CN1321393C - A method for calibrating and protecting digital images using image geometry - Google Patents

Abstract

The invention relates to the field of multimedia signal processing, and is a method for image geometry calibration and digital image protection. After spread spectrum modulation and interleaving, the information watermark embedded in the image DWT domain realizes resynchronization detection by the matching template embedded in the image DFT domain and the training sequence embedded in the image DWT domain. Under the condition that the watermarked image undergoes JPEG compression and geometric deformation at the same time, error-free detection of meaningful information can still be realized. The invention can protect the digital image or video data transmitted on the network. The image geometric calibration method proposed in the present invention can also be used in other occasions that require image synchronization, such as satellite imaging, interactive digital maps, digital watermarks, and the like.

Description

Adopt the method for image geometry calibration and protection digital image

Technical field

The present invention relates to a kind of field of multimedia signal processing, is the method for a kind of image geometry calibration and protection digital image.

Background technology

At present, watermark Against Geometrical Attacks has the method for non-blind Detecting and blind Detecting.In general, the robustness of blind Detecting water mark method is relatively poor.But owing to do not need raw image, its range of application is wider, also has more challenge.It can be divided into two classes again: first kind method is that watermark is embedded in the image feature territory with geometric invariance, as the Fourier-Mellin territory, makes geometry deformation not influence the extraction of watermark information.But these class methods all can only be resisted RST, also have difficulties in realization, for example when the amplitude spectrum of visual DFT conversion being LPM (Log Polar Mapping), because interpolation error can cause the serious decline of image quality with ILPM (Inverse LogPolar Mapping).Second class is to cause under the situation of watermark detection step-out at geometry deformation, and it is synchronous with the weight of realizing watermark detection to manage to carry out geometry correction earlier before watermark detection.This need also embed a geometric correction model watermark (template watermark) except that hiding the watermark (information watermark) of carrying user profile in image.Up to the present, also there are many problems in Against Geometrical Attacks, comprises that the hiding data amount is few, relatively poor, the anti-JPEG of the invisibility of watermark compresses and anti-affined transformation ability is weak, can not resist the combination attacks of JPEG compression and geometric transformation etc. simultaneously.In addition, many methods adopt and embed the information watermark in the DFT territories, and DFT (Discrete FourierTransform) compares with DWT (Discrete Wavelet Transform), has the weakness of self, is difficult to become the basis of main flow water mark method.

Summary of the invention

The objective of the invention is to propose a kind of sane stealthy visual blind Detecting water mark method that can resist general signal Processing such as JPEG compression and geometric transformation simultaneously, and the quantity of information of hiding (quiet lotus) is big (can reach more than 200 bits), and the invisibility of watermark is better.

The method of calibration of the inventive method image geometry and protection digital image, at first the information watermark is embedded in the visual DWT territory with a training sequence through band spectrum modulation with after interweaving, again a matching template is embedded into visual DFT territory, be embedded in the heavy synchronous detection of the information watermark in the visual DWT territory after realizing band spectrum modulation and interweave by matching template that is embedded in visual DFT territory and the training sequence that is embedded in visual DWT territory at last, specific practice is: 1) meaningful information b to be embedded (L bits) is at first carried out the spread spectrum coding modulation with a pseudo-random binary PN sequence that is produced by key, can obtain scale-of-two watermark data W to be embedded like this, and then interweave; To adopt the method that quantizes modulation directly to be embedded in the LL that visual DWT conversion obtains by the training sequence T that the pseudo-random code that key produces is formed ₃In the central row of subband and central series sub-band coefficients or other sub-band coefficients, and the method that adopts quantification to modulate in all the other sub-band coefficients embeds the scale-of-two watermark data W after interweaving; By 2-D IDWT obtain embedding the watermark of DWT territory visual f ' (x, y); 2) (x y) carries out the DFT conversion, and the method that we adopt addition to embed in the amplitude spectrum coefficient of DFT conversion embeds a matching template, and template point position is produced by a cipher controlled with f '; 3) testing process is the inverse process of telescopiny, detect matching template earlier, and after relatively obtaining the transformation matrix of the affined transformation that image stood and do inverse transformation recovering geometric configuration with original matching template, do visual translation calibration back according to training sequence again and extract watermark information.

For detect the image stand the transformation matrix of affined transformation and recover its original geometry do inverse transformation, can embed a matching template that is made of a little louder local pole in the amplitude spectrum coefficient of visual DFT conversion, a little bigger position of local pole can be produced by a cipher controlled; Embed a training sequence simultaneously in visual DWT territory and be used for the translation calibration.

