CN101039371A - Novel method of digital watermarking for protecting literary property of presswork - Google Patents

Abstract

The present invention belongs to the fields of information concealing and digital watermark in multimedia information security, and provides a new method for digital watermark in presswork copyright protection aiming at the widely used half-color image in pressing process. The method is based on nerve network (NN) capable of realizing synchronous accomplishment of watermark embedding and half-color treating through measures such as a filter with function of self-adapted regulation of error diffusion, etc. In particular, the method does not need the original carrier image when extracting the watermark. The invention can be applied in bio-value output devices such as a printer, an electrograph, etc.

Description

The novel method of digital watermarking that is used for protecting literary property of presswork

Technical field

The invention belongs to Information hiding and digital watermark technology field in the multi-media information security, be specifically related to a kind of robustness half tone image watermark new method, have great practical value for the copyright protection of printed matter.

Background technology

Along with the develop rapidly of network technology (particularly Internet technology) with the multimedia messages treatment technology, the intellectual property protection of copyright becomes the key issue that presses for solution.Under this background, digital watermark technology has been subjected to people's common concern, and has become a focus of network information security research field.In recent years, the image digital watermark Study on Technology has obtained remarkable progress, greatly abundant and perfect digital watermarking correlation theory has been proposed successively such as spatial domain, transform domain, compression domain, based on statistics, based on multiple digital watermarking algorithms such as physiological models.Yet existing most image watermark algorithms but can't directly apply to half tone image, this be because: (1) existing algorithm is not considered characteristics such as half tone image (belonging to bianry image) data redundancy is less, can the amount of hiding Info limited; (2) existing algorithm is not considered the self-characteristic (be that factors such as the water absorption of laser beam diffusion, paper and smoothness cause the halftoning spot to change easily, thereby cause the output image smudgy) of printout process; (3) existing algorithm all can't be resisted scanning etc. and causes image that the special attack of distortion takes place easily.

Half-tone picture similarly is a kind ofly to show the particular image of multi-level gray scale with two rank gray scales, but when watching outside certain distance, thinks that still it is a width of cloth continuous-tone image, is often used in fields such as printing, publication.Up to now, be broadly divided into three classes at the digital watermarking algorithm that half tone image proposed: the random error diffusion method, can revise error-diffusion method and condition probability method.Wherein, the random error diffusion method needs original image when extracting watermark, promptly can't realize blind check; It is less to revise the open ended amount of information of error-diffusion method; Condition probability method is not considered human vision model (HVS), and the half tone image of generation is second-rate.

Pertinent literature is as follows:

[1]Ming Sun Fu，O.C.Au.“Data hiding watermarking for halftone images，”In：IEEETransactions on Image Processing，2002，11(4)：477-484.

[2]Xu chao-yong.Digital halftoning image watermarking based on conditional probability[D].Kaohsiung：National Kaohsiung First University of Science and Technology，2002.

Summary of the invention

In view of above-mentioned existing in prior technology problem; the present invention introduces half tone image watermark field with the neural net correlation theory; and a kind of sane half tone image watermark new method that is used for protecting literary property of presswork proposed; this method is based on neural net; can realize finishing synchronously of watermark embedding and halftone process by measures such as self adaptation regulating error diffusion filters.This method does not need initial carrier when detecting digital watermarking.

Basic functional principle of the present invention is: the error diffusion algorithm of Adaline neuron with classics combined, construct a kind of based on neuronic error diffusion filter.Specifically describe as follows:

$a(i,j)=W^T\·X=\left[\begin{array}{cccc}{w}_1& w_2& w_3& w_4\end{array}\right]\·\left[\begin{array}{c}e(i-1,j-1)\\ e(i-1,j)\\ e(i-1,j+1)\\ e(i,j-1)\end{array}\right]---\left(1\right)$

f(i，j)＝g(i，j)+a(i，j) (2)

$y(i,j)=\frac{255}{2}\left\{\mathrm{sqn}\right[f(i,j)-T]+1\},$ $\mathrm{sqn}\left(t\right)=\left\{\begin{array}{c}1,t\&GreaterEqual;0\\ -1,t<0\end{array}\right.---\left(3\right)$

E (i, j)=f (i, j)-y (i, j) (4) wherein, the input vector of filter is sent in X representative, (i j) locates the error that other pixel is produced in the pixel neighborhood and forms in the position by original-gray image; W representation vector, its initial value is made as [1 53 7] according to the distribution principle of error diffusion algorithm ^T/ 16; F (i, j) the clean input of filter is sent in representative, is made up of two parts: a part be a (i, j) (by the inner product generation of input vector and weight vector), another part be image in the position (i, the gray value g that j) locates (i, j); (i j) represents error to e.

