the coefficient of Low Medium Frequency part, in the medium and low frequency coefficient, obtain a proper vector V (j) of this image by symbolic operation.This proper vector of evidence has the resist geometric attacks ability.

FF(i，j)＝DFT2(F(i，j))

V(j)＝-Sign(FF(i，j))

2) according to a plurality of watermark W that will embed ^g, g=1, the proper vector V of 2... and text image (j) generates the corresponding two-valued function sequence key relevant with watermark ^g(j), g=1,2.....

${\mathrm{key}}^g\left(j\right)=V\left(j\right)\⊕W^g\left(j\right)$

Key ^g(j) be by the phase characteristic vector V (j) of image and the watermark W that will embed ^g(j), generate by cryptography HASH function commonly used.Preserve key ^gTo use when (j), extracting watermark below.By with Key ^g(j) apply for to the third party as key,, reach the purpose of protection copyright to obtain the entitlement of original works.

Second portion: the extraction of many watermarks;

3) (j) to establish testing image be that (i j), is that (i, j) same as above, a proper vector V ' who obtains testing image (j) for FF ' through the matrix of coefficients that obtains after the DFT conversion to F ' to the proper vector V ' that obtains a resist geometric attacks of text image to be measured.

FF’(i，j)＝DFT2(F’(i，j))

V’(j)＝-Sign(FF’(i，j))

4) utilize two-valued function sequence key ^g(j) and the proper vector V ' of testing image (j), extract corresponding watermark W ^g' (j)

$W^{g^{,}}\left(j\right)={\mathrm{key}}^g\left(j\right)\⊕V^{,}\left(j\right)$

According to the key that when many watermarks embed, generates ^g(j) and the phase characteristic of testing image vector V ' (j), utilize Hash function character can extract corresponding a plurality of watermark W ^g' (j).Again according to W ^g(j) and W ^g' whether (j) degree of correlation is differentiated have watermark to embed.

The present invention and existing text digital watermark relatively have following advantage:

1) resist geometric attacks ability is preferably arranged.This water mark method has robustness preferably to local nonlinearity geometric attacks such as RBA.Because the embedding of watermark is based on the feature of bianry image, and during the conversion of image generation local nonlinearity, the essential characteristic of image does not change.Therefore the present invention has better robustness.

2) embedding of many watermarks does not influence picture quality.Owing to just utilize the feature of image, employing be zero digital watermark, so embed watermark does not influence original image quality, can realize the embedding of many watermarks.

3) extraction of using the present invention to carry out watermark does not need original image and extraction rate fast, can be implemented in the watermark detection of line.

Below we from the explanation of theoretical foundation and test figure:

1) discrete Fourier transform (DFT) (DFT)

In digital watermarking, DFT is widely used, and by this conversion, we can obtain the amplitude subspace and the phason space of image.

Suppose that (m n) is a two-dimensional function in the discrete space to f, and then the positive and negative transformation for mula of two-dimensional discrete Fourier is as follows.

The direct transform formula (DFT) of discrete Fourier:

$F(p,q)=\underset{m=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}f(m,n)e^{-j(2\mathrm{\π}/M)\mathrm{pm}}{e}^{-j(2\mathrm{\π}/N)\mathrm{qn})}$

p＝0，1，Λ，M-1；q＝0，1，Λ，N-1；

Inverse transformation formula (IDFT):

$f(m,n)=\frac{1}{\mathrm{MN}}\underset{p=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{N-1}{\mathrm{\Σ}}}F(p,q)e^{j(2\mathrm{\π}/M)\mathrm{pm}}{e}^{j(2\mathrm{\π}/N)\mathrm{qn})}$

m＝0，1，Λ，M-1；n＝0，1，Λ，N-1；

(p q) is called the discrete Fourier transform (DFT) coefficient to F.

M wherein, n is the spatial domain sampled value; P, q are the frequency field sampled value, and digital picture is represented with the pixel square formation usually, i.e. M=N.

