CN1321393C - Watermark method using geometry calibrated anti-geometry conversion image - Google Patents

Watermark method using geometry calibrated anti-geometry conversion image Download PDF

Info

Publication number
CN1321393C
CN1321393C CNB021496145A CN02149614A CN1321393C CN 1321393 C CN1321393 C CN 1321393C CN B021496145 A CNB021496145 A CN B021496145A CN 02149614 A CN02149614 A CN 02149614A CN 1321393 C CN1321393 C CN 1321393C
Authority
CN
China
Prior art keywords
image
watermark
embedded
template
dwt
Prior art date
Application number
CNB021496145A
Other languages
Chinese (zh)
Other versions
CN1414778A (en
Inventor
康显桂
黄继武
Original Assignee
中山大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中山大学 filed Critical 中山大学
Priority to CNB021496145A priority Critical patent/CN1321393C/en
Publication of CN1414778A publication Critical patent/CN1414778A/en
Application granted granted Critical
Publication of CN1321393C publication Critical patent/CN1321393C/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T1/00General purpose image data processing
    • G06T1/0021Image watermarking
    • G06T1/005Robust watermarking, e.g. average attack or collusion attack resistant
    • G06T1/0064Geometric transfor invariant watermarking, e.g. affine transform invariant
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2201/00General purpose image data processing
    • G06T2201/005Image watermarking
    • G06T2201/0052Embedding of the watermark in the frequency domain

Abstract

本发明涉及一种多媒体信号处理领域,是一种图象几何校准和保护数字图象的方法。 The present invention relates to a multimedia signal processing, an image is geometrically calibrating the image and protect digital method. 经过扩频调制和交织后嵌入于图象DWT域中的信息水印由嵌入于图象DFT域的匹配模板和嵌入于图象DWT域的训练序列来实现重同步检测。 After interleaving and spread spectrum modulation on the watermark information embedded in an image embedded in the DWT domain by a DFT domain matching template image and an image embedded in the DWT domain training sequence to achieve resynchronization detected. 在水印图象同时经过JPEG压缩和几何变形的情况下,仍然可实现有意义信息的无差错检测。 In the case of the watermark image at the same time through the JPEG compression and geometric distortion, error detection can still be achieved without meaningful information. 本发明可使通过网络上传播的数字图象或视频数据获得保护。 The invention can be protected by the digital image or video data propagated on the network. 本发明中提出的图象几何校准方法还可用于其他需要图象同步的场合,如卫星成象、交互式数字地图、数字水印等。 Image geometric calibration method proposed in the present invention may also be used in other applications requiring synchronous image, such as satellite imaging, interactive digital map, a digital watermark.

Description

采用图象几何校准和保护数字图象的方法 Calibration and geometry of the image using digital image protection method

技术领域 FIELD

本发明涉及一种多媒体信号处理领域,是一种图象几何校准和保护数字图象的方法。 The present invention relates to a multimedia signal processing, an image is geometrically calibrating the image and protect digital method.

