CN102892048A - Video watermark anti-counterfeiting method capable of resisting geometric attacks - Google Patents

Abstract

The invention belongs to the fields of computer vision and image understanding, and relates to a video watermark anti-counterfeiting method capable of resisting geometric attacks, which comprises the steps of: first, extracting a maximally stable extrema region (MSER) of a video frame; second, screening fitted elliptic regions, eliminating regional overlapping, and selecting an elliptic region for final watermark embedding; and third, by modifying a horizontal high frequency vertical low-frequency band (HL) and a horizontal low frequency vertical low-frequency band (LH) of a wavelet transform coefficient, implementing the watermark embedding on the selected characteristic region. The ability of resisting affine geometric attacks is enhanced. Furthermore, the invention discloses an inverse process of the watermark embedding, namely a method of watermark extracting. According to the method, the modification on the original video is effectively reduced; and an excellent vision effect is guaranteed even under malicious geometric attacks.

Description

A kind of video watermark method for anti-counterfeit of resist geometric attacks

Affiliated technical field

The invention belongs to computer vision and image understanding field, relate to a kind of video watermark method for anti-counterfeit of resist geometric attacks.

Background technology

In recent years, more senior and more easy-operating video editing/distort/attack tool is popularized just gradually, causes the news of forging, the problems such as evidence video that are used for judiciary that malice is distorted to emerge in an endless stream; On the other hand, become the Information Communication main carriers take forum, blog, blog, microblogging, social networks as the Society information net that representative was consisted of, its distinguishing feature is exactly that Information Communication is rapider.The safety of video content is being challenged in both combinations worse.Video watermark is one of effective technology means of protection video content safety and authenticity.Video watermark refers to watermark is embedded in the video sequence to prevent that video from suffering fraudulent copying and modification.For copyright protection, the video watermark system is intended to resist various malice and non-malicious attack.Although the research and development of video watermark technology for some time, the resist geometric attacks ability remains the difficult point of video watermark technology.

Early stage video watermark process is embedded into spatial domain by the pixel of revising the video single frames with watermark.These class methods are processed very sensitive to common geometric attack and image.Subsequently watermark begins to be embedded into frequency domain and strengthens the robustness that watermaking system is processed image.These class methods are come watermarked usually by the conversion coefficient of revising frame sequence, common transform method has discrete cosine transform, discrete Fourier transform, wavelet transform etc.But the method that is based on frequency domain is still attacked responsive to geometric transformation.

The video watermark embedding grammar of existing resist geometric attacks roughly can be divided into three major types: 1. how much not political reforms; 2. how much restoring methods; 3. based on the watermark method of feature.In first kind method, watermark carrier is how much fields of invariants.In order to reach the purpose of anti-translation, rotation, change of scale, general stochastic transformation territory is used to watermarked and can keeps synchronously [1] under geometric transformation.Certainly, also have many other how much fields of invariants [2] to be suggested, as: logarithm-polar coordinates territory, common square, Zernike square.Although these methods can be resisted rotation to a certain extent and yardstick is attacked, watermark is subject to the local geometric conversion and damages.The Equations of The Second Kind method is come watermarked by the parameter of estimating geometric transformation, these parameters are used to before watermark is extracted picture be carried out inverse transformation.Typical geometric transformation parameter can be estimated with template [3] [4].But the major defect that is based on template method is template to be easy under attackly and remove, and these fatal errors are so that based on the method for recovering for how much and impracticable.The 3rd class digital watermark take " based on the localized watermark of feature " as main feature is at first early proposed by [5] [6] [7] [8] such as Kutter.The main points of view of these class methods is impacts that picture feature is not subjected to geometric attack, and watermark information and picture feature are associated the integrality that can guarantee watermark information.But utilize the duplicate detection of picture feature to determine the position that watermark embeds and extracts based on the watermark embedded technology of feature.This class technology is extract minutiae from picture at first, then makes up the local feature zone based on characteristic point and comes watermarked.

The people such as Bas [9] propose content-based synchronization watermarking mechanism.They use the Harris detector to extract picture feature point and make up the Delaunay segmentation, then come watermarked with classical additivity mechanism in each spatial domain trigonum.This mechanism detected characteristics point and the original Delaunay of reconstruct after being attacked segment to detect watermark.Yet the Harris point is very sensitive to dimensional variation, so this method is subject to the impact that yardstick is attacked.