Weight vector training criterion adopts LMS (Least Mean Square) algorithm:

$W_{k+1}=W_k+\mathrm{\&eta;}\&CenterDot;\frac{{\mathrm{\&epsiv;}}_k\&CenterDot;X}{{\left|\right|X\left|\right|}^2},{\mathrm{\&epsiv;}}_k=d_k-y_k---\left(5\right)$

Wherein, k represents the self adaptation periodicity; W _kRepresentation vector currency; W _K+1Next is worth the representation vector constantly; η representative training constant, its value decision training speed; ε _kRepresent desirable output d _kWith reality output y _kPoor.

After each process training, all can cause weight vector W to change, so just be difficult to guarantee that each coefficient sum of filter is 1,, need carry out normalized to W with formula (6) in order to keep overall picture quality.

$W^{*}=W/\underset{i=1,...4}{\mathrm{\&Sigma;}}{w}_i---\left(6\right)$

Wherein, the weight vector after the W representative changes, W ^*Represent the weight vector after the normalized.

Original-gray image, embedded location and original watermark information are sent into filter together, just can obtain having the half tone image of watermark.Simultaneously; emulation experiment also shows; the novel method of digital watermarking that is used for protecting literary property of presswork that is proposed not only has the good transparency, and attacks such as JBIG compression, superimposed noise, how much shearings, scribble and printing-scanning are all had robustness preferably.Especially, this method does not need the initial carrier image when extracting digital watermarking.

Description of drawings

The present invention has six accompanying drawings, wherein,

Fig. 1: based on neuronic error diffusion filter.

Fig. 2: the embedding flow process of watermark.

Fig. 3: the testing process of watermark.

Fig. 4 (a): initial carrier image (Lena).

Fig. 4 (b): original watermark image (person of outstanding talent).

Fig. 4 (c): the half tone image after watermarked (PSNR=6.7052dB).

Fig. 4 (d): the digital watermarking that from contain the watermark half tone image, extracts (NC=1.00) (not under fire).

Fig. 5 (a) and (b), (c), (d), (e), (f) and (g): the half tone image after watermarked respectively through JBIG2 compression, noise stack (0.05), noise stack (0.10), shear (1/4), shear (1/2), scribble arbitrarily and print-scanning attack.

Fig. 6 (a) and (b), (c), (d), (e), (f) and (g): from the watermarking images of being attacked that is extracted the watermark half tone image (being Fig. 5) that contains.

Specific embodiments

If initial carrier be 256 grades of gray level image I={g (i, j), 1≤i≤M, 1≤j≤N}, digital watermarking be bianry image S={s (i, j), 1≤i≤P, 1≤j≤Q}.Wherein, g (i, j) and s (i j) represents the i of initial carrier image and binary bitmap capable respectively, j row pixel gray value.Then the half tone image watermark embed process (committed step) based on neural net can be described below.

1, the scramble conversion and the dimensionality reduction of image

In order to eliminate the pixel space correlation of binary bitmap, improve robustness of the present invention, guarantee still can recover watermark whole or in part after an image part is damaged, should at first carry out the scramble conversion to binary bitmap.The present invention adopts Arnold transfer pair binary bitmap to carry out the scramble conversion.

Next, again the watermarking images behind the scramble is utilized line scanning to form one-dimensional vector, and label is 1,2 successively ..., P * Q promptly obtains the one dimension digital watermarking sequence that is converted by original binary bitmap S:

M＝{m(k)，1≤k≤P×Q，m(k)∈{0，255}}

2, the embedding of watermark

The embedding work of digital watermarking is finished in the halftone process process.

At first, utilize pseudo-random seed produce a size for the pseudo random sequence of P * Q as the digital watermarking embedded location.