The information watermark that is embedded in the visual DWT territory through band spectrum modulation with after interweaving realizes heavy synchronous detection by matching template that is embedded in visual DFT territory and the training sequence that is embedded in visual DWT territory; The detection two big steps that are divided into the embedding and the watermark of watermark.

1, the embedding of watermark:

The present invention is embedded in the information watermark in the low frequency sub-band coefficient in visual DWT territory, and the template watermark is embedded in the amplitude spectrum intermediate frequency coefficient after the visual DFT conversion, and two parts watermark is not disturbed mutually like this, can obtain better robustness again.The image watermark embedding grammar block scheme that the present invention proposes as shown in Figure 1.

2: the detection of watermark:

Do not need the auxiliary of raw image, the inventive method can detect the data of hiding and obtain from the watermark image that may suffer geometric attack and JPEG compression simultaneously.Testing process is as follows:

A) when watermark detection, at first want the watermark of application training Sequence Detection whether synchronous.If asynchronous, then must be earlier through the overweight synchronous image g that obtains synchronously ^*(x, y).It is heavy that to comprise that synchronously plate watermark detection, contrary affined transformation, the translation of application training sequence are touched in the DFT territory synchronous.If synchronously, directly do a step.

B) to synchronous image g ^*(x y) makes DWT territory watermark detection, has obtained actual hiding data.

Watermark embed process mainly contains the embedding of the DWT territory watermark pre-service of (comprising information watermark, training sequence), the watermark of DWT territory and embedding three parts that the plate watermark is touched in the DFT territory.

The pre-service of 1) DWT territory watermark (information watermark, training sequence): direct sequence spread spectrum is encoded, is interweaved.

The present invention introduces the robustness and the secret that hide Info with enhancing in the image watermark method with technology commonly used in some communication theories (as the direct sequence spread spectrum modulation and interweave).

Suppose that the raw image size is 512 * 512.Application length is N ₁PN sign indicating number sequence m={m _jJ=1 ..., N ₁Information b{b to embedding _iI=1 ..., L} (b wherein _j∈ 0,1}) carry out the spread spectrum coding modulation." 1 " is modulated to m (bipolar sequence, m _j∈ 1, positive sequence 1}), i.e. {+1 * m _iJ=1 ..., N ₁, " 0 " is modulated to the anti-phase sequence of m, i.e. { 1 * m _jJ=1 ..., N ₁.15 PN sign indicating number sequence is produced by PN sign indicating number sequencer by a key.Can obtain scale-of-two watermark data W to be embedded like this:

$b_i\stackrel{\mathrm{DSSS}coding}{\&RightArrow;}{W}_i\{w_{\mathrm{ij}};w_{\mathrm{ij}}\&Element;\{-1,+1\}1\&le;j<N_1,1\&le;i<L\}$

Training sequence is that can the information watermark realize the key that translation is synchronous, and for make its influence that is subjected to visual cutting less as far as possible, it should be embedded in needs to lay special stress on protecting the low frequency sub-band position of part correspondence or the central row and the central series of low frequency sub-band in the image.Be embedded in 32 row and 32 row of low frequency sub-band as shown in Figure 2.And in the remainder of low frequency sub-band, embed through the scale-of-two watermark data W after interweave (adopting 2-dimensional interleaving technology or other interleaving technologies).

Deposit 127 training sequences on 32 row of one 64 * 64 two-dimensional matrix and 32 column positions, all the other sequence of positions are deposited the scale-of-two watermark data W afterwards that interweaves, and the two-dimensional matrix that obtains is become an one-dimension array by line scanning, are designated as X.

The embedding and the detection method of 2) DWT territory watermark (information watermark, training sequence)

(x y) carries out three grades of DWT and decomposes, low frequency sub-band LL the embedding raw image f of DWT territory watermark ₃Coefficient becomes one-dimension array by line scanning, is designated as C.By formula (1), we are added to binary data X on the low frequency coefficient C, obtain new low frequency coefficient c ':

0≤i＜4096 wherein, C (i), C ' (i), x _iBe respectively i the element of C, C ', X.α represents embedment strength, is satisfying under the prerequisite of invisibility, selects maximum round values as far as possible.With the wavelet coefficient behind the embed watermark carry out IDWT obtain embedding the watermark of DWT territory visual f ' (x, y).