A kind of local invariant feature that cries yardstick invariant features conversion (SIFT) [10] is arranged, because its consistency to rotation, yardstick is widely used in content-based digital watermark [11].In [12], the SIFT feature is used to generate the circular piece as embedded unit.Watermark is embedded in the spatial domain of these unit, and this mechanism can be resisted more generally geometric attack.The Harris-Laplace detector also is used to make up border circular areas based on the Harris angle point with watermarked.But SIFT detector and Harris-Laplace detector [13] are subject to the impact of length-width ratio distortion.The weakness that above-mentioned content-based watermark embeds mechanism is that the embedding zone is very responsive to asymmetric change of scale.In order to address this problem, affine invariant point detector has further been put forward.The affine area detector of Harris is used for extracting affine covariant zone by people such as Hefei Ling in [14].

Maximum stable extremal region (MSER) possesses rotation, yardstick, Aspect Ratio variation, illumination variation consistency.In this piece patent, we have proposed the video watermark mechanism based on the MSER local feature of a stalwartness.We select MSER to be because: the first, absolutely mostly counting in the situations, such as visual angle change, dimensional variation, illumination variation etc., it is more stable that MSER compares the further feature territory; The second, MSER efficient is high, is close to the Pixel-level linear complexity in the leaching process.In addition, MSER has two good characteristics: the first, and it is closed that the MSER of acquisition is integrated under the continuous transformation of picture coordinate; The second, it is closed that the MSER of acquisition is integrated under the monotonic transformation of picture intensity.One efficient (being close to linear complexity) and practical fast algorithm of detecting are at Matas[15] be suggested in the original text.

In view of the multi-resolution characteristics of wavelet field can provide spatial domain and frequency domain feature to make it compatible with human visual system (HVS), we concentrate on wavelet transform (DWT) and add watermark to wavelet field HL and LH band rather than spatial domain.

The main reference document

[1] Dimitrios Simitopoulos, Dimitrios E.Koutsonanos, Michael Gerassimos Strintzis " based on the picture watermark of general stochastic transformation " image processes 2003.

[2] Picard Justin, Zhao Jian video watermark: the U.S., US2009220070A1[P] 2009-09-03.

[3] Shelby Pereira, Thierry Pun " based on the anti-affine transformation picture watermark of template matches " image processes 2000.

[4] Macy William W, Holliman Matthew J watermark decode synchronization template: the U.S., US6707926B1[P] 2004-03-16.

[5] M.Kutter, S.K.Bhattacharjee, T.Ebrahimi " towards second generation watermark mechanism " image is processed international conference 1999.

[6] Chih-Wei Tang, Hsueh-Ming Hang " a kind of digital picture watermark mechanism based on feature " signal processes 2003.

[7] Xiang-yang Wang, Li-min Hou, Jun Wu " a kind of digital picture watermark of the resist geometric attacks based on feature " picture and vision calculate 2008.

[8] N.R.Nantha Priya, S.Lenty Stuwart " based on the picture watermark process of feature " computer application International Periodicals 2010.

[9] Patrick Bas, Jean-Marc Chassery, and

Macq " the constant watermark of the geometry of use characteristic point " image processes 2002.

[10] David G.Lowe " based on the object identification of local yardstick invariant features " computer vision 1999.

[11] Lee Hae Yeoun, Lee Heung Kyu uses the picture watermark of yardstick invariant features conversion: Korea S, KR20070073332A[P] 2007-07-10.

[12] Viet Quoc PHAM, Takashi MIYAKI, Toshihiko YAMASAKI, Kiyoharu AIZAWA

The watermark mechanism based on how much constant objects with the SIFT feature " image processing international conference 2007.

[13] Krystian Mikolajczyk, Cordelia Schmid " yardstick and affine constant point of interest detector " computer vision International Periodicals 2004.

[14] Hefei Ling, Liyun Wang, Fuhao Zou, Zhengding Lu, Ping Li " based on the compressed domain video watermark of affine field of invariants " signal processes 2006.

[15] J Matas, O Chum, M Urban, T P ajdla " based on the wide Baseline Stereo of maximum stable extremal region " image and vision calculate 2004.

