CN101556686A - Robust image watermark method based on local Tchebichef moment - Google Patents

Abstract

The invention discloses a robust image watermark new method based on a local Tchebichef moment, mainly solving the problem that a watermark arithmetic based on image characteristic at present can not effectively resist general image processing and geometry attack simultaneously. The method comprises the steps as follows: firstly, extracting characteristic points of an original image by utilizing a Harris-Laplace detection operator, and obtaining a group of stable and mutually independent round characteristic areas through characteristic selection; then obtaining a geometrically unchanged round characteristic area through principle direction alignment; and finally, calculating the Tchebichef moment in an inscribed square image block of the round characteristic area, and embedding a watermark through dithering quantized modulation to modify an amplitude value with low-level Tchebichef moment. In detection, the modified low-level Tchebichef moment is obtained through the dithering quantized modulation, and a minimum distance is utilized to realize the blind extraction of the watermark. The invention has favorable nonvisibility, excellent robust performance for general image processing, geometry attach and combined attach and superior performance to similar methods and can be used for copyright protection for different images.

Description

Robust image watermark method based on local Tchebichef moment

Technical field

The present invention relates to field of information security technology; specifically a kind of digital figure watermark embeds and blind checking method; this method is highly resistant to geometric attack and normal image is handled, and can be used for copyright protection, entitlement checking, the copy control field of digital picture works on the internet.

Background technology

The access that is digitized as multimedia messages of multi-medium data such as image, video, audio frequency etc. provides very big facility, has also greatly improved the efficient and the accuracy of information representation simultaneously.Universal day by day along with the internet, the interchange of multimedia messages has reached the unprecedented degree of depth and range especially, and its issue form is also more and more abundanter.People can carry out copy, download and the issue of Digital Media easily by the internet.This gives people's the live and work condition of providing convenience on the one hand, has improved work efficiency, also provides convenience for the piracy of Digital Media on the other hand.Some lawless persons arbitrarily copy and propagate the Digital Media publication of copyright protection under the situation that does not obtain the copyright owner of Digital Media mandate, therefrom speculate.In addition, public document, bank statement and personal credit data are maliciously tampered on the internet, make E-Government, ecommerce can not obtain smooth promotion and application.Therefore, how to protect the copyright problem of Digital Media and information security to become an extremely urgent realistic problem on the internet.Digital watermarking becomes the common focus of paying close attention to of business circles and academia recent years as a very potential solution.Digital watermark technology has remedied the defective of cryptographic technique on the one hand, because it can provide further protection for the data after the deciphering; On the other hand, digital watermark technology has also remedied the defective of digital signature technology because it can be in raw data a large amount of secret information of disposable embedding.

The basic thought of digital watermark technology is to embed secret information in multi-medium data, so that the true and reliable property of the copyright of protection digital product, proof product, follow the tracks of copy right piracy or the additional information of product is provided.The Image Watermarking Technique of robust must possess the ability that the multiple watermark of opposing is attacked.Handle as noise, filtering, compression etc. with respect to normal image, geometric attack such as translation, rotation, convergent-divergent etc. are difficult to resist more.Geometric attack does not destroy image watermark itself, but has destroyed the synchronized relation between watermarking images to be detected and the embed watermark information, causes watermark detector can't detect watermark information.

The image watermark method of resist geometric attacks can be divided into non-blind Detecting water mark method and blind Detecting water mark method.Non-blind Detecting water mark method need be used and be subjected to bigger restriction by original image or under fire watermarking images not.The blind Detecting water mark method roughly is divided into based on how much fields of invariants, based on supplementary with based on the water mark method of characteristics of image.Wherein, preceding two kinds of methods are only to global change---and rotation, convergent-divergent and translation are effective, can't resist comparatively complicated geometric attack such as local bending, shearing and combination attacks.By contrast, water mark method based on characteristics of image belongs to second generation digital watermark technology, its basic thought is to utilize that metastable unique point identifies the watermark embedded location in the image, and in the regional area corresponding with each unique point embed watermark independently, still utilize unique point to locate and detect watermark during detection, thereby reach the purpose of opposing geometric attack.

Have the characteristics of image of geometric invariance by structure, identify the accurate position that watermark embeds and detects then, can offset the watermark synchronization error that geometric attack causes.Characteristics of image can be overall, also can be local.In the resist geometric attacks water mark method based on global characteristics, general method is to utilize square that image is carried out normalized, and watermark is embedded in the image after the normalization.Document M.Alghoniemy and A.H.Tewfik.Geometric invariance inimage watermarking.IEEE Trans.Image Process. for example, 2004,13 (2), 145-153. utilize geometric moment that image is carried out normalization, carry out watermark in the image after normalization and embed and detect and to resist attacks such as convergent-divergent, rotation.Yet the maximum deficiency of these class methods is to resist shearing attack, and this is that the image section contents lost will inevitably cause the calculating of square value very mistake to occur because the calculating of square depends on all pixel values of entire image.

In resist geometric attacks image watermark method based on local feature, by the location of space characteristics point, with the unique corresponding sub-image area of unique point in to embed and extract watermark be common method.Because the unique point of choosing in the image has enough stability and distribution is more even, thereby these class methods can effectively be resisted shearing attack.Document J.S.Jeo and C.D.Yoo.Localized image watermarking based on feature points of scale-space representation.Pattern Recognit. for example, 2004,37 (7): 1365-1375. adopts Harris-Laplace to detect the operator extraction unique point, with the unique point is central configuration circular feature zone, and the two-dimensional circular watermark is directly embedded in the characteristic area.Water mark method based on the local feature zone is being obtained effect preferably aspect the resist geometric attacks, but has following problem: (1) geometric attack causes the extraction of image characteristic point skew to occur; (2) interpolation error of geometric attack process generation can change the pixel value of image.These problems will influence the performance of watermark detector, cause the verification and measurement ratio of watermark lower.