The detection of DWT territory watermark is synchronous visual g ^*(x, the low frequency sub-band LL after y) DWT decomposes ₃Coefficient becomes one-dimension array by line scanning, is designated as C ^*The binary data that extracts is designated as X ^*={ x _i ^-, it is as follows to extract formula:

$\left\{\begin{array}{c}{x}_i^{*}=+1,\left(C^{*}\left(i\right)\mathrm{mod}a\right)\&GreaterEqual;\frac{a}{2}\\ x_i^{*}=-1,\mathrm{otherwise}\end{array}---\left(2\right)\right.$

0≤i＜4096 wherein, α is an embedment strength.With the binary data X that extracts ^*Carry out reciprocal cross and knit the binary data sequence W that (inverse process that interweaves) recovers embedding ^*W then ^*Carry out segmentation by 15 bits, every section is carried out relevantly with the sequence m of 15bits, if correlation is greater than 0, then to embed information bit be " 1 " in judgement, otherwise judgement embedding information bit is " 0 ".Just obtain the embedding information recovered after the despreading.

3) embedding of DFT territory template and detection

(x, DFT territory y) embeds the synchronizing information after a template is out of shape as watermark image to visual f ' after embedding DWT territory watermark.

The embedding of template divides following four steps:

A) (x, y) (512 * 512) extend to 1024 * 1024 around filling with average with f '.

B) do the DFT conversion, get the fourier coefficient range weight.(normalized frequency is 0.20～0.30) embeds 28 template points in the intermediate frequency zone, and being evenly distributed in inclination angle, DFT territory is θ ₁And θ ₂Two straight lines, 14 points on the every line, Fig. 2 is the synoptic diagram of embedded template, the situation of 14 the template points of poincare half plane that only draw among the figure, lower half-plane also embeds 14 template points about former point symmetry.The inclination angle of straight line and the utmost point of template point footpath are produced by a key pseudorandom.

C) the mould value of increase template point place fourier coefficient, (can adopt radius is the circular window of R, maximum value as shown in Figure 3) to make it to become regional area.The change amount is standard with invisible, and generally getting maximum value is that local mean values adds the variance about several times to tens times.

D) calculate inverse fourier transform (IDFT) and obtain final watermark image f " (x, y).

Image will produce the corresponding linear conversion in the DFT territory in the linear transformation that spatial domain is subjected to, so just can determine the geometric deformation that image is experienced by the transformation relation of template point position.If the linear transformation of square image below spatial domain has taken place:

$\left[\begin{array}{c}x\\ y\end{array}]\&RightArrow;B\left[\begin{array}{c}x\\ y\end{array}]---\left(3\right)\right.\right.$

Be equivalent to so do following linear transformation in the DFT territory:

$\left[\begin{array}{c}u\\ v\end{array}]\&RightArrow;{\left(B^{-1}\right)}^T\left[\begin{array}{c}u\\ v\end{array}\right]\right.---\left(4\right)$

For the template point on the template line, after the experience linear transformation, they still coexist one and cross on the initial point straight line.There is certain relation (as r '=Kr, K is a certain constant) in the coordinate of new template point (as utmost point footpath r ') with the coordinate (as utmost point footpath r) of original template point, and this can be used for the quick matching judgment of search procedure.

The step of template detection is as follows:

A) treating mapping resembles g (x y) does Barlette filtering.

The same during b) with embedded template, filtered image to be measured is extended to 1024 * 1024.

C) do the DFT conversion.With a radius be R ' (R '＜R, the windows radius of R when embedding) circular window (as regional area) in the poincare half plane of fourier coefficient magnitude matrix, search for, extract all local maximum points.Is DFT coefficient amplitude matrix poincare half plane that the summit is divided into N with the initial point _b(N _b=360 or 180 or other values) individual sector region, each fan-shaped drift angle is 0.5 ° or 1 °.By angle all local maximum points are included into each sector region respectively again.

D) find with two and touch corresponding possible of printed line and touch the set of plate point.

In each sector region, at K _Min＜K＜K _MaxThe such K value of search in the scope: it makes to have N at least in this sector _mIndividual local maximum point satisfies | r _Li-Kr _Tj' |＜threshold, wherein N _mBe a number of predesignating, r _LiBe the utmost point footpath (i=l...N of Local Extremum among the i of sector _b), r _Tj' be the utmost point footpath of touching plate point on the grand master pattern printed line j (j=1,2), threshold＞0 is a threshold value.We get N in the experiment _m=5, threshold=0.002, K _Min=0.5 and K _Max=2.0 (are 2～0.5 corresponding to the zooming parameter on the spatial domain).If find such K value, we just get off corresponding Local Extremum coordinate record.