[16] Chun-Shien Lu, Hong-Yuan Mark Liao " based on the watermark of VS: the mechanism of a kind of anti-rotation and upset " image is processed international conference 2001.

Summary of the invention

The objective of the invention is to overcome the above-mentioned deficiency of prior art, propose a kind of at the superior video watermark method for anti-counterfeit of anti-affine geometry attack performance.Technical scheme of the present invention is as follows:

A kind of video watermark method for anti-counterfeit of resist geometric attacks comprises the following steps:

The first step: the maximum stable extremal region MSER that extracts frame of video:

(1) use the BINSORT sort algorithm that pixel in the frame of video according to pixels is worth size and arrange, and set up location index for the pixel after the ordering, each pixel after the ordering is considered as being communicated with assembly;

(2) use and to unite-to search algorithm and carry out merging growth to being communicated with assembly, record the area that each is communicated with assembly;

(3) select MSER.Under assign thresholds, from the connection assembly that (2) obtain, pick out the assembly area change and obtain the connection assembly of local minimum with changes of threshold, obtain maximum stable extremal region.

(4) for each maximum stable extremal region, computation of mean values u and covariance ∑, recycling elliptic region equation (x-u) ^T∑ ^-1(x-u)=1, fit to elliptic region according to the actual area shape;

Second step: the elliptic region that simulates is screened, eliminate region overlapping, choose the elliptic region that embeds for final watermark:

(1) the excessive too small elliptic region of filter area, the elliptic region area of reservation satisfies τ ₁＜| R _i|＜τ ₂, wherein, | R _i| represent the area of i elliptic region, τ ₁And τ ₂The lower bound and the upper bound that represent respectively size.

(2) design of graphics model, in graph model, the corresponding summit V of each elliptic region i _iThe summit is to V _i, V _jBetween by limit E _IjConnect; Limit E _IjWeight w _IjEqual the summit to V _i, V _jBetween distance, (V, E) consists of undirected complete graph G;

(3) using minimum spanning tree at undirected complete graph G) algorithm obtains minimum spanning tree T;

(4) judge according to elliptic equation whether every pair of characteristic area is overlapping among the T, if overlapping, then these two elliptic regions put identical cluster under, otherwise, put different clusters under;

(5) pick out the elliptic region of area maximum in each cluster, these elliptic regions will embed for final watermark;

The 3rd step: by revising the vertical low-frequency band of horizontal high frequency (HL) and the vertical high frequency band of horizontal low frequency (LH) of wavelet conversion coefficient, the characteristic area that selects is carried out watermark embed, method is as follows:

(1) elliptic region normalization: at first, to the contrary ∑ of the covariance of the elliptic region picked out ^-1Carry out singular value decomposition, ∑ ^-1=Q Λ Q ^-1, wherein, Q represents real Orthogonal Symmetric unitary matrice, Q ^-1Expression Q's is contrary, and Λ represents diagonal matrix.Then, make z=Q ^T(x-μ) obtains z ^TΛ z=1, wherein, ^TThe transposition of representing matrix, z represent the circle coordinates after the elliptic coordinates x normalization, seek transformation equation z=Q ^T(x-μ) normalizes to circle coordinates z with elliptic coordinates x;

(2) outside zero padding of the border circular areas after normalization is filled to external square area;

(3) utilize Algorithms of Discrete Wavelet Transform to resulting external square area advanced every trade conversion carry out again rank transformation, obtain the wavelet conversion coefficient matrix;

(4) adaptive wavelet coefficient modifying, method is as follows: the vertical low-frequency band LH of horizontal high frequency (x, y) and the vertical high frequency band HL of the horizontal low frequency (x of note wavelet conversion coefficient matrix, y) difference is D (x, y), and establishing the robustness threshold value is α, when watermark bit w (x, y) was embedded into, the coefficient difference must satisfy lower relation: D (x, y)＞=and α, if w (x, y)=1, D (x, y)＜-a, if w (x, y)=0;

(5) amended wavelet coefficient matrix is carried out reverse wavelet transform, obtain amended square features zone;

(6) get inscribed circle and oppositely normalization at the square features intra-zone, obtain the oval feature zone after watermarked;

(7) characteristic area is replaced: remember the characteristic area R after watermarked ₂With primitive character zone R ₁Difference be R _d, with R _dBe added on the original video frame.