Summary of the invention

The object of the present invention is to provide a kind of based on local Tchebichef moment (Local Techebichef Moments, LTMs) robust image watermark method, can not resist shearing attack and the water mark method watermark detection rate problem on the low side of utilizing local feature to overcome the above-mentioned water mark method of global characteristics that utilizes, improve robustness geometric attack and normal image processing.

Realize that technical scheme of the present invention is: the advantage of above two class water mark methods is organically combined.When watermark embeds, at first utilize Harris-Laplace to detect the operator extraction image characteristic point, obtain one group of stable and circular feature independent of each other zone by the feature selecting strategy; Each circular feature zone direction of passage alignment is obtained geometric invariance; Utilize the Tchebichef square to extract the global characteristics in each circular feature zone, range value by jitter quantisation modulation low order Tchebichef square is realized embed watermark, and the range value by jitter quantisation low order Tchebichef square, adopt the mode of minor increment decoding to detect watermark.Detailed process is as follows:

One, based on the robust image watermark embedding grammar of local Tchebichef moment, comprises the steps:

A. generate the pseudorandom watermark sequence b={b of a two-value by key ₁, b ₂..., b _L, b _i∈ 0,1};

B. utilize Harris-Laplace to detect operator extraction image primitive character point, the unique point that will belong in the intermediate features range scale according to the characteristic dimension of primitive character point chooses, and is central configuration candidate circles characteristic area with these unique points;

C. by feature selecting strategy the candidate circles characteristic area is carried out cluster according to the distance restraint condition based on the minimum spanning tree clustering algorithm, to each category feature zone, only select the zone of intensity maximum, obtain stable and circular feature independent of each other zone thus;

D. utilize the single order local derviation to calculate the main gradient direction in each circular feature zone, and these main gradient directions are alignd by rotation;

E. get from postrotational circular feature zone and connect square in it, calculating connects foursquare Tchebichef square in being somebody's turn to do, and utilizes the mode of jitter quantisation modulation watermark to be embedded in the range value of low order Tchebichef square;

F. to the Tchebichef square be reconstructed obtain to contain watermark in connect square, and with these contain watermark in connect square image block and replace one by one and connect square image block in original, obtain containing watermarking images.

In the above-mentioned image watermark embedding grammar, step B described " is central configuration candidate circles characteristic area with these unique points ", carry out as follows:

B1. utilize Harris-Laplace to detect the primitive character point of operator extraction image, be designated as set omega ₁

B2. put pairing characteristic dimension according to primitive character, select the unique point in the medium scale scope, be designated as set omega ₂

B3. with set omega ₂In each unique point be the center, be radius structure circular feature zone with the τ of each unique point characteristic dimension doubly, its radius is expressed as:

Here, the characteristic dimension of σ presentation video unique point,

Expression rounds operation, and τ is a positive integer.

In the above-mentioned image watermark embedding grammar, step e described " utilizing the mode of jitter quantisation modulation that watermark is embedded in the range value of low order Tchebichef square ", carry out as follows:

The maximum order of E1. establishing low order Tchebichef square is O _Max, the set that these Tchebichef squares are formed is expressed as:

S＝{T _pq，p+q＜O _max，p，q≠0}， (2)

Wherein, p+q represents exponent number, T _PqThe Tchebichef square of representing the p+q rank;

E2. pass through key K ₁From S set, select L low order Tchebichef square at random: $T=(T_{p_1{\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">q</figure-callout>}}_1},...,T_{p_L{q}_L});$

E3. for watermark sequence b={b ₁, b ₂..., b _LIn every watermark b _i, the mode that adopts jitter quantisation to modulate is revised the range value of low order Tchebichef square, to realize the embedding of watermark, to revise rule is:

$\left|{\tilde{T}}_{p_i{q}_i}\right|=\left[\frac{\left|T_{p_i{q}_i}\right|-d_i\left(b_i\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(b_i\right),---\left(3\right)$

Wherein, [] is the operation that rounds up, and Δ is a quantization step, d _i() is i quantization function, and satisfies d _i(1)=Δ/2+d _i(0); Vector (d ₁(0) ..., d _L(0)) passes through key K ₂Produce, and upward obey even the distribution in interval [0, Δ].

Above-mentioned image watermark embedding grammar, step F described " to the Tchebichef square be reconstructed obtain to contain watermark in connect square ", carry out as follows:

F1. the range value by L amended low order Tchebichef square obtains L amended Tchebichef square, and it is expressed as:

$T_{p_i{q}_i}=\frac{\left|{\tilde{T}}_{p_i{q}_i}\right|}{\left|T_{p_i{q}_i}\right|}{T}_{p_i{q}_i},$ i＝1，...，L， (4)

Wherein,

Be the range value of amended i low order Tchebichef square,

Range value for i original low order Tchebichef square;

F2. in S set, utilize in first group of the Tchebichef square reconstruct that be not modified and meet square image block f _Rest(x, y):

f _rest(x，y)＝f(x，y)-f _T(x，y)， (5)

Wherein, (x is to connect square image block, f in original y) to f _T(x, y) be by Tchebichef square to be revised reconstruct in connect square image block;

F3. in S set, utilize in second group of L the amended Tchebichef square reconstruct and connect square image block

F4. in first group, meet square image block f _Rest(x y) He in second group connects square image block