2) an anti-local geometric attack graph choosing as proper vector

Present most digital watermarking algorithm is that watermark directly is embedded in the pixel or conversion coefficient of image.When watermarking images is subjected to slight bird caging, usually cause the unexpected variation of pixel value or transform coefficient values, the watermark of Qian Ruing is just attacked easily like this.And in fact we find that watermarking images does not significantly change for how much.Hayes etc. studies show that for characteristics of image, phase place is more important than amplitude, as long as image is that its similar phase place should not have too big difference.Observe by DFT coefficient (medium and low frequency) great amount of images, find a phenomenon, (projective transformation when an image being carried out multiple local nonlinearity geometric attack, RBA, bird caging etc.), some variations generally can may take place in medium and low frequency coefficient (comprising real part and imaginary part two parts) value size, but the symbol of Low Medium Frequency coefficient remains unchanged substantially, if DFT coefficient (containing real part and imaginary part two parts), on the occasion of using " 1 " expression, negative value or null value are used " 0 " expression, can obtain a binary sequence, with the proper vector of this sequence as image, and this vector has the ability that anti-local geometric is attacked.Promptly, attacks the urtext image when suffering local geometric, this proper vector of this image remains unchanged substantially, with the proper vector degree of correlation of former figure all greater than 0.5, concrete data are shown in Table 1, here when calculating related coefficient, the length of the symbol sebolic addressing of test pattern is the same with the test portion of back, gets 128bit.

Table 1 two-value text image is subjected to local geometric to attack the degree of correlation of front and back proper vector

	Urtext image Text512	Projective transformation is attacked	RBA attacks	Ripple distortion (200%)	Extruding distortion (40%)	Rotation distortion (40 degree)
							PSNR(dB)		5.63	6.26	8.82	6.14	6.36
The degree of correlation	1.0	0.76	0.75	0.87	0.60	0.65

In order to determine that further this symbol sebolic addressing can be used as the proper vector of this image, we are again common different test patterns (seeing Fig. 2 (a)-(f)), carry out the DFT conversion, obtain " the coefficient symbols sequence " of each test pattern earlier, calculate the related coefficient of different test patterns " coefficient symbols sequence " again, result of calculation sees Table 2.Find different test patterns from table 2, the related coefficient of " coefficient symbols sequence " is very little, so the symbol sebolic addressing of DFT can be used as a proper vector of this image.

Related coefficient between the characteristics of image vector of the different two-value text test patterns of table 2

	V1	V2	V3	V4	V5	V6
							V1	1.00	0.09	0.11	0.03	0.17	0.22
V2	0.09	1.00	0.29	0.34	0.29	0.12
							V3	0.11	0.29	1.00	0.26	0.15	0.13
V4	0.11	0.34	0.26	1.00	0.26	0.18
							V5	0.17	0.29	0.15	0.26	1.00	0.01
V6	0.22	0.12	0.13	0.18	0.01	1.00

^*V1-V6 is the characteristics of image vector of corresponding diagram 2 (a)-Fig. 2 (f) respectively.

3) position of watermark embedding and the length of disposable embedding

According to human visual system (HVS, Human Vision System), the Low Medium Frequency signal is bigger to people's visual impact, the main contour feature of representative's image.Therefore the image of our selected image is the symbol of DFT Low Medium Frequency coefficient, the robustness of the size of the definite and original image of the number L of Low Medium Frequency coefficient and the quantity of information of disposable embedding and requirement is relevant, the L value is more little, and the quantity of information of disposable embedding is few more, but robustness is high more.Take all factors into consideration, the length of choosing L here is 128bit.

In sum, we are by to the analysis of DFT coefficient, utilize the symbol sebolic addressing of image Low Medium Frequency coefficient to obtain the characteristics of image vector of a bianry image, and this vector have stronger anti-local geometric attacking ability;

Description of drawings

Fig. 1 (a) is original two-value text image.

Fig. 1 (b) is the image of attacking through projection.

Fig. 1 (c) is the image of attacking through RBA.

Fig. 1 (d) is the image through the ripple distortion.

Fig. 1 (e) is the image through the extruding distortion.