The watermark extracting step can for:

(1) to being embedded with the frame of video of watermark, extracts the maximum stable extremal region MSER of the frame of video that is embedded with watermark according to the method for the first step;

(2) carry out the elliptic region screening according to the method for second step, obtain to comprise the elliptic region of watermark;

(3) to comprising the elliptic region of watermark, carry out successively (1), (2) and (3) in the 3rd step;

(4) watermark bit blind Detecting: note w ' (x, y) detects watermark bit at (x, y), and then the mode of watermark bit blind Detecting is as follows:

$w^{\&prime;}(x,y)=\left\{\begin{array}{cc}1,\mathrm{if}& D^{\&prime;}(x,y)\&GreaterEqual;0\\ 0,\mathrm{if}& D^{\&prime;}(x,y)<0\end{array}\right.,$ Wherein, D ' (x, y)=HL ' (x, y)-LH ' (x, y), HL ' (x, y) and LH ' (x, y) are illustrated in the wavelet coefficient that position (x, y) is located;

(5) calculate recovery rate, if the rate of getting is less than certain threshold tau _Nc, be considered as in this zone not with watermarked informationly, get τ among the present invention _Nc=0.68;

(6) final watermark w ^*(x, y) ballot obtains according to following ballot formula:

$w^{*}(x,y)=\left\{\begin{array}{cc}1,\mathrm{if}& {\mathrm{Num}}_1(x,y)\&GreaterEqual;{\mathrm{Num}}_0(x,y)\\ 0,\mathrm{if}& {\mathrm{Num}}_0(x,y)<{\mathrm{Num}}_1(x,y)\end{array}\right.$

Wherein, w ^*(x, y) is illustrated in the watermark bit voting results of locating position (x, y), Num ₁(x, y) and Num ₀(x, y) is illustrated respectively in the number of " 1 " that position (x, y) detects and the number of " 0 ".

Above-mentioned robustness is put threshold alpha can be 0.09.

The present invention has following technique effect:

1. the present invention is based on the MSER Affinely invariant region and carry out the watermark embedding, synchronization watermarking embeds and leaching process, has strengthened anti-affine geometry attacking ability.

2. the present invention revises wavelet coefficient to carry out the watermark embedding by the vertical lower frequency region of horizontal high frequency (HL) and the vertical high-frequency domain of horizontal low frequency (LH) at DWT, can effectively reduce the modification to original video, even suffering also to keep fabulous visual effect in the malice geometric attack situation.

3. the algorithm realized of the present invention is close to linear complexity, its high efficiency and Real-time ensuring technology applying in the actual production life.

Description of drawings

The watermark that Fig. 1 the present invention proposes embeds flow chart.

R*r square region after Fig. 2 normalization.

Square region after Fig. 3 fills.

The MSER zone that Fig. 4 is original.

The ellipse that Fig. 5 simulates.

MSER zone after Fig. 6 area filters.

The regional minimum spanning tree that Fig. 7 constructs.

Fig. 8 MSER region clustering.

The final selection result of Fig. 9.

Video sequence signal to noise ratio curve chart after Figure 10 is watermarked.

Figure 11 dimensional variation is attacked lower robust analysis curve chart.

Robust analysis curve chart under Figure 12 rotation attack.

Embodiment

The present invention is described in detail below in conjunction with drawings and Examples.

Below the present invention is described in detail:

Step 1: generating pictures local affine invariant covariant zone

MSER is communicated with by analyzing, the layering set of weighted graph detects, and mainly divides three steps to carry out: to arrange pixel intensity; Merge and be communicated with assembly; Detect extremal region and confirm maximum stable extremal region.Concrete steps are as follows:

1. detection maximum stable extremal region.At first use the BINSORT algorithm according to pixels to be worth and arrange the picture pixel.After the ordering, all pixels have carried out mark in the picture, are communicated with assembly by uniting-search algorithm and grow and merges, and have preserved the area of each connection assembly.In extremal region, the maximum stable zone refers to that those are under assign thresholds, the device region relative area changes the device region that the function that changes as relative threshold is obtained local minimum, also namely, MSER refers to can keep stable picture region in larger threshold range in picture binaryzation process.