Merge, obtain containing watermark in connect square image block

$\tilde{f}(x,y)=f_{\mathrm{rest}}(x,y)+f_{\tilde{T}}(x,y).---\left(6\right)$

Two, based on the robust image watermark detection method of local Tchebichef moment, comprise the steps:

G. utilize Harris-Laplace to detect operator and from image to be detected, extract image characteristic point,, obtain stable and circular feature independent of each other zone through after the feature selecting;

H. each circular feature zone is calculated its main gradient direction and is carried out main gradient direction alignment by rotation, in postrotational circular feature zone, get and connect square in it, and calculate should in connect foursquare Tchebichef square;

I. select L low order Tchebichef square, by minor increment decoding extract watermark b '=b ' ₁, b ' ₂..., b ' _L;

J. compare original watermark b={b ₁, b ₂..., b _LWith the watermark b ' that extracts=b ' ₁, b ' ₂..., b ' _L, whether the watermark figure place r that obtains mating, and compare with predefined matching detection threshold value T judges in this circular feature zone embed watermark, when r 〉=T, then this circular feature zone has embedded watermark; As r＜T, then this circular feature zone does not have embed watermark.In the above-mentioned image watermark detection method, step I carries out as follows:

I1. use the key K identical with telescopiny ₁, select L low order Tchebichef square in the square image block interior connecing: $T^{\&prime;}=(T_{p_1{\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">q</figure-callout>}}_1}^{\&prime;},...,T_{p_L{q}_l}^{\&prime;});$

I2. utilize identical key K ₂Produce quantization vector (d ₁(0) ..., d _LAnd (d (0)) ₁(1) ..., d _L(1));

I3. utilize the jitter quantisation modulation system identical, to each low order Tchebichef square with watermark embed process

Range value make amendment:

${\left|T_{p_i{q}_i}^{\&prime;}\right|}_j=\left[\frac{\left|T_{p_i{q}_i}^{\&prime;}\right|-d_i\left(j\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(j\right),---\left(7\right)$

Wherein, Δ is a quantization step, and [] is rounding operation, i=1 ..., L, j=0,1;

I4. by comparing

With its two groups quantification formulas

With

Between distance, extract L position watermark information:

$b_i^{\&prime;}=\arg \underset{j\&Element;\left\{\mathrm{0,1}\right\}}{\min }{({\left|T_{p_i{q}_i}^{\&prime;}\right|}_j-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,i=\mathrm{1,2},...,L---\left(8\right)$

Wherein, will

And two groups of distances that quantize formula are defined as respectively $\mathrm{dist}0={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_0-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2$ With $\mathrm{<figure-callout\; id="1"\; label="dist"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">dist</figure-callout>}1={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_1-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,$ And calculate t=dist0-dist1, if t＜0, then b ' _i=0, otherwise b ' _i=1.

The present invention is because the local feature of combining image is regional and global characteristics is represented, utilize Harris-Laplace to detect operator and the feature selecting strategy obtains one group of stable and circular feature independent of each other zone, effectively improved system the geometric attack robustness of shearing attack particularly; Simultaneously owing to utilize the Tchebichef square to represent the global statistics characteristic in circular feature zone, not only overcome because of the low problem of watermark detection rate that unique point is offset and interpolation error causes, and utilized good noise immunity of Tchebichef square and resistance to compression to strengthen the resistivity that normal image is handled; Adopt the method for jitter quantisation modulation when embedding and detecting owing to watermark in addition, both realized the blind Detecting of watermark, improved the accuracy of detection of watermark again.

Description of drawings

Fig. 1 is a The general frame of the present invention;

Fig. 2 is that watermark of the present invention embeds FB(flow block);

Fig. 3 is the synoptic diagram of the present invention's circular feature zone selection course in watermark embed process, wherein Fig. 3 (a) is an original image, 3 (b) are original circular feature areal map, Fig. 3 (c) is the circular feature areal map in the medium scale scope, and Fig. 3 (d) is the circular feature areal map after selecting through the minimum spanning tree clustering method;

Fig. 4 is the synoptic diagram of the embedding zone forming process of the present invention in watermark embed process, wherein Fig. 4 (a) is that desirable watermark embeds the zone, Fig. 4 (b) is original external square image block, and Fig. 4 (c) is the image block that " zero padding " forms, and Fig. 4 (d) connects square image block in being;

Fig. 5 is a watermark detection FB(flow block) of the present invention;

Fig. 6 is the synoptic diagram that concerns between PSNR value of the present invention, quantization step and the watermark length three;

Fig. 7 is that watermark of the present invention embeds the effect synoptic diagram, and wherein Fig. 7 (a) is an original image, and Fig. 7 (b) is for containing watermarking images, and Fig. 7 (c) is an error image;

Fig. 8 is that detection threshold of the present invention and false-alarm probability concern synoptic diagram.

Specific embodiments

With reference to The general frame shown in Figure 1, the present invention at first utilizes the unique point of Harris-Laplace detection operator extraction original image, and obtains one group of stable and circular feature independent of each other zone by feature selecting; Obtain constant circular feature zone how much by the principal direction alignment then; At last, in the circular feature zone, connect and calculate the Tchebichef square in the square image block, and realize that by the range value of jitter quantisation modulation modification low order Tchebichef square watermark embeds; During detection, obtain amended low order Tchebichef square and utilize the minor increment decoding to extract watermark by the jitter quantisation modulation, and compare to judge whether to exist watermark with predefined detection threshold.It is following that the present invention is described in further detail with reference to accompanying drawing.