E) by above-mentioned steps, obtain the set of possible matched line, be called " accurate matched line ", the Local Extremum on the line is called " accurate match point ", and coordinate is designated as (x _Ij, y _Ij).The coordinate of the corresponding primary template point of image poincare half plane is designated as (x _Ij', y _Ij'), wherein { 1,2} represents i bar template (coupling) line to i ∈, and { 1,2, Λ, 7} represent j template (coupling) point to j ∈.Take out a set and concentrate another set of taking-up from concentrating corresponding to the accurate match point of template line 2 corresponding to the accurate match point of template line 1.A possible transformation matrix A who calculates according to the point and the corresponding relation between template point of these two set.Seek the minimum A of average error MAE (Mean Absolute Error).

$\mathrm{MAE}=\frac{1}{\mathrm{nummatches}}\left|\right|{A\left(\begin{array}{cc}{x}_{11}& y_{11}\\ M& M\\ x_{11}& y_{11}\\ x_{21}& y_{21}\\ M& M\\ x_{21}& y_{21}\end{array}\right)}^T-{\left(\begin{array}{cc}{x}_{11}^{\&prime;}& y_{11}^{\&prime;}\\ M& M\\ x_{11}^{\&prime;}& y_{11}^{\&prime;}\\ x_{21}^{\&prime;}& y_{21}^{\&prime;}\\ M& M\\ x_{21}^{\&prime;}& y_{21}^{\&prime;}\end{array}\right)}^T\left|\right|---\left(5\right)$

Wherein template point is (x _Ij', y _Ij') and " accurate match point " be (x _Ij, y _Ij), nummatches is the match point number, is the error matrix of one 2 row among the operational symbol ‖ Λ ‖.

F) will add 180 ° corresponding to the accurate match point of template line 1, repeat e), determine last frequency domain transform matrix A by the MAE value of minimum.Can get spatial domain transformation matrix B=A by formula (3) and (4) ^T

Use the training sequence S of extraction and the related coefficient of original training sequence T and determine whether image reaches the translation synchronization parameter of image synchronously.

Whether when watermark detection, it is synchronous at first will to detect watermark.If asynchronous, then must just can carry out watermark detection behind the heavy synchronization watermarking.If synchronously, then directly extract low frequency sub-band LL ₃Subband hiding data and decode information.

Whether detect watermark synchronous: (x, y) resetting size is 512 * 512, then it is carried out 3 grades of DWT and decomposes, from LL with visual g to be measured ₃Extract training sequence S in 32 row of subband and 32 row, calculate the related coefficient of it and original training sequence T $P_{T,S}\left(0\right)=\frac{1}{127}\underset{n=1}{\overset{127}{\mathrm{\&Sigma;}}}\left(T_n{S}_n\right),$ Whether see 〉=threshl.If we think that S is real training sequence, and watermark is synchronous, can directly carry out the extraction and the decoding of the watermark of DWT territory.If＜threshl thinks that then watermark is nonsynchronous, must be earlier through overweight extraction and the decoding that just can carry out the watermark of DWT territory synchronously.General desirable 0.56 (by the definite value of experiment) of threshl.Pseudo-synchronous probability promptly appears in the appearance false-alarm can be by calculating $P_{\mathrm{fp}}=\frac{1}{2^{127}}\underset{k=127-e}{\overset{127}{\mathrm{\&Sigma;}}}{C}_{127}^k=8.59\&times;{10}^{-9}$ E=round (127 * (1-threshl)/2) wherein.Round represents round.

The heavy synchronous first step is: recover original geometry.(x detects the template watermark of embedding in y), and it and original template compared obtains the affine transformation matrix B that image is stood from visual g to be measured.After obtaining affine transformation matrix B, with visual g (x to be measured, y) carry out the visual g ' (x that the image geometry inverse transformation reverts to M * N size, y) (Fig. 5 b), and then fill 0 one-tenth 512 * 512 size visual I (x, y), filled with 0 by the part of cutting, (x is y) at visual I (x, y) center (Fig. 5 c) for g '.

The heavy second synchronous step is: translation is synchronous.Promptly determine the translation synchronization parameter of image with the related coefficient of training sequence S that extracts and T.

Translation synchronously adoptable a kind of way is, with visual I (x, y) do following all possible translation:

I _t(x，y)＝I(x-x _t)mod512，(y-y _t)mod512)；

$\{-\frac{1}{2}(512-M)\&le;x_1<\frac{1}{2}(512-M);-\frac{1}{2}(512-N)\&le;y_1<\frac{1}{2}(512-N)\}---\left(6\right)$

Image after each translation is made DWT and is decomposed, from LL ₃32 row of subband extract training sequence S in being listed as with 32.Can determine translation parameters (x according to the training sequence that extracts and the related coefficient maximum of original training sequence _t, y _t).