Fig. 1 (f) is the image through the rotation distortion.

Fig. 2 (a) is standardized test chart Text_1.

Fig. 2 (b) is standardized test chart Text_2.

Fig. 2 (c) is standardized test chart Text_3.

Fig. 2 (d) is standardized test chart Text_4.

Fig. 2 (e) is standardized test chart Text_5.

Fig. 2 (f) is standardized test chart Text_6.

Fig. 3 (a) is the two-value text image that does not add when disturbing.

Fig. 3 (b) is the grid image that does not add when disturbing.

Fig. 3 (c) is the many watermark detection of bianry image that do not add when disturbing.

Fig. 4 (a) is the bianry image that projective transformation is arranged.

Fig. 4 (b) is the grid image that projective transformation is arranged.

Fig. 4 (c) is the many watermark detection of bianry image that projective transformation is arranged.

Fig. 5 (a) is the bianry image that is not subjected to the band grid of RBA attack.

Fig. 5 (b) is the watermarking images (stirmark4.0) that the band grid of RBA attack is arranged.

Fig. 5 (c) is the many watermark detection of bianry image that have RBA to attack.

Fig. 6 (a) is the watermarking images (distortion quantity is 200%) that the ripple distortion is arranged.

Fig. 6 (b) is the grid that the ripple distortion is arranged.

Fig. 6 (c) is the many watermark detection of bianry image that the ripple distortion is arranged.

Fig. 7 (a) is the watermarking images (distortion quantity is 40%) that the extruding distortion is arranged.

Fig. 7 (b) is the grid that the extruding distortion is arranged.

Fig. 7 (c) is the many watermark detection of bianry image that the extruding distortion is arranged.

Fig. 8 (a) is many watermarking images (distortion quantity is 30%) that the sphere distortion is arranged.

Fig. 8 (b) is the grid that the sphere distortion is arranged.

Fig. 8 (c) is many watermark detection that the sphere distortion is arranged.

Fig. 9 (a) is the watermarking images (distortion angle is 40 degree) that the rotation distortion is arranged.

Fig. 9 (b) is the grid that the rotation distortion is arranged.

Fig. 9 (c) has the image multi-watermark of rotation distortion to detect.

Figure 10 (a) is the image (pond ripple type, distortion quantity is 30%) that the ripples distortion is arranged.

Figure 10 (b) is the grid that the ripples distortion is arranged.

Figure 10 (c) has the image multi-watermark of ripples distortion to detect.

Embodiment

Emulation platform is Matlab6.1, use 1000 groups of independently binary pseudo-random (value is+1 or-1), every group of sequence length is 128bit, and in these 1000 groups of data, we appoint four watermark sequences that extract four groups (we select the 200th, 400,600 and 800 group) conduct embedding here.Two-value text image is elected Text512 (512x512x2) as, sees Fig. 3 (a), original image be expressed as F (i, j), wherein the DFT transform coefficient matrix of 1≤i≤512,1≤j≤512 correspondences be FF (i, j), 1≤i≤512,1≤j≤512 wherein.Carry out symbolic operation by coefficient, form the proper vector of image its Low Medium Frequency part.After detecting watermark W ', judged whether that by calculating normalized correlation coefficient NC (Normalized Cross Corrclation) watermark embeds.

Normalized correlation coefficient NC (Normalised Cross-Correlation) wherein

$\mathrm{NC}=\frac{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{W}_{(i,j)}{W^{\text{'}}}_{(i,j)}}{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{W^2}_{(i,j)}}$

NC is worth big I according to this and reflects whether watermark exists as the output of watermark detector.Watermark text image when not adding external disturbance is seen Fig. 3 (a), clear picture, and corresponding grid is seen Fig. 3 (b), Fig. 3 (c) is seen in the output of watermark detector, can see, obviously detects the existence of 4 watermarks, and NC1=NC2=NC3=NC4=1.0.

Judge the anti-local geometric attacking ability of this watermark below by test.

1) projective transformation

Projective transformation is a kind of local nonlinearity conversion.