See figures.1.and.2, watermark embed step of the present invention is as follows:

Step 1: utilize Harris-Laplace to detect the operator extraction image characteristic point.

To original image Fig. 3 (a), utilize Harris-Laplace to detect operator extraction circular feature zone, the initial circular characteristic area is shown in Fig. 3 (b).Harris-Laplace detects operator and comprises dimension self-adaption Harris operator and automatic yardstick selection technology two parts, and the unique point of extraction has robustness preferably to rotation, convergent-divergent, translation and noise etc., and concrete steps are:

At first, given metric space, ${\mathrm{\&sigma;}}_I^{\left(n\right)}={1.4}^n\&CenterDot;{\mathrm{\&sigma;}}_0,$ ${\mathrm{\&sigma;}}_D^{\left(n\right)}=0.7{\&CenterDot;\mathrm{\&sigma;}}_I^{\left(n\right)}$ (n=1,2 ..., 15) and threshold value 1000, wherein σ ₀The expression initial gauges utilizes dimension self-adaption Harris operator to calculate the candidate feature point of image, and the autocorrelation matrix M of dimension self-adaption Harris operator is expressed as:

$M(x,y,{\mathrm{\&sigma;}}_I,{\mathrm{\&sigma;}}_D)={\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="D"\; filenames="A20091002174400182.png,A20091002174400192.png"\; state="\{\{state\}\}">D</figure-callout>}}^2\&CenterDot;g\left({\mathrm{\&sigma;}}_I\right)\&CircleTimes;\left[\begin{array}{cc}{L}_x^2(x,y,{\mathrm{\&sigma;}}_D)& L_x{L}_y(x,y,{\mathrm{\&sigma;}}_D)\\ L_x{L}_y(x,y,{\mathrm{\&sigma;}}_D)& L_y^2(x,y,{\mathrm{\&sigma;}}_D)\end{array}\right]---\left(1\right)$

$L(x,y,{\mathrm{\&sigma;}}_D)=g\left({\mathrm{\&sigma;}}_D\right)\&CircleTimes;I,---\left(2\right)$

Wherein, σ _IAnd σ _DRepresent integral scale and differential yardstick respectively; G (σ) expression average is that 0 variance is the Gaussian function of σ; L is that Gauss's metric space of image is represented; L _sThe partial derivative of expression L on the s direction; I is an input picture;

The expression linear convolution.Given σ _IAnd σ _D, can determine point (x, characteristic strength R y):

R(x，y，σ _I，σ _D)＝Det(M(x，y，σ _I，σ _D))-κ·Tr ²(M(x，y，σ _I，σ _D))， (3)

Here, the determinant and the mark of Det () and Tr () difference representing matrix; κ is a constant, gets 0.04 usually.On each yardstick plane, unique point can be according to following Rule Extraction:

$R(x,y,{\mathrm{\&sigma;}}_I,{\mathrm{\&sigma;}}_D)>R(x,y,{\mathrm{\&sigma;}}_I,{\mathrm{\&sigma;}}_D)\&ForAll;(\hat{x},\hat{y})\&Element;A,$ (4)

R(x，y，σ _I，σ _D?)≥Th

Wherein, A remarked pixel point (x, a certain neighborhood y); Th represents threshold value.

Then, to each candidate feature point, adopt process of iteration to determine unique point and characteristic dimension thereof:

Step 1: establish x _kBe the candidate image unique point, whether check LoG operator can obtain local extremum at this some place in whole yardstick hunting zone, if can not obtain extreme value, then give up this point, and the yardstick hunting zone is defined as ${\mathrm{\&sigma;}}_I^{(k+1)}=t{\&CenterDot;\mathrm{\&sigma;}}_I^{\left(k\right)},$ T=0.7 ..., 1.4, the LoG operator representation is:

$\left|\mathrm{LoG}(x,y,{\mathrm{\&sigma;}}_I)\right|={\mathrm{\&sigma;}}_I^2|L_{\mathrm{xx}}(x,y,{\mathrm{\&sigma;}}_I)+L_{\mathrm{yy}}(x,y,{\mathrm{\&sigma;}}_I)|,---\left(5\right)$

Here, L _XxThe second-order partial differential coefficient of expression L on the x direction, L _YyThe second-order partial differential coefficient of expression L on the y direction;

Step 2: the image characteristic point x that can obtain extreme value for the LoG operator _k, the unique point x of search characteristics intensity maximum in this neighborhood of a point _K+1, if x _K+1Exist and then give up x _k

Step 3: repeat Step1～Step2, up to ${\mathrm{\&sigma;}}_I^{(k+1)}={\mathrm{\&sigma;}}_I^{\left(k\right)}$ Or x _K+1=x _kTill.

Step 2: the structure of part-circular characteristic area and selection.

Unique point only comprises positional information and characteristic dimension information, can not carry out the embedding and the extraction of watermark, need be central configuration suitable feature zone with the unique point therefore.Here, be the center of circle with the unique point, and be radius with the multiple of the pairing characteristic dimension of unique point, structure circular feature zone, the radius in this circular feature zone is:

Wherein, the characteristic dimension of σ presentation video unique point;

Expression rounds operation; τ is a positive integer, is used to regulate the size in circular feature zone.Because characteristic dimension can reflect the variation of picture size, thereby the circular feature of structure zone can covariant change in the convergent-divergent of image.