2. fitted ellipse is regional.Original MSER has irregularly shaped and very difficult description, so must make up the region shape descriptor.Usually extremal region can fit to ellipse according to the actual area shape.At first calculate average u and the covariance ∑ of each maximum stable extremal region R:

$u=\frac{1}{\left|R\right|}\underset{x\&Element;R}{\mathrm{\&Sigma;}}x$

$\mathrm{\&Sigma;}=\frac{1}{\left|R\right|}\underset{x\&Element;R}{\mathrm{\&Sigma;}}(x-u){(x-u)}^T.$

Wherein x=(x, y) represents the position of pixel, and R represents area pixel point location sets, | R| represents the number of pixel among the MSER.Oval local affine invariant covariant characteristic area is by formula (x-u) ^T∑ ^-1(x-u)=1 determine.

Step 2: maximum stable extremal region filters.

The zone that the MSER detector extracts is the whole picture of the more overlapping covering of quantity (seeing accompanying drawing 4) usually.After the over-fitting step, many elliptic regions have overlapping (seeing accompanying drawing 5).This causes algorithm can't accurately extract watermark so that our watermark embedding method is repeatedly watermarked in the overlapping region.In order to eliminate this ambiguity, the overlapping region need to be filtered.

Based on the MSER provincial characteristics, the excessive too small elliptic region of area at first is filtered.As pre-treatment step, the elliptic region area that we keep satisfies τ ₁＜| R _i|＜τ ₂, wherein | R _i| represent the area of i MSER elliptic region, τ ₁And τ ₂The lower bound and the upper bound that represent respectively area.τ ₁And τ ₂Usually obtained by experiment experience, we get τ in the present invention ₁=2000, τ ₂=5000.(seeing accompanying drawing 6)

In order to filter all overlapping regions, we design of graphics category of model MSER is regional.In graph model, each elliptic region is regarded as a summit V _iBe connected to each other between every pair of summit, (V, E) consists of undirected complete graph G.Every limit E _jWeight w _jEqual the distance between the corresponding vertex.Our target is that undirected complete graph cutting is generated unconnected graph.Each connected subgraph represents a cluster, and each summit only belongs to a specific cluster.In order to obtain stable zero lap MSER zone, we remove the weight limit based on minimum spanning tree (MST) algorithm.Minimum spanning tree is the spanning tree (seeing accompanying drawing 7) with minimal weight summation.The MST algorithm tends to delete the weight larger skirt, and two of apart from each other MSER zones can be divided in the different clusters like this.By threshold tau is set, we further delete the larger limit of weight, make every weight limit satisfy w _i＜w _τSo after further filtering, we obtain a dark woods of spanning tree or without connected graph (seeing accompanying drawing 8).Like this, we obtain the cluster set of a quantification, a plurality of MSER of each set-inclusion zone.At last, final select (the seeing accompanying drawing 9) of the MSER of Retention area maximum zone conduct in each cluster.

The MSER filter algorithm divided for four steps carried out:

1. filter out the excessive or too small MSER zone of area;

2. use the MST algorithm at the complete graph that makes up and generate minimum spanning tree;

3. candidate's MSER set is classified to obtain the cluster of quantification;

4. the MSER zone that keeps area maximum in each cluster is selected as final.

Step 3: watermark mechanism.

1. watermark carrier channel selecting.

The YUV color space typically is used for coloured image pipeline, the Human Perception effect of energy efficient coding colour picture or video.Y represents brightness component, and U and V are color component.The sampling ratio Y of U and V is low, and this is called " color down-sampling " usually, and purpose is in order to improve compression efficiency.In order to keep chromaticity, we only add watermark at the Y passage.

2. the constant normalization of rotation-yardstick.

In order to obtain yardstick and affine consistency, the MSER elliptic region of each selection need to be normalized into circular block.The oval Cheng Yuan of normalization is equivalent to and seeks a transformation equation.At first, we describe the ellipse of match by average μ and covariance ∑, satisfy (x-μ) ^T∑ ^-1(x-μ)=, wherein, x=(x, y).Oval for normalization, ∑ is pressed singular value decomposition:

∑ ^-1＝QΛQ ^-1

Wherein, Q represents real Orthogonal Symmetric unitary matrice, and Λ represents diagonal matrix.