Fig. 4 (a) is through the watermark text image after the projective transformation, PSNR=5.63dB at this moment, and numerical value is lower;

Fig. 4 (b) is that corresponding grid changes;

Fig. 4 (c) is watermark detector response, can obviously detect the existence of four watermarks from Fig. 4 (c), NC1=0.765 at this moment, NC2=0.765, NC3=0.772, NC4=0.764.

2) random distortion is attacked (RBA, Random bending attack)

Stirmark is in the digital watermarking research, the software that detects relatively more commonly used.Here using this software (Stirmark4.0) that text image is carried out a kind of random distortion attacks.For observing conveniently, we add grid in watermarking images.

Fig. 5 (a) is the original watermark image of band grid;

Fig. 5 (b) is an image (LATEST RNDDIST_1.1) of attacking through random distortion, PSNR=6.26dB;

Fig. 5 (c) is the testing result of watermark, can obviously detect the existence of four watermarks, NC1=0.749 at this moment, NC2=0.751, NC3=0.750, NC4=0.748.

3) other common bird caging is attacked

For the convenience and the repeatability of testing, we use the distortion module of Adobe photoshop6.0 to realize following common bird caging attack.

A) ripple distortion:

Fig. 6 (a) carries out the ripple distortion to watermarking images, and distortion quantity is (200%), at this moment watermark

The PSNR=8.823dB of image;

Fig. 6 (b) is corresponding grid image;

Fig. 6 (c) watermarking detecting results.Can obviously detect the existence of four watermarks, NC1=0.874 at this moment, NC2=0.873, NC3=0.875, NC4=0.878.By table 3 can see when distortion quantity up to 800% the time, still can detect the existence of many watermarks.So this paper watermarking algorithm has stronger anti-ripple distortion ability.

The anti-ripple torsion test of table 3 watermark data

Distortion quantity (%)	100	200	400	600	800
						PSNR(dB)	11.764	8.823	6.929	6.380	6.070
NC1	0.937	0.875	0.889	0.781	0.796
						NC2	0.937	0.873	0.889	0.779	0.795
NC3	0.937	0.874	0.891	0.780	0.796
						NC4	0.937	0.878	0.892	0.794	0.796

B) extruding distortion:

Fig. 7 (a) is the watermarking images of distortion of being squeezed, and distortion quantity is 40%, the PSNR=6.14dB of watermarking images at this moment, and signal to noise ratio (S/N ratio) is lower;

Fig. 7 (b) is corresponding grid image;

Fig. 7 (c) is a watermarking detecting results.Can obviously detect the existence of many watermarks, NC1=0.608 at this moment, NC2=0.609, NC3=0.611, NC4=0.609.Can see when distortion quantity is 50% by table 4, still can detect the existence of many watermarks, so this watermarking algorithm has robustness preferably to the extruding distortion.

The anti-extrusion torsion test data of table 4 watermark

Distortion quantity (%)	10	20	30	40	50
						PSNR(dB)	49.60	6.64	6.292	6.14	5.95
NC1	1.000	0.811	0.686	0.608	0.514
						NC2	1.000	0.814	0.686	0.609	0.514
NC3	1.000	0.813	0.686	0.611	0.517
						NC4	1.000	0.815	0.688	0.609	0.514

C) sphere distortion:

Fig. 8 (a) carries out sphere distortion to watermarking images, and distortion quantity is 30%, the PSNR=5.33dB of watermarking images at this moment, and signal to noise ratio (S/N ratio) is lower;

Fig. 8 (b) is corresponding grid image;

Fig. 8 (c) is a watermarking detecting results.Can obviously detect the existence of many watermarks, NC1=0.609, NC2=0.609, NC3=0.608, NC4=0.608.Can see when distortion quantity is 40% by table 5, still can detect the existence of many watermarks.