Harris-Laplace detects operator and is used for images match and Target Recognition at first, and detected unique point is very intensive, and the characteristic area of formation overlaps serious.In order to strengthen robustness of the present invention, need to select one group of stable and circular feature independent of each other zone to be used for watermark and embed and detect, this selection course is as follows:

Step 1: utilize Harris-Laplace to detect the initial characteristics point of the original image of operator extraction shown in Fig. 3 (a), be designated as set omega ₁, shown in Fig. 3 (b);

Step 2: the unique point that characteristic dimension is little, and its stability is lower, and the characteristic area that the big unique point of characteristic dimension forms overlaps seriously, therefore at first selects the unique point in the medium scale scope, is designated as set omega ₂, shown in Fig. 3 (c);

Step 3: with set omega ₂Interior unique point is the center, τ with its characteristic dimension doubly is radius structure circular feature zone, less zone can increase the robustness of watermark, but can limit watermark capacity, though and bigger zone can increase the embedding capacity but can reduce the robustness of watermark, so τ should the compromise robustness and the watermark capacity of watermark;

Step 4: in order to obtain characteristic area independent of each other, distance between the unique point also needs to consider: characteristic distance is little, region overlapping is serious, characteristic distance is big, regional number deficiency, and the present invention introduces distance restraint D, utilize the minimum spanning tree clustering method that unique point is carried out cluster, to same category feature point, select the unique point of characteristic strength maximum and form the circular feature zone, shown in Fig. 3 (d).

Step 3: the circular feature zone that constructive geometry is constant

Through after the above-mentioned steps, obtained one group of stable and circular feature independent of each other zone, these zones have unchangeability to translation, convergent-divergent.In order to realize the unchangeability of circular feature zone to rotation, the present invention calculates each circular feature zone principal direction by the gradient direction distribution character of unique point neighborhood territory pixel, and alignment realizes rotational invariance through principal direction.During actual computation, in the neighborhood window that with the unique point is the center, sample, and with the gradient direction of statistics with histogram neighborhood territory pixel, histogrammic peak value has been represented the principal direction of this unique point, also as the principal direction in corresponding circular feature zone.Image I mid point (x ₀, y ₀) gradient located is:

$\&dtri;I(x_0,y_0)={[(\&PartialD;I/\&PartialD;x),(\&PartialD;I/\&PartialD;y)]|}_{({\mathrm{<figure-callout\; id="0"\; label="x"\; filenames="A20091002174400192.png,A20091002174400212.png"\; state="\{\{state\}\}">x</figure-callout>}}_0,y_0)},---\left(7\right)$

Here, the amplitude of gradient is

The direction of gradient is

Step 4: choose the circular feature zone in connect square image block and calculate the Tchebichef square

For embed watermark in the circular feature zone shown in Fig. 4 (a), during actual treatment, at first obtain comprising the external square image block in circular feature zone, shown in Fig. 4 (b); The image block of " zero padding " acquisition shown in Fig. 4 (c) around passing through then.Consider that " zero padding " zone should keep the covariant characteristic to the influence of calculating Tchebichef square and the zone of final embed watermark, get in the circular feature zone of the present invention after the principal direction alignment and connect square region in it, shown in Fig. 4 (d).Consider that Tchebichef has good feature representation ability and stronger noise inhibiting ability, the present invention calculates the Tchebichef square interior connecing in the square region.

Step 5: embed watermark in low order Tchebichef square

Step 1. generates the pseudorandom watermark sequence b={b of a two-value by key ₁, b ₂..., b _L, b _i∈ 0,1};

The maximum square exponent number of Step 2. low order Tchebichef squares is O _Max, then the set of these Tchebichef squares formations is expressed as:

S＝{T _pq，p+q＜O _max，p，q≠0}， (8)

For invisibility and the robustness of compromising, should select suitable low order Tchebichef square embed watermark;

Step 3. passes through key K ₁From S set, select L low order Tchebichef square at random: $T=(T_{p_1{\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">q</figure-callout>}}_1},...,T_{p_L{q}_L})$ Be used for embed watermark;

Step 4. is for each watermark bit b _i, adopt the method for jitter quantisation modulation low order Tchebichef square range value to realize that watermark embeds, its embedding rule is:

$\left|{\tilde{T}}_{p_i{q}_i}\right|=\left[\frac{\left|T_{p_i{q}_i}\right|-d_i\left(b_i\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(b_i\right),---\left(9\right)$

In the formula, [] is the operation that rounds up; Δ is a quantization step; d _i() is i quantization function, and satisfies d _i(1)=Δ/2+d _i(0); Vector (d ₁(0) ..., d _L(0)) passes through key K ₂Produce, and upward obey even the distribution in interval [0, Δ];

Step 5. is expressed as amended Tchebichef square:

${\tilde{T}}_{p_i{q}_i}=\frac{\left|{\tilde{T}}_{p_i{q}_i}\right|}{\left|T_{p_i{q}_i}\right|}{T}_{p_i{q}_i},$ i＝1，...，L， (10)

Wherein,

Be the range value of amended i low order Tchebichef square, Range value for i original low order Tchebichef square.

Step 6: will contain watermark circular feature zone and replace original circular feature zone one by one, and obtain to contain watermarking images.

At first, will contain watermark in connect square image block and be divided into two groups: first group of Tchebichef square reconstruct of not being modified of serving as reasons obtain in meet square image block f _Rest(x, y):

f _rest(x，y)＝f(x，y)-f _T(x，y)， (11)

Here, (x is to connect square image block in original y) to f; f _T(x is by connecing square image block in the Tchebichef square reconstruct to be revised y).Second group connects square image block in being obtained by the reconstruct of amended Tchebichef square

Then, merge connecing square image block in above two groups, obtain containing watermark in connect square image block

$\tilde{f}(x,y)=f_{\mathrm{rest}}(x,y)+f_{\tilde{T}}(x,y).---\left(12\right)$

At last, with all contain watermark in connect square image block and replace and connect square image block in original, and " zero-suppressing " obtains containing watermark circular feature zone around will containing the external square image block of watermark, recover through principal direction and replace all original circular feature zones after can obtain containing watermarking images.