Another method that the present invention proposes is that (x y) does maximum 8 * 8=64 time translation and gets final product, thereby can reduce calculated amount greatly with visual I.According to the time-frequency local character of DWT, LL ₃Each coefficient of subband is all corresponding to a part of image.Can prove (our experiment has also proved this point), if when DWT, adopt a tight wavelet filter and adopt periodic extension mode (if other continuation modes of employing then except image border, also satisfy following relationship), visual I (x, y) translation 8 * x _T1Row and 8 * y _T1Row (x _T1, y _T1Be integer), obtain translation image I _t(x, y):

I _t(x，y)＝I((x-8×x _t1)mod512，(y-8×y _t1))mod512) (7)

LL after then three grades of DWT of image decompose ₃Subband is translation x also _T1Row and y _T1Row:

LL _3t(x，y)＝LL ₃((x-x _t1)mod 64，(y-y _t1)mod64) (8)

LL wherein ₃(x, y) and LL _3t(x, y) be respectively visual I (x, y) and I _t(x, y) LL ₃Sub-band coefficients.LL ₃Translation also takes place in the training sequence that the translation of subband causes embedding.The character that application of formula 7 and 8 provides, we can be only to I (x, y) do maximum 8 * 8 translations:

I _t(x，y)＝I((x-x _t)mod 512，(y-y _t)mod 512)；{-4≤x _t，y _t＜4 (9)

Every translation once is DWT and is decomposed, and obtains LL ₃Subband LL _3t(x, y).With LL _3t(x, y) do translation:

LL _3t′(x，y)＝LL _3t((x-x _t1)mod64，(y-y _t1)mod64)；{-T ₁≤x _t1＜T ₁；-T ₂≤y _t1＜T ₂ (10)

In the following formula, T ₁=round (0.5 * (512-M)/8), T ₂=round (0.5 * (512-N)/8).Each translation is from LL _3t'; (x, 32 row y) and 32 row extractions training sequence S, according to and original training sequence T between maximum related value determine translation parameters.Can determine the translation parameters (8 * x of image after maximum 64 translations search _t+ x _T1, 8 * y _t+ y _T1), thereby the visual g after the acquisition translation calibration ^*(x, y).

The present invention has the following advantages:

1) the digital watermarking blind checking method of the DFT-DWT compositum of the present invention's proposition resists normal signal processing aspect and affined transformation aspect at the same time and has all reached stronger robustness (table 1).When compressibility factor is 15 JPEG compression (JPEG_15), can realize that zero defect detects, can resist other geometric transformations except that Rand Bending among the international watermark test platform StirMark 3.1, as to rotation (auto crop, auto scale), jitter, scaling, shearing, general linear transform etc. can both realize that zero defect detects, and can resist the combination attacks of JPEG compression and geometric transformation, as resisting combination attacks such as JPEG_50 compression, rotation, convergent-divergent, cutting, translation simultaneously.

2) water mark method that proposes of the present invention can be hidden the above information of 264 bits, and more than 40dB, the invisibility of watermark is better with respect to the PSNR of raw image for watermark image.

3) the image geometry collimation technique among the present invention, the degree of accuracy height, and can avoid in image, embedding a visable indicia.

Table 1 the present invention proposes the result that method water seal test platform StirMark 3.1 carries out the robustness test.

StirMark functions	Lena	Baboon
			JPEG 15～100	0	0
Gauss filtering	0	0
			sharpening	0	0.02
jitter	0	0
			scaling	0	0
aspect ratio	0	0
			cropping 25	0	0
rotation(auto-crop，scale)	0	0
			general linear transform	0	0
shearing	0	0

Description of drawings:

Fig. 1 is that watermark embeds block diagram.

Fig. 2 is training sequence and is embedded in LL ₃The position of subband.

Fig. 3 is the position view that embeds 14 template points at DFT conversion amplitude spectrum poincare half plane.

Fig. 4 is the image that has embedded the watermark of DWT-DFT territory.

Fig. 5 is the robustness test that the inventive method antagonism JPEG compression and RST (rotation, cutting, convergent-divergent and translation combination attacks) gang up against.

Among Fig. 1, the 1st, need hiding bit information, the 2nd, hide Info through direct sequence spread spectrum, the 3rd, interweave, the 4th, raw image, the 5th, DWT, the 6th, data are embedded in the visual DWT sub-band coefficients, the 7th, IDWT, the 8th, embedded the watermark of DWT territory visual f ' (x, y), the 9th, with f ' (x, y) carry out the DFT conversion, the 10th, in the DFT territory, embed and touch the plate watermark.The 11st, IDFT.The 12nd, embedded the visual f of DWT-DFT territory compound watermark " (x, y).The 13rd, the training sequence that embeds.

Among Fig. 2, T ₁... T ₁₂₇Be that 127 training sequences are at LL ₃Certain delegation in the subband and the position of a certain row.Row 32 expressions 32 row, Column 32 expressions 32 row.

Some results that are to use reference image Lena and Baboon test that provide among Fig. 4,5.

Among Fig. 4, a) Lena watermark image (PSNR=40.1dB); B) Baboon watermark image (PSNR=39.6dB).

Among Fig. 5, and the visual g after a) the Lena watermark image is subjected to JPEG compression and RST and gangs up against (x, y).

B) figure a through the visual g ' after the calibration affined transformation (x, y).Image size is 504 * 504.

C) figure mend around the b 0 to 512 * 512 big or small I (x, y).

D) the visual g after figure c proofreaies and correct through translation ^*(x, y) image size is 512 * 512, filling part is visual g (x, average y).264 bits of hiding still can detect by zero defect.

Embodiment

Embodiment:

The development of computing machine, printer and high-speed transmission equipment makes and becomes very convenient at enterprising row image of network and transmission of video signals.But the serious problems that transmission and storage faced of electron image, video etc. are that their duplicate and original paper is just the same, thereby copyright owner's be unwilling to propagate by this way their material.Because the Internet is increasingly extensive in the application of commercial field, therefore need a kind of means that can protect electronic data in a hurry.Digital watermarking is to carry a group of owner's copyright information to distinguish means.Digital watermarking for good and all is embedded into to be used for copyright protection and to check whether data are destroyed in the multi-medium data.But present digital watermark technology just detects not come out mostly behind some very little geometry deformations of image process.Therefore great advantage of the present invention is to make the digital watermarking of embedding can resist the attack etc. of geometry deformation and image compression coding simultaneously.

The digital watermark technology that provides anti-geometric transformation below uses and tests some results that obtain at reference image Lena and Baboon.

(be on 512 * 512 * 8bits) the image and embed information watermark, matching template and a training sequence that comprises 44 characters (264 bit) respectively, the PSNR value is respectively 40.1dB and 39.6dB (Fig. 4) at Lena and Baboon for we.We get L=264 in the test, N ₁=15, α=56, threshl=0.56, threshold=0.002, N _b=180, N _m=5, K _Min=0.5 and K _Max=2.0, and in to visual wavelet transform, use Daubechies 9/7 biorthogonal wavelet wave filter, training sequence is embedded in LL ₃32 row of subband and 32 row.Watermark embedding method of the present invention need be less than 3 second time on the P4 of 1.7G computing machine (windows platform, VC++ language), and detection method needs about 1～13 second.As can be seen, calculated amount is very not big.

Table 1 is the situation that information watermark opposing StirMark3.1 attacks.Wherein BER (Bit Error Rate) is meant bit error rate.Watermark can both realize that to geometric attacks such as convergent-divergent, rotation+cutting, rotation+cutting+convergent-divergent, general linear transformation, shearing, jitter, change length breadth ratio among the StirMark3.1 zero defect detects.Can realize when cropping_25, JPEG_15 attack in StirMark3.1 that zero defect detects.To Gaussian filtering, sharpening, the watermark character string also can the zero defect rate or is extracted with extremely low error rate.

In addition, we also tested be subjected at watermark image rotating arbitrarily+to shear+situation of convergent-divergent+translation (RST) under the robustness of watermark.As shown in Figure 5.Fig. 5 (a) be after the Lena watermark image is subjected to JPEG compression and RST and gangs up against visual g (x, y).Fig. 5 (b) be Fig. 5 (a) through recover after the calibration affined transformation behind its original-shape visual g ' (x, y).Image size is 504 * 504.Fig. 5 (c) is that (x, (x y), mends 0 back former visual g ' (x, the visual I that y) is positioned at (x, central authorities y) y) to fill up the visual I of 0 to 512 * 512 size on every side at visual g '.Fig. 5 (d) is that (x is y) through the visual g after the translation correction for visual I ^*(x, y), image size is 512 * 512.During watermark detection, (x, fill, rather than with 0 value filling, the result who does like this can improve the detection performance by average y) with visual g for filling part.Test result is that 264 bits of hiding still can detect by zero defect.

Claims (2)

1. the method for image geometry calibration and protection digital image; it is characterized in that this method at first is embedded into the information watermark in the visual DWT territory with a training sequence through band spectrum modulation with after interweaving; again a matching template is embedded into visual DFT territory; be embedded in the heavy synchronous detection of the information watermark in the visual DWT territory after being realized band spectrum modulation and interweaved by matching template that is embedded in visual DFT territory and the training sequence that is embedded in visual DWT territory at last, specific practice is:

1) watermark embeds: watermark embed process mainly contains the pre-service of the DWT territory watermark that comprises information watermark, training sequence, the embedding of DWT territory watermark and embedding three parts that the plate watermark is touched in the DFT territory;

I) pre-service of DWT territory watermark: direct sequence spread spectrum is encoded, is interweaved;

Application length is N ₁PN sign indicating number sequence m={m _jJ=1 ..., N ₁Information b{b to embedding _iI=1 ..., L} is b wherein _i{ 0,1} carries out the spread spectrum coding modulation to ∈; " 1 " is modulated to the positive sequence of m, promptly+and 1 * m _jJ=1 ..., N ₁, " 0 " is modulated to the anti-phase sequence of m, i.e. { 1 * m _jJ=1 ..., N ₁, can obtain scale-of-two watermark data W to be embedded like this;

One with the central row of the two-dimensional matrix of the identical size of visual DWT low frequency sub-band and central series on or need to lay special stress on protecting the corresponding low frequency sub-band position of part in the image and deposit the pseudorandom training sequence T that produces by key, scale-of-two watermark data W after all the other sequence of positions are deposited and interweaved, the two-dimensional matrix that obtains is become an one-dimension array by line scanning, be designated as X;

Ii) (x y) carries out three grades of DWT and decomposes, low frequency sub-band LL the embedding raw image f of DWT territory watermark ₃Coefficient becomes one-dimension array by line scanning, is designated as C; By following formula, we are added to binary data X on the low frequency coefficient C, obtain new low frequency coefficient C ':

Wherein C (i), C ' (i), x _iBe respectively i the element of C, C ', X; α represents watermark embed strength; With the wavelet coefficient behind the embed watermark carry out IDWT obtain embedding the watermark of DWT territory visual f ' (x, y);

The iii) embedding of DFT territory template: (two straight lines of edge mistake initial point increase the mould value of some intermediate-frequeney point place fourier coefficients to visual f ' after embedding DWT territory watermark, make these points become the maximum value of regional area for x, DFT territory y); The change amount is standard with invisible, and generally getting maximum value is that local mean values adds the variance about several times to tens times; The local maximum point that distributes along two straight lines of these embeddings constitutes a template, as the synchronizing information after the watermark image distortion; The position of these template points can be produced by a cipher controlled; These two straight lines of crossing initial point are called the template line;

2) detection of watermark:

Testing process is as follows:

A) when watermark detection, at first want the watermark of application training Sequence Detection whether synchronous; If asynchronous, then must earlier the test image be obtained synchronous image g synchronously through weighing ^*(x, y), heavy to comprise that synchronously plate watermark detection, contrary affined transformation, the translation of application training sequence are touched in the DFT territory synchronous; If synchronously, directly do next step;

B) to synchronous image g ^*(x y) makes DWT territory watermark detection, has obtained actual hiding data;

The heavy synchronous first step is: recover original geometry; (x detects the template watermark of embedding in y), and it and original embedded template compared obtains the affine transformation matrix B that image is stood from visual g to be measured;

Suppose that the raw image size is M ₁* M ₁, the step of template watermark detection is as follows:

A) treating mapping resembles g (x y) does Barlette filtering;

The same during b) with embedded template, filtered image to be measured is extended to 1024 * 1024;

C) do the DFT conversion; With the circular window that a radius is R ', in the poincare half plane of fourier coefficient magnitude matrix, search for, extract all local maximum points; Is DFT coefficient amplitude matrix poincare half plane that the summit is divided into N with the initial point _bIndividual sector region, each fan-shaped drift angle are 0.5 ° or 1 °; By angle all local maximum points are included into each sector region respectively again;

D) find and two possible template point set that the template line is corresponding;

In each sector region, at K _Min＜K＜K _MaxThe such K value of search in the scope: it makes to have N at least in this sector _mIndividual local maximum point satisfies $|r_{\mathrm{li}}-K{r}_{\mathrm{Tj}}^{\&prime;}|<\mathrm{threshold},$ N wherein _mBe a number of predesignating, r _LiBe the utmost point footpath (i=1...N of Local Extremum among the i of sector _b), r _Tj' be the utmost point footpath of touching plate point on the grand master pattern printed line j (j=1,2), threshold＞0 is a threshold value.If find such K value, we just get off corresponding Local Extremum coordinate record;

By above-mentioned steps, obtain the set of possible matched line, be called " accurate matched line ", the Local Extremum on the line is called " accurate match point ", and coordinate is designated as (x _Ij, y _Ij); The coordinate of the corresponding primary template point of image poincare half plane is designated as (x _Ij', y _Ij'), wherein { 1,2} represents i bar template matches line to i ∈, and j ∈ { 1,2, Λ } represents j template matches point; Take out a set and concentrate another set of taking-up from concentrating corresponding to the accurate match point of template line 2 corresponding to the accurate match point of template line 1; A possible transformation matrix A who calculates according to the point and the corresponding relation between primary template point of these two set; Seek 4 of average error MAE minimum;

$\mathrm{MAE}=\frac{1}{\mathrm{nummatches}}\left|\right|A{\left(\begin{array}{cc}{x}_{11}& y_{11}\\ M& M\\ x_{1j}& y_{1j}\\ x_{21}& y_{21}\\ M& M\\ x_{2j}& y_{2j}\end{array}\right)}^T-{\left(\begin{array}{cc}{x}_{11}^{\&prime;}& y_{11}^{\&prime;}\\ M& M\\ x_{1j}^{\&prime;}& x_{1j}^{\&prime;}\\ x_{21}^{\&prime;}& y_{21}^{\&prime;}\\ M& M\\ x_{2j}^{\&prime;}& y_{2j}^{\&prime;}\end{array}\right)}^T\left|\right|$

Wherein template point is (x _Ij', y _Ij') and " accurate match point " be (x _Ij, y _Ij), nummatches is the match point number, operational symbol || Λ || in be the error matrix of one 2 row;

F) will add 180 ° corresponding to the accurate match point of template line 1, repeat e), determine last frequency domain transform matrix A by the MAE value of minimum; Can get spatial domain transformation matrix B=A ^T

After obtaining affine transformation matrix B, (x y) carries out the image geometry inverse transformation and reverts to the visual g ' of M * N size (x y), and then fills 0 one-tenth M with visual g to be measured ₁* M ₁(x y), is filled with 0 by the part of cutting the visual I of size, and (x is y) at visual I (x, y) center for g;

The heavy second synchronous step is: translation is synchronous, promptly determines the translation synchronization parameter of image with the related coefficient of training sequence S that extracts and original training sequence T;

The translation method for synchronous be with visual I (x, y) do maximum 8 * 8=64 time translation and get final product:

I _t(x，y)＝I((x-x _t)mod N ₂，(y-y _t)mod N ₂)；{-4≤x _t，y _t＜4}

Every translation once is DWT and is decomposed, and obtains LL ₃Subband LL _3t(x, y); With LL _3t(x, y) do translation:

$L{L}_{3t}^{\&prime;}(x,y)=L{L}_{3t}((x-x_{t1})\mathrm{mod}64,(y-y_{t1})\mathrm{mod}64);$ {-T ₁≤x _t1＜T ₁；-T ₂≤y _t1＜T ₂}

In the following formula, T ₁=round (0.5 * (M ₁-M)/8), T ₂=round (0.5 * (M ₁-N)/8), each translation is from LL ₃', (x, central row y) and central series extract training sequence S, according to and original training sequence T between maximum related value determine translation parameters, can determine visual translation parameters (8 * x after maximum 64 translations search _t+ x _T1, 8 * y _t+ y _T1), thereby the synchronous image g after the acquisition translation calibration ^*(x, y);

The detection of DWT territory watermark: synchronous image g ^*(x, the low frequency sub-band LL after y) DWT decomposes ₃Coefficient becomes one-dimension array by line scanning, is designated as C ^*The binary data that extracts is designated as X ^*={ x _i ^*, it is as follows to extract formula:

$\left\{\begin{array}{c}{x}_i^{*}=+1,\left(C^{*}\left(i\right)\mathrm{mod}\mathrm{\&alpha;}\right)\&GreaterEqual;\frac{\mathrm{\&alpha;}}{2}\\ x_i^{*}=-1,\mathrm{otherwise}\end{array}\right.$

With the binary data X that extracts ^*Carry out reciprocal cross and knit the binary data sequence W that (inverse process that interweaves) recovers embedding ^*W then ^*Press N ₁The position bit carries out segmentation, and every section is carried out relevantly with sequence m, if correlation is greater than 0, then to embed information bit be " 1 " in judgement, is " 0 " otherwise adjudicate the embedding information bit; Just obtain the embedding information recovered after the despreading.

2. the method for a kind of image geometry calibration as claimed in claim 1 and protection digital image, it is characterized in that in the amplitude spectrum coefficient of visual DFT conversion embedding a matching template that constitutes a little louder by local pole, be used to detect image stand the transformation matrix of affined transformation and recover its original geometry do inverse transformation; Embed a training sequence simultaneously in visual DWT territory and be used for visual translation calibration.

Images