背景技术 Background technique

目前,水印对抗几何攻击有非盲检测和盲检测的方法。 Currently, there are ways geometric watermark against non-blind and blind detection to detect attacks. 一般来说,盲检测水印方法的稳健性较差。 In general, blind detection robustness of the watermarking method is poor. 但由于不需要原始图象,它的应用范围更广,也更具有挑战性。 But because without the original image, its wider range of applications, but also more challenging. 它又可分为两类:第一类方法是将水印嵌入于具有几何不变性的图象特征域,如Fourier-Mellin域,使得几何变形不影响水印信息的提取。 It can be divided into two categories: the watermark is embedded in the image field of characteristic geometric invariant, such as Fourier-Mellin domain, such deformation does not affect the geometry of the extracted watermark information. 但这类方法都只能抵抗RST,在实现上也存在困难,例如在将图象DFT变换的幅度谱做LPM(Log Polar Mapping)与ILPM(Inverse LogPolar Mapping)时,由于插值误差会导致图象质量的严重下降。 However, these methods can only resist the RST, there are difficulties in the realization, for example, the amplitude of the DFT spectrum image do when LPM (Log Polar Mapping) and ILPM (Inverse LogPolar Mapping), since interpolation errors cause image a serious decline in quality. 第二类是在几何变形导致水印检测失步的情况下,在水印检测前设法先进行几何校正以实现水印检测的重同步。 The second category is the geometric deformation results in the case of step-out of the watermark detection, the watermark detection prior to trying to achieve a geometric correction resynchronization watermark detection. 这需要在图象中除隐藏携带用户信息的水印(信息水印)外还嵌入一个几何校正模型水印(模板水印)。 This requires, in addition to the image carries hidden information outside user watermark (watermark information) embedded in a geometric calibration model further watermark (watermark template). 到目前为止,对抗几何攻击还存在许多问题,包括隐藏数据量少、水印的不可见性较差、抗JPEG压缩和抗仿射变换能力较弱、不能同时抵抗JPEG压缩和几何变换的组合攻击等。 So far, the fight against geometric attacks are still many problems, including hidden data less, the watermark is invisible poor, anti-JPEG compression and anti-affine transformation capability is weak, can not resist the combination of JPEG compression attack and geometric transformations, etc. . 另外,许多方法采用在DFT域中嵌入信息水印,而DFT(Discrete FourierTransform)与DWT(Discrete Wavelet Transform)相比,存在自身的弱点,难以成为主流水印方法的基础。 Further, watermark information is embedded using a number of methods in the DFT domain and DFT (Discrete FourierTransform) and the DWT (Discrete Wavelet Transform) as compared to the presence of their own weaknesses, it is difficult to form the basis of the main watermarking method.

发明内容 SUMMARY

本发明的目的是提出一种能同时抵抗一般的信号处理如JPEG压缩和几何变换的稳健的隐形图象盲检测水印方法,并且隐藏的信息量(静荷)要较大(可达200比特以上),水印的不可见性较好。 Object of the present invention is to provide an apparatus capable of blind detection while resisting a latent image watermarking method generally JPEG compression signal processing as geometric transformation and robust, and hidden information (static load) to a large (up to 200 bits ), invisible watermark is good.

本发明方法图象几何校准和保护数字图象的方法,首先将信息水印经扩频调制和交织后与一个训练序列一起嵌入到图象DWT域中,再将一个匹配模板嵌入到图象DFT域,最后由嵌入于图象DFT域的匹配模板和嵌入于图象DWT域的训练序列来实现扩频调制和交织后嵌入于图象DWT域中的信息水印的重同步检测,具体做法是:1)将待嵌入的有意义信息b(L bits)首先用一个由密钥产生的伪随机二进制PN序列进行扩频编码调制,这样可得到待嵌入的二进制水印数据W,然后再进行交织;将由密钥产生的伪随机码组成的训练序列T采用量化调制的方法直接嵌入于图象DWT变换得到的LL3子带的中心行和中心列子带系数或其他子带系数中,而在其余子带系数中采用量化调制的方法嵌入交织后的二进制水印数据W;通过2-D IDWT得到嵌入DWT域水印的图象f′(x,y);2)将f′(x,y)进行DFT变换, The method of the present invention is an image geometry and the calibration method of protecting digital image, by the first watermark information is embedded in the interleaving and spread spectrum modulated with the training sequence with a DWT to the image domain, and then embedded in a matching template image DFT domain Finally, the image embedded in the DFT domain matching template image and embedded in the DWT domain training sequence spread spectrum modulation and interleaving to achieve a weight embedded in the DWT domain image synchronization detection watermark information, which would be: 1 ) to be embedded meaningful information b (L bits) first with a pseudo-random binary sequence generated by the PN spreading code modulation key, thus obtaining binary watermark data to be embedded in W, then interleaving; adhesion by T training sequence consisting of a pseudo-random code key generated using the modulation quantization method directly embedded in the central row and central column sub band coefficients other subband coefficients or subband LL3 DWT transform of the image, and in the remaining sub-band coefficients using quantization index modulation of the interleaving method of embedding binary watermark data W is; obtained by 2-D IDWT DWT domain watermark embedded image f '(x, y); 2) the f' (x, y) for the DFT, DFT变换的幅度谱系数中我们采用加法嵌入的方法嵌入一个匹配模板,模板点位置由一个密钥控制产生;3)检测过程是嵌入过程的逆过程,先检测匹配模板,并与原始的匹配模板比较得到图象所经受的仿射变换的变换矩阵并作逆变换恢复几何形状后,再根据训练序列作图象平移校准后提取水印信息。 The amplitude of the DFT spectral coefficients we use the addition embedding method of embedding a matching template, the template generating a key control point position by a; 3) the detection process is the reverse process of the embedding process, the first detection template matching, template matching with the original comparison transformation matrix to obtain an image subjected to affine transformation and inverse transformation as geometries recovery, then in accordance with training sequences extracted watermark image after calibration translated.