With reference to Fig. 1 and Fig. 5, watermark detection step of the present invention is as follows:

Step 1: utilize Harris-Laplace to detect operator and from image to be detected, extract image characteristic point and carry out feature selecting.

Utilize Harris-Laplace to detect operator and from image to be detected, extract image characteristic point.Because image may suffer certain attack, the unique point that extracts when therefore unique point of extracting and not sum embed is in full accord.Through after the identical feature selecting strategy, obtain one group of stable and circular feature independent of each other zone.Test verifiedly, the characteristic area that obtains through the feature selecting strategy has re-detection rate preferably, even image has suffered certain geometric attack, the circular feature zone that major part contains watermark still can be detected.

Step 2: the circular feature zone is carried out the principal direction alignment and calculated the Tchebichef square.

In order to realize the unchangeability of circular feature zone to rotation, the gradient direction distribution character by the unique point neighborhood territory pixel calculates each circular feature zone principal direction, and alignment realizes rotational invariance through principal direction.Get in the circular feature zone after each principal direction alignment and connect square image block in it and calculate the Tchebichef square.

Step 3: watermark detection

Step 1. uses the key K identical with telescopiny ₁, select L low order Tchebichef square $T^{\&prime;}=(T_{p_1{\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">q</figure-callout>}}_1}^{\&prime;},...,T_{p_L{q}_l}^{\&prime;})$ Be used for watermark extracting;

Step 2. utilizes the key key K identical with telescopiny ₂Produce quantization vector (d ₁(0) ..., d _LAnd (d (0)) ₁(1) ..., d _L(1));

Step 3. adopts the jitter quantisation modulation formula identical with telescopiny, to each low order Tchebichef square The amount of carrying out is revised:

${\left|T_{p_i{q}_i}^{\&prime;}\right|}_j=\left[\frac{\left|T_{p_i{q}_i}^{\&prime;}\right|-d_i\left(j\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(j\right),---\left(6\right)$

Wherein, [] is rounding operation, i=1 ..., L, j=0,1;

Step 4. relatively

And two groups of quantification formulas

With Between distance, extract watermark information:

$b_i^{\&prime;}=\arg \underset{j\&Element;\left\{\mathrm{0,1}\right\}}{\min }{({\left|T_{p_i{q}_i}^{\&prime;}\right|}_j-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,$ i＝1，2，...，L， (7)

Wherein, will

And two groups of distances that quantize formula are defined as respectively $\mathrm{dist}0={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_0-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2$ With $\mathrm{<figure-callout\; id="1"\; label="dist"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">dist</figure-callout>}1={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_1-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,$ And calculate t=dist0-dist1, if t＜0, then b ' _i=0, otherwise b ' _i=1.

Step 4: compare original watermark b={b ₁, b ₂..., b _LWith the watermark b ' that extracts=b ' ₁, b ' ₂..., b ' _L, obtain match bit array r, and compare with predefined detection threshold T, whether judge in this circular feature zone embed watermark, when r 〉=T, then this circular feature zone has embedded watermark; As r＜T, then this circular feature zone does not have embed watermark.The step of detection threshold T is as follows:

Step 1. is for the image of embed watermark not, the leaching process of watermark can be regarded the Bernoulli test as, therefore, the watermark bit of extracting is the pseudo-random variable of matching probability p=0.5, watermark sequence and the original watermark sequence extracted are compared, when coupling number during greater than certain detection threshold, just can think to have watermark information in the product to be detected, so the false-alarm error probability in each circular feature zone is:

$P_{\mathrm{FP}}=\underset{r=T_h}{\overset{L}{\mathrm{\&Sigma;}}}{\left(\frac{1}{2}\right)}^L\left(\frac{L!}{r!(L-r)!}\right),---\left(8\right)$

Wherein, T _hThe expression detection threshold, L represents watermark length, r represents the bit number of watermark matches;

Step 2. by law of great numbers as can be known, when L was very big, it was that m=Lp, variance are that the probability distribution of binary random variable is approximately average $\widehat{\mathrm{\&sigma;}}=\mathrm{Lp}(1-p)$ Gaussian distribution, therefore, formula (8) approximate representation is:

$P_{\mathrm{FP}}\&cong;\left\{\begin{array}{cc}1-\frac{1}{2\mathrm{\&pi;}}\frac{e^{-T^{\&prime;2}/2}}{(a-1)T^{\&prime;}+a\sqrt{T^{\&prime;2}+b}}& T_h<\mathrm{pL}\\ 0.5& T_h=\mathrm{<figure-callout\; id="1"\; label="pL"\; filenames="A20091002174400201.png"\; state="\{\{state\}\}">pL</figure-callout>}\\ \frac{1}{2\mathrm{\&pi;}}\frac{e^{-T^{\&prime;2}/2}}{(1-a)T^{\&prime;}+a\sqrt{T^{\&prime;2}+b}}& T_h>\mathrm{pL}\end{array}\right.---\left(9\right)$

Here, $T^{\&prime;}=(T_h-m)/\widehat{\mathrm{\&sigma;}},$ a＝1/π，b＝2π；

Step 3. is according to formula (9), can draw as shown in Figure 8 detection threshold and the relation curve of false-alarm probability, by this curve as can be known, as detection threshold T _h=140 o'clock, false-alarm probability P _Fp=9.13 * 10 ^-5, when promptly the coupling number of watermark of Ti Quing and original watermark reached 140, the false-alarm probability in each circular feature zone was P ^Fp=9.13 * 10 ^-5

Advantage of the present invention can further specify by following experiment:

The present invention has carried out test experiments on a large amount of standard grayscale images, comprising benchmark test image B aboon, Lena, Peppers.Baboon comprises smooth region and complex texture zone; Lena has the characteristics of image of number of different types; On behalf of brightness, Peppers change.With invisibility and robustness evaluation and test foundation as performance quality of the present invention.