Then, make z=Q ^T(x-μ) obtains z ^TΛ z=1, wherein z represents the coordinate that obtains from x normalization.Find transformation equation z=Q ^T(x-μ) afterwards, can calculate and derive thus in the normalization space.

In order to obtain the rotation attack consistency, we need to compose a fixed-direction to each normalization zone.To each normalization zone compute gradient and make up direction histogram in circular window.Histogrammic peak point is exactly the principal direction of characteristic area.

3. frequency domain.

The present invention is based on discrete wavelet transformer and brings the interpolation watermark to HL and the LH band of wavelet field.The main advantage that wavelet transform is compared Fourier transform is its simultaneously capture frequency and spatial positional information.In the present invention, the DWT algorithm by first line translation again the mode of rank transformation be applied on the normalized circular block.Most concentration of energy of wavelet transformation are at the LL subband and only account for the overall small part of wavelet coefficient.The reason that watermark is embedded into small echo HL and LH subband has two:

1) the LL subband comprises the most energy in the signal.Therefore, have strenuous exercise in frame of video, the watermark of embedding can't correctly be extracted after under attack.

2) the HH subband comprises the detailed information of video, in the watermarked easy destruction video visual effect of this subband.

4. watermark embeds.

Watermark embed process is from decoded video obtains original frame sequence.For each frame, we utilize MSER detector and filter algorithm to extract and select the MSER characteristic area.Watermark is repeated to be embedded in these MSER zones.Detailed watermark embed process is as follows:

1) use the MSER detector to extract local affine invariant covariant zone from every frame.An independent frame comprises many MSER zone and overlaps each other.The overlapping region is used and is eliminated based on the feature filter algorithm of minimum spanning tree.

2) in order to obtain affine and the yardstick consistency, we become the unit border circular areas at each elliptic region to be embedded of normalization.Regional compute gradient direction after each normalization and principal direction one are obtained rotational invariance.For obtaining can be as the square region of DWT input, the zero padding operation is carried out on the border of our team's border circular areas.

3) each square region obtains the DWT coefficient after wavelet transform.Watermark is embedded in HL and the LH subband.In the present invention, Wavelet Coefficient Blocks is of a size of 2r*2r, so watermark is scaled to r*r to keep onesize with LH or HL subband.We get r and equal 32 in the experiment.After the watermark adjusted size is complete, carry out the watermark embedding by revising HL and LH sub-band coefficients.

4) last, based on wavelet transform, we are embedded into the watermark self adaptation in the wavelet coefficient.Here, we remember that the difference of LH (x, y) and HL (x, y) is:

D(x，y)＝HL(x，y)-LH(x，y)

Select α as the robustness threshold value, when watermark bit w (x, y) was embedded into, the coefficient difference must satisfy following relation:

D(x，y)＞＝α，if?w(x，y)＝1，D(x，y)＜-a，if?w(x，y)＝0

Larger threshold alpha can improve the robustness of algorithm but also reduce simultaneously the perceived effect of video.Balance robustness and perceived effect, we arrange threshold alpha is 0.09.

5. watermark extracting.

Similar watermark embeds, and the first step of watermark extracting is to analyze every content frame to extract local affine invariant covariant zone.Then, the use characteristic filter algorithm is selected Non-overlapping Domain.Last watermark is obtained from these zones.In order to obtain affine and yardstick consistency, it is necessary that each elliptic region that is embedded with watermark is carried out normalization, and the zone after each normalization is filled to the square (seeing accompanying drawing 2,3) of 2r*2r.

We are from LH and the HL subband extraction watermark information of wavelet coefficient.The mode of watermark bit blind Detecting is as follows:

$w^{\&prime;}(x,y)=\left\{\begin{array}{c}1,\mathrm{if}{D}^{\&prime;}(x,y)\&GreaterEqual;0\\ 0,\mathrm{if}{D}^{\&prime;}(x,y)<0\end{array}\right.$

Wherein, D ' (x, y)=HL ' (x, y)-LH ' (x, y), HL ' (x, y) and LH ' (x, y) are illustrated in the wavelet coefficient that position (x, y) is located.

After detecting watermark w ', we calculate the normalization relating value z between w and the w ' _Nc:

$z_{\mathrm{nc}}=\frac{(w\&CenterDot;w^{\&prime;})}{\left|w\right|}$

Wherein, | w| represents the length of w.If z _NcLess than threshold tau _Nc, we think not with watermarked information in this zone.

In fact, we have embedded watermark in a plurality of invariant regions, even like this so that also can detect watermark in the situation that video deforms.Final watermark result is by choosing in a vote.Note Num ₁(x, y) and Num ₀(x, y) is illustrated respectively in the number of " 1 " that position (x, y) detects and the number of " 0 ".Then, watermark result is following determines:

$w^{*}(x,y)=\left\{\begin{array}{c}1,\mathrm{if}{\mathrm{Num}}_1(x,y)\&GreaterEqual;{\mathrm{Num}}_0(x,y)\\ 0,\mathrm{if}{\mathrm{Num}}_0(x,y)<{\mathrm{Num}}_1(x,y)\end{array}\right.$

The present invention tests the Foreman video the most frequently used in the watermark field, and it mainly weighs the visual quality of watermark video and the robustness of watermark extracting under various attack.Y-PSNR (PSNR) is used to measure the visual quality of frame of video, and recovery rate (Normalized correlation value) is used to measure the robustness of watermark mechanism.

Figure 10 is the visual effect that Y-PSNR (PSNR) is used for weighing video.From the experimental results, there are not vision difference in watermark video and original video.Figure 10 has showed the Y-PSNR value of watermark video frame, and the average peak signal to noise ratio of watermark video frame is 43.763dB.

Figure 11 attacks for yardstick, and watermark video is carried out dimensional variation between 0.6 to 2.0, and experimental result is illustrated among Figure 11, and attack has robustness to yardstick can to see the present invention.

Figure 12 rotation attack is rotated the watermark video frame between 180 degree at 0 degree, is lower than 100 when spending when the anglec of rotation as seen from Figure 12, and the watermark extracting rate is greater than 90%, and along with the increase of the anglec of rotation, recovery rate is rapidly decline.

Table 1 is that method for anti-counterfeit multiclass of the present invention is attacked recovery rate contrast situation lower and list of references [16], can find out that the present invention has surmounted [16] under the various attack type.

Table 1 multiclass is attacked recovery rate contrast lower and list of references [16]

Claims (3)

1. the video watermark method for anti-counterfeit of a resist geometric attacks comprises the following steps:

The first step: the maximum stable extremal region MSER that extracts frame of video:

(1) use the BINSORT sort algorithm that pixel in the frame of video according to pixels is worth size and arrange, and set up location index for the pixel after the ordering, each pixel after the ordering is considered as being communicated with assembly;

(2) use and to unite-to search algorithm and carry out merging growth to being communicated with assembly, record the area that each is communicated with assembly;

(3) select MSER.Under assign thresholds, from the connection assembly that (2) obtain, pick out the assembly area change and obtain the connection assembly of local minimum with changes of threshold, obtain maximum stable extremal region.

(4) for each maximum stable extremal region, computation of mean values u and covariance ∑, recycling elliptic region equation (x-u) ^T∑ ^-1(x-u)=1, fit to elliptic region according to the actual area shape;

Second step: the elliptic region that simulates is screened, eliminate region overlapping, choose the elliptic region that embeds for final watermark;

(1) the excessive too small elliptic region of filter area, the elliptic region area of reservation satisfies τ ₁＜| R _i|＜τ ₂, wherein, | R _i| represent the area of i elliptic region, τ ₁And τ ₂The lower bound and the upper bound that represent respectively size.

(2) design of graphics model, in graph model, the corresponding summit V of each elliptic region i _iThe summit is to V _i, V _jBetween by limit E _IjConnect; Limit E _IjWeight w _IjEqual the summit to V _i, V _jBetween distance, (V, E) consists of undirected complete graph G;

(3) use minimal spanning tree algorithm to obtain minimum spanning tree T at undirected complete graph G;

(4) judge according to elliptic equation whether every pair of characteristic area is overlapping among the T, if overlapping, then these two elliptic regions put identical cluster under, otherwise, put different clusters under;

(5) pick out the elliptic region of area maximum in each cluster, these elliptic regions will embed for final watermark;

The 3rd step: by revising the vertical low-frequency band of horizontal high frequency (HL) and the vertical high frequency band of horizontal low frequency (LH) of wavelet conversion coefficient, the characteristic area that selects is carried out watermark embed, method is as follows:

(1) elliptic region normalization: at first, to the contrary ∑ of the covariance of the elliptic region picked out ^-1Carry out singular value decomposition, ∑ ^-1=Q Λ Q ^-1, wherein, Q represents real Orthogonal Symmetric unitary matrice, Q ^-1Expression Q's is contrary, and Λ represents diagonal matrix.Then, make z=Q ^T(x-μ) obtains z ^TAz=1, wherein, ^TThe transposition of representing matrix, z represent the circle coordinates after the elliptic coordinates x normalization, seek transformation equation z=Q ^T(x-μ) normalizes to circle coordinates z with elliptic coordinates x;

(2) outside zero padding of the border circular areas after normalization is filled to external square area;

(3) utilize Algorithms of Discrete Wavelet Transform to resulting external square area advanced every trade conversion carry out again rank transformation, obtain the wavelet conversion coefficient matrix;

(4) adaptive wavelet coefficient modifying, method is as follows: the vertical low-frequency band LH of horizontal high frequency (x, y) and the vertical high frequency band HL of the horizontal low frequency (x of note wavelet conversion coefficient matrix, y) difference is D (x, y), and establishing the robustness threshold value is α, when watermark bit w (x, y) was embedded into, the coefficient difference must satisfy lower relation: D (x, y)＞=and α, if w (x, y)=1, D (x, y)＜-a, if w (x, y)=0;

(5) amended wavelet coefficient matrix is carried out reverse wavelet transform, obtain amended square features zone;

(6) get inscribed circle and oppositely normalization at the square features intra-zone, obtain the oval feature zone after watermarked;

(7) characteristic area is replaced: remember the characteristic area R after watermarked ₂With primitive character zone R ₁Difference be R _d, with R _dBe added on the original video frame.

2. the video watermark method for anti-counterfeit of resist geometric attacks according to claim 1 is characterized in that, described method for anti-counterfeit also comprises following watermark extracting step:

(1) to being embedded with the frame of video of watermark, extracts the maximum stable extremal region MSER of the frame of video that is embedded with watermark according to the method for the first step;

(2) carry out the elliptic region screening according to the method for second step, obtain to comprise the elliptic region of watermark;

(3) to comprising the elliptic region of watermark, carry out successively (1), (2) and (3) in the 3rd step;

(4) watermark bit blind Detecting: note w ' (x, y) detects watermark bit at (x, y), and then the mode of watermark bit blind Detecting is as follows:

$w^{\&prime;}(x,y)=\left\{\begin{array}{cc}1,\mathrm{if}& D^{\&prime;}(x,y)\&GreaterEqual;0\\ 0,\mathrm{if}& D^{\&prime;}(x,y)<0\end{array}\right.,$ Wherein, D ' (x, y)=HL ' (x, y)-LH ' (x, y), HL ' (x, y) and LH ' (x, y) are illustrated in the wavelet coefficient that position (x, y) is located;

(5) calculate recovery rate, if the rate of getting is less than certain threshold tau _Nc, be considered as in this zone not with watermarked informationly, get τ among the present invention _Nc=0.68;

(6) final watermark w ^*(x, y) ballot obtains according to following ballot formula:

$w^{*}(x,y)=\left\{\begin{array}{cc}1,\mathrm{if}& {\mathrm{Num}}_1(x,y)\&GreaterEqual;{\mathrm{Num}}_0(x,y)\\ 0,\mathrm{if}& {\mathrm{Num}}_0(x,y)<{\mathrm{Num}}_1(x,y)\end{array}\right.$

Wherein, w ^*(x, y) is illustrated in the watermark bit voting results of locating position (x, y), Num ₁(x, y) and Num ₀(x, y) is illustrated respectively in the number of " 1 " that position (x, y) detects and the number of " 0 ".

3. the video watermark method for anti-counterfeit of resist geometric attacks according to claim 1 is characterized in that, it is 0.09 that robustness is put threshold alpha.

Images