为了检测图象所经受仿射变换的变换矩阵并作逆变换恢复其原始几何形状,可在图象DFT变换的幅度谱系数中嵌入由局部极大点构成的一个匹配模板,局部极大点位置可由一个密钥控制产生;同时在图象DWT域嵌入一个训练序列用于平移校准。 In order to detect the image transformation matrix is ​​subjected to affine transformation and inverse transformation for return to its original geometry, may be embedded in a matching template consisting of a local maximum in the log magnitude spectrum in the DFT image, the position of the local maximum point by generating a control key; while a training sequence embedded in the image and DWT for translating calibration.

经过扩频调制和交织后嵌入于图象DWT域中的信息水印由嵌入于图象DFT域的匹配模板和嵌入于图象DWT域的训练序列来实现重同步检测;分为水印的嵌入和水印的检测两大步骤。 After interleaving and spread spectrum modulation on the watermark information embedded in an image embedded in the DWT domain by a DFT domain matching template image and an image embedded in the DWT domain training sequences to achieve a weight synchronous detection; watermark embedding and watermark into the detection of two steps.

1、水印的嵌入:本发明将信息水印嵌入在图象DWT域的低频子带系数中,而模板水印嵌入在图象DFT变换后的幅度谱中频系数中,这样两部分水印互相不干扰,又能获得更好的稳健性。 1, the embedded watermark: The present invention watermark information is embedded in the frequency subband DWT domain image coefficients, and the template watermark embedding amplitude spectrum after DFT of an image intermediate frequency coefficients, so that the two parts do not interfere with each other watermark, and get more robustness. 本发明提出的图象水印嵌入方法方框图如图1所示。 Image watermark embedding method of the present invention proposed a block diagram shown in Fig.

2:水印的检测:不需要原始图象的辅助,本发明方法可以将隐藏的数据从可能同时遭到几何攻击和JPEG压缩的水印图象中检测得到。 2: detection of the watermark: the auxiliary without the original image, the method of the present invention may be hidden data may also suffer from the geometrical attacks and JPEG compression watermark detection image obtained. 检测过程如下:a)在水印检测时,首先要应用训练序列检测水印是否同步。 The detection process is as follows: a) in the watermark detection, we must first apply whether the training sequence to detect the watermark synchronization. 若不同步,则必须先经过重同步得到同步图象g*(x,y)。 If synchronization, re-synchronization must first be synchronized image g * (x, y). 重同步包括DFT域摸板水印检测、逆仿射变换、应用训练序列平移同步。 DFT domain formwork comprising resynchronization watermark detection, the inverse affine transformation, application translation synchronization training sequence. 若同步,直接做一步。 If synchronized directly to do one step.

b)对同步图象g*(x,y)作DWT域水印检测,得到了实际隐藏的数据。 b) synchronization image g * (x, y) for the DWT domain watermark detection to obtain the actual hidden data.

水印嵌入过程主要有DWT域水印(包括信息水印、训练序列)的预处理、DWT域水印的嵌入和DFT域摸板水印的嵌入三部分。 The watermark embedding