(1) invisibility

The present invention is with the foundation of objective indicator PSNR as the evaluation invisibility.By analyzing as can be known, the PSNR value among the present invention mainly is subjected to the influence of two factors.When watermark length fixedly the time, the quantization step Δ in the jitter modulation influences the PSNR value, and the big more watermark strength of Δ is just big more, but the PSNR value on the contrary can be more little; When fixedly Δ or watermark strength, the watermark figure place is long more, and PSNR will be low more.

Fig. 6 has provided the relation between mean P SNR value, quantization step Δ and the watermark figure place three of image.When quantizing step delta=18, watermark length L=224 o'clock, whole PSNR value is all greater than 50dB.

Fig. 7 (a) is the original image of Baboon, Lena and Peppers three width of cloth images; Fig. 7 (b) is for containing watermarking images; Fig. 7 (c) is original image and the error image that contains watermarking images, the i.e. influence that watermark produces original image.Fig. 7 further illustrates the present invention and has good invisibility.

(2) robustness

Utilize 4.0 couples of evaluating tool Stirmark to contain watermarking images and carry out a series of attack experiments, to test robustness of the present invention.Table 1 and table 2 have provided the watermarking detecting results of the present invention to normal image processing and geometric attack respectively.As the index of estimating robustness of the present invention, the mark in the form is verification and measurement ratio with verification and measurement ratio for we.

$\mathrm{DR}=\frac{\#\det \mathrm{ected}\_\mathrm{number}}{\#\mathrm{original}\_\mathrm{number}}---\left(10\right)$

Here, #detected_number represents successfully to detect in the attack graph picture characteristic area number of watermark information; #original_number represents the characteristic area number of original image embed watermark.

The resistivity that table 1 digital watermarking is handled normal image

Table 2 digital watermarking is to the resistivity of geometric attack

Method 1 is a test result of the present invention in table 1 and the table 2, the result that method 2 obtains according to method in " C.W.Wang; H.M.Hang.Afeature-based robust digital image watermarking scheme.IEEE Trans.Signal Processing.; 2003; 51 (4): 950-959. ", the result that method 3 obtains according to method in " J.S.Seo; C.D.Yoo.Image watermarkingbased on invariant regions of scale-space representation.IEEE Trans.Signal Processing; 2006,54 (4): 1537-1549. ".

As can be seen from Table 1, the present invention is highly resistant to normal image and handles, and especially adds to make an uproar and the JPEG compression.This be because: the unique point that extract (1) has good stability, and after handling through normal image, the embedded location of watermark does not change, and has guaranteed that watermark information can correctly extract; (2) local Tchebichef moment itself is made an uproar and JPEG compression has good resistibility to adding.By contrast, need to carry out normalized before method 2 embed watermarks, will cause watermark information energy greater loss, reduced the recall rate of method.Method 3 is by the spatial domain direct embed watermark that superposes, thereby the resistibility that normal image is handled is relatively poor.

As can be seen from Table 2, this method has robustness preferably to geometric attack (particularly aspect ratio changes, ranks remove and random bend) and combination attacks, has all obtained higher recall rate under all situations.Method 2 is used fixing local feature area size, does not have covariance, greatly reduces the robustness of this method to wide-angle rotation, convergent-divergent and aspect ratio change and combination attacks.Though method 3 can be resisted most of geometric attack, but embed the recall rate that the tactful deficiency that exists has reduced this method, and change do not have robustness to aspect ratio.

Claims (6)

1. the robust image watermark embedding grammar based on local Tchebichef moment comprises the steps:

A. generate the pseudorandom watermark sequence b={b of a two-value by key ₁, b ₂..., b _L, b _i∈ 0,1};

B. utilize Harris-Laplace to detect operator extraction image primitive character point, the unique point that will belong in the intermediate features range scale according to the characteristic dimension of primitive character point chooses, and is central configuration candidate circles characteristic area with these unique points;

C. by feature selecting strategy the candidate circles characteristic area is carried out cluster according to the distance restraint condition based on the minimum spanning tree clustering algorithm, to each category feature zone, only select the zone of intensity maximum, obtain stable and circular feature independent of each other zone thus;

D. utilize the single order local derviation to calculate the main gradient direction in each circular feature zone, and these main gradient directions are alignd by rotation;

E. get from postrotational circular feature zone and connect square in it, calculating connects foursquare Tchebichef square in being somebody's turn to do, and utilizes the mode of jitter quantisation modulation watermark to be embedded in the range value of low order Tchebichef square;

F. to the Tchebichef square be reconstructed obtain to contain watermark in connect square, and with these contain watermark in connect square image block and replace one by one and connect square image block in original, obtain containing watermarking images.

2. robust image watermark embedding grammar according to claim 1, step B described " is central configuration candidate circles characteristic area with these unique points " wherein, carry out as follows:

B1. utilize Harris-Laplace to detect the primitive character point of operator extraction image, be designated as set omega ₁

B2. put pairing characteristic dimension according to primitive character, select the unique point in the medium scale scope, be designated as set omega ₂

B3. with set omega ₂In each unique point be the center, be radius structure circular feature zone with the τ of each unique point characteristic dimension doubly, its radius is expressed as:

Here, the characteristic dimension of σ presentation video unique point,

Expression rounds operation, and τ is a positive integer.