The anti-sphere torsion test of table 5 watermark data

Distortion quantity (%)	5	10	20	30	40
						PSNR(dB)	8.59	6.68	6.043	5.62	5.33
NC1	0.939	0.887	0.791	0.609	0.514
						NC2	0.936	0.874	0.780	0.609	0.518
NC3	0.939	0.876	0.780	0.607	0.514
						NC4	0.936	0.877	0.784	0.607	0.513

D) rotation distortion:

Fig. 9 (a) is rotated distortion to watermarking images, and the anglec of rotation is 40 degree, the at this moment PSNR=6.367dB of watermarking images;

Fig. 9 (b) is corresponding grid image;

Fig. 9 (c) is a watermarking detecting results.Can obviously detect the existence of many watermarks, NC1=0.655, NC2=0.661, NC3=0.663, NC4=0.655.By table 6 can see when distortion angle be 50 when spending, still can detect the existence of a plurality of watermarks, so this watermarking algorithm has stronger anti-rotation distortion ability.

The anti-rotation of table 6 watermark torsion test data

Distortion angle (degree)	10	20	30	40	50
						PSNR(dB)	7.92	6.99	6.59	6.36	6.33
NC1	0.905	0.843	0.781	0.655	0.594
						NC2	0.904	0.842	0.782	0.661	0.599
NC3	0.912	0.849	0.782	0.663	0.609
						NC4	0.905	0.843	0.780	0.655	0.592

E) ripples distortions (pond ripple):

Figure 10 (a) is that watermarking images carries out ripples distortion (parameter is chosen as: distortion quantity is 30%, rises and falls 5%).At this moment the PSNR=6.66dB of watermarking images, signal to noise ratio (S/N ratio) is lower.

Figure 10 (b) is corresponding grid image.

Figure 10 (c) is many watermarking detecting results, can obviously detect the existence of four watermarks, NC1=0.534, NC2=0.541, NC3=0.531, NC4=0.529.Can see when distortion angle is 40% by table 7, still can detect the existence of watermark, so this watermarking algorithm has stronger anti-ripples distortion ability.

The anti-ripples torsion test of table 7 watermark data

Ripples distortion quantity (%)	5	10	20	30	40
						PSNR(dB)	8.78	7.57	6.828	6.656	6.438
NC1	0.858	0.749	0.608	0.534	0579
						NC2	0.859	0.751	0.607	0.541	0581
NC3	0.859	0.750	0.608	0.531	0.576
						NC4	0.860	0.757	0.613	0.529	0.587

By top campaign, learn that the present invention has stronger robustness to local geometric attack, and the employed commercial watermark software Digimarc of Adobe Photoshop6.0 can't carry out the embedding and the extraction of watermark to bianry image.

Claims (1)

1. the two-value text image multi-watermark implementation method of a robust, it is characterized in that: based on the two-value text image feature, digital watermark and cryptography are organically combined, realize that many watermarks of bianry image embed and extract, this digital watermark method amounts to four steps altogether in two sub-sections:

First is that many watermarks embed:

1) former figure is carried out the DFT conversion, in DFT Low Medium Frequency coefficient, obtain the proper vector V (j) of an anti-local geometric attack according to the symbol sebolic addressing of these Low Medium Frequency coefficients;

2) a plurality of watermark Wg (j) that utilize the Hash function and will embed, g=1,2 ...; Obtain corresponding two-valued function sequence keyg (j), g=1,2 ...;

$\mathrm{keyg}\left(j\right)=V\left(j\right)\⊕\mathrm{Wg}\left(j\right);$ g＝1，2，...；

Preserve keyg (j), will use when extracting watermark below, by keyg (j) is applied for to the third party as key, to obtain entitlement former figure;

Second portion is many watermark extracting:

3) testing image is carried out the DFT conversion; In the DFT coefficient, a proper vector V ' who goes out testing image according to the symbol extraction of Low Medium Frequency coefficient (j);

4) utilize Hash function character to extract watermark, ${\mathrm{Wg}}^{,}\left(j\right)=\mathrm{keyg}\left(j\right)\⊕V^{,}\left(j\right);$ Wg (j) and Wg ' (j) are carried out degree of correlation test, determine the entitlement of image.

Images