Then, utilize formula (1), (2), (3) and (4) that image I is carried out Filtering Processing.For general position, the filtering result is final input; For selected watermark embedded location, its real output value y (i, j) be that 255 and 0 probability is 0.5, watermark information position m (k) of desiring to embed in this position simultaneously is that 255 and 0 probability also all is 0.5, be that (i is 0.5 with the corresponding to probability of watermark information position m (k) j) to real output value y.Therefore, can be with watermark information position m (k) as desirable output, i.e. d (k)=m (k).(i j) with m (k) when inconsistent, trains it with formula (5), makes it equal as y.Method following (false code):

An if d (k)!=y (i, j)/the desirable output of * and the inconsistent * of actual output/

then{

ε=d (k)-y (i, j)/* utilize formula (5) calculate the difference * of desirable output and actual output/

repeat{

W _K+1=W _k+ (η ε X)/‖ X ‖ ²/ * utilize formula (5) to weight vector train */

a ^*(i，j)＝X ^T·W _k+1

if g(i，j)+a ^*(i，j)≥T

then y ^*(i，j)＝255

else y ^*(i，j)＝0

ε ^*＝d(k)-y ^*(i，j)}

until ε ^*！＝0}

else no change

Each through after the training, need W to be carried out normalized with formula (6).

By aforementioned calculation, just can obtain containing the half tone image I of digital watermarking ^*

3, the detection of watermark

The digital watermark detection method of discussion of the present invention belongs to object detection method, does not promptly need original carrier image when detecting digital watermarking.If half tone image to be detected is I ^*, then digital watermarking testing process is as follows:

At first, produce the random sequence that size is P * Q with identical pseudo-random seed, and with the extracting position as digital watermarking.And directly take out the pixel value of these positions, just can obtain the one dimension binary sequence:

M ^*＝{m ^*(k)，1≤k≤P×Q，m ^*(k)∈{0，255}}。

Then, the one dimension binary sequence M to being extracted ^*Carry out rising the random conversion (according to the inverse operation of watermark embed process (step 2)) of dimension and inverted, just can obtain binary bitmap:

S ^*＝{s ^*(i，j)，1≤i≤P，1≤j≤Q}

In addition, for the masters such as experience, physical qualification, experiment condition and equipment that eliminate the observer, the influence of objective factor, (Normalized Cross-Correlation NC) carries out quantitative assessment to the watermark of extraction and the similitude of original watermark to need to adopt normalizated correlation coefficient.

4, emulation experiment

In order to verify high efficiency of the present invention, below provided the experimental result that detects performance test, anti-attack ability test respectively, and contrast with document [2] method, comprising: JBIG2 compression, noise stack, shear for how much, scribble and printing-scanning etc.In the experiment, selected initial carrier is 512 * 512 * 8bit standard grayscale image Lena, Mandrill and Barbara.32 * 32 two-value pattern " person of outstanding talent " has been adopted in digital watermarking.The training selection of constant is η=0.6.

(1) JBIG2 compression.

(2) noise stack.To the half tone image that is embedded with watermark add that average is 0, mean square deviation is 0.05 and 0.10 salt-pepper noise.

(3) shear for how much.Respectively the half tone image that is embedded with watermark being carried out 1/4 and 1/2 shears.

(4) scribble.The half tone image that is embedded with watermark is altered arbitrarily.

(5) printing-scanning (D/A-A/D conversion).The half tone image that is embedded with watermark is used the printout on common paper of HP4VC laser printer earlier.Re-use scanner scanning output (resolution is 400dpi).Use Photoshop software to carry out angle correct to the digital picture of obtaining again then, and make the sampling resolution of image become 72dpi again, it is 256 * 256 that size is also readjusted.

Table 1 and table 2 have provided the contrast and experiment of two kinds of half tone image water mark methods.

Table 1 contains the Y-PSNR (dB) between watermark half tone image and initial carrier

	The present invention	Document [2] method
			Lena Mandrill Barbara	6.7052 6.6679 7.0474	6.5270 6.2480 6.8812

Table 2 digital watermarking is handled and the resistivity (NC) of attacking common image

Attack pattern	Processing parameter	Lena		Mandrill		Barbara
								The present invention	Document [2] method	The present invention	Document [2] method	The present invention	Document [2] method
		Not attacking how much shearing scribbles of JBIG2 compression stack salt-pepper noise attacks	0.05 1/1st/0.10th four at random	1 0.9821 0.9403 0.8819 0.7220 0.5923 0.9584	1 0.9800 0.9327 0.8667 0.7169 0.5881 0.8614	1 0.9708 0.9370 0.8828 0.7220 0.5923 0.9248	1 0.9619 0.9131 0.8601 0.7083 0.5607 0.8297							1 0.9810 0.9403 0.8819 0.7220 0.5923 0.9223	1 0.9771 0.9326 0.8664 0.7169 0.5881 0.8527

Claims (2)

1. The present invention proposes a kind of new method that is used for protecting literary property of presswork, its claim mainly comprises following two aspects:

   1, based on neuronic adaptive error diffusion filter:

   The Adaline neuron is combined with error diffusion algorithm, construct a kind ofly based on neuronic error diffusion filter, this filter can pass through self adaptation regulating error diffusion filter based on neural net.Specifically describe as follows:

   $a(i,j)=W^T\&CenterDot;X=\left[\begin{array}{cccc}\frac{1}{16}& \frac{5}{16}& \frac{3}{16}& \frac{7}{16}\end{array}\right]\&CenterDot;\left[\begin{array}{c}e(i-1,j-1)\\ e(i-1,j)\\ e(i-1,j+1)\\ e(i,j+1)\end{array}\right]---\left(1\right)$

   f(i，j)＝g(i，j)+a(i，j) (2)

   $y(i,j)=\frac{255}{2}\left\{\mathrm{sqn}\right[f(i,j)-T]+1\},\mathrm{sqn}\left(t\right)=\left\{\begin{array}{c}1,t\&GreaterEqual;0\\ -1,t<0\end{array}\right.---\left(3\right)$

   e(i，j)＝f(i，j)-y(i，j) (4)

   Wherein, the input vector of filter is sent in the X representative, and (i j) locates the error that other pixel is produced in the pixel neighborhood and forms in the position by original-gray image; W representation vector, its initial value is made as [1537] according to the distribution principle of error diffusion algorithm ^T/ 16; F (i, j) the clean input of filter is sent in representative, is made up of two parts: a part be a (i, j) (by the inner product generation of input vector and weight vector), another part be image in the position (i, the gray value g that j) locates (i, j); (i j) represents error to e.

   Weight vector training criterion adopts LMS (Least Mean Square) algorithm:

   $W_{k+1}=W_k+\mathrm{\&eta;}\&CenterDot;\frac{{\mathrm{\&epsiv;}}_k\&CenterDot;X}{{\left|\right|X\left|\right|}^2},{\mathrm{\&epsiv;}}_k=d_k-y_k---\left(5\right)$

   Wherein, k represents the self adaptation periodicity; W _kRepresentation vector currency; W _K+1Next is worth the representation vector constantly; η representative training constant, its value decision training speed; ε _kRepresent desirable output d _kWith reality output y _kPoor.

   After each process training, all can cause weight vector W to change, so just be difficult to guarantee that each coefficient sum of filter is 1,, need carry out normalized to W with formula (6) in order to keep overall picture quality.

   $W^{*}=\frac{W}{\underset{i=1,...,4}{\mathrm{\&Sigma;}}{W}_i}---\left(6\right)$

   Wherein, the weight vector after the W representative changes, W ^*Represent the weight vector after the normalized.
2. 2, utilize sef-adapting filter proposed by the invention to realize the finishing synchronously of embedding of halftone process and watermark:

   At first, utilize pseudo-random seed produce a size for the pseudo random sequence of P * Q as the digital watermarking embedded location.

   Then, utilize formula (1), (2), (3) and (4) that image I is carried out Filtering Processing.For general position, the filtering result is final input; For selected digital watermarking embedded location, its real output value y (i, j) be that 255 and 0 probability is 0.5, watermark information position m (k) of desiring to embed in this position simultaneously is that 255 and 0 probability also all is 0.5, be that (i is 0.5 with the corresponding to probability of watermark information position m (k) j) to real output value y.Therefore, can be with watermark information position m (k) as desirable output, i.e. d (k)=m (k).(i j) with m (k) when inconsistent, trains it with formula (5), makes it equal as y.In order to keep overall picture quality, after each training, utilize formula (6) that new weight vector is carried out normalized.

Images