3. robust image watermark embedding grammar according to claim 1, step e described " utilizing the mode of jitter quantisation modulation that watermark is embedded in the range value of low order Tchebichef square " wherein, carry out as follows:

The maximum order of E1. establishing low order Tchebichef square is O _Max, these Tchebichef squares are constituted set:

S＝{T _pq，p+q＜O _max，p，q≠0}，

Wherein, p+q represents exponent number, T _PqThe Tchebichef square of representing the p+q rank;

E2. pass through key K ₁From S set, select L low order Tchebichef square at random: $T=(T_{p_1{q}_1},...,T_{p_L{q}_L});$

E1. for watermark sequence b={b ₁, b ₂..., b _LIn every watermark b _i, the mode that adopts jitter quantisation to modulate is revised the range value of low order Tchebichef square, to realize the embedding of watermark, to revise rule is:

$\left|{\tilde{T}}_{p_i{q}_i}\right|=\left[\frac{\left|T_{p_i{q}_i}\right|-d_i\left(b_i\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(b_i\right),$

Wherein, [] is the operation that rounds up, and Δ is a quantization step, d _i() is i quantization function, and satisfies d _i(1)=Δ/2+d _i(0), vector (d ₁(0) ..., d _L(0)) passes through key K ₂Produce, and upward obey even the distribution in interval [0, Δ].

4. robust image watermark embedding grammar according to claim 1, wherein step F described " to the Tchebichef square be reconstructed obtain to contain watermark in connect square ", carry out as follows:

F1. the range value by L amended low order Tchebichef square obtains L amended Tchebichef square, and it is expressed as:

${\tilde{T}}_{p_i{q}_i}=\frac{\left|{\tilde{T}}_{p_i{q}_i}\right|}{\left|T_{p_i{q}_i}\right|}{T}_{p_i{q}_i},i=1,...,L,$

Wherein,

Be the range value of amended i low order Tchebichef square,

Range value for i original low order Tchebichef square;

F2. in S set, utilize the Tchebichef square reconstruct that is not modified to obtain connecing in first group square image blocks f _Rest(x, y):

f _rest(x，y)＝f(x，y)-f _T(x，y)，

Wherein, (x is to connect square image blocks, f in original y) to f _T(x, y) be by Tchebichef square to be revised reconstruct in connect square image blocks;

F3. in S set, utilize L amended Tchebichef square reconstruct to obtain connecing square image blocks in second group

F4. in first group, meet square image blocks f _Rest(x y) He in second group connects square image blocks

Merge, obtain containing watermark in connect square image blocks

$\tilde{f}(x,y)=f_{\mathrm{rest}}(x,y)+f_{\tilde{T}}(x,y).$

5. the robust image watermark detection method based on local Tchebichef moment comprises the steps:

G. utilize Harris-Laplace to detect operator and from image to be detected, extract image characteristic point,, obtain stable and circular feature independent of each other zone through after the feature selecting;

H. each circular feature zone is calculated its main gradient direction and is carried out main gradient direction alignment by rotation, in postrotational circular feature zone, get and connect square in it, and calculate should in connect foursquare Tchebichef square;

I. select L low order Tchebichef square, by minor increment decoding extract watermark b '=b ' ₁, b ' ₂..., b ' _L;

J. compare original watermark b={b ₁, b ₂..., b _LWith the watermark b ' that extracts=b ' ₁, b ' ₂..., b ' _L, whether the watermark figure place r that obtains mating, and compare with predefined detection threshold T judges in this circular feature zone embed watermark, when r 〉=T, then this circular feature zone has embedded watermark; As r＜T, then this circular feature zone does not have embed watermark.

6. robust image watermark detection method according to claim 5, step I carries out as follows:

I1. use the key K identical with telescopiny ₁, select L low order in the square image blocks interior connecing

The Tchebichef square: $T^{\&prime;}=(T_{p_1{q}_1}^{\&prime;},...,T_{p_L{p}_l}^{\&prime;});$

I2. utilize identical key K ₂Produce quantization vector (d ₁(0) ..., d _LAnd (d (0)) ₁(1) ..., d _L(1));

I3. utilize the jitter quantisation modulation system identical, to each low order Tchebichef square with watermark embed process

Range value make amendment:

${\left|T_{p_i{q}_i}^{\&prime;}\right|}_j=\left[\frac{\left|T_{p_i{q}_i}^{\&prime;}\right|-d_i\left(j\right)}{\mathrm{\&Delta;}}\right]\mathrm{\&Delta;}+d_i\left(j\right),$

Wherein, Δ is a quantization step, and [] is rounding operation, i=1 ..., L, j=0,1;

I4. by comparing

With its two groups quantification formulas

With

Between distance, extract L position watermark information:

$b_i^{\&prime;}=\arg \underset{j\&Element;\left\{\mathrm{0,1}\right\}}{\min }{({\left|T_{p_i{q}_i}^{\&prime;}\right|}_j-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,i=\mathrm{1,2},...,L,$

Wherein, will

And two groups of distances that quantize formula are defined as respectively $\mathrm{dist}0={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_0-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2$ With $\mathrm{dist}1={({\left|T_{p_i{q}_i}^{\&prime;}\right|}_1-|T_{p_i{q}_i}^{\&prime;}\left|\right)}^2,$ And calculate t=dist0-dist1, if t＜0, then b ' _i=0, otherwise b ' _i=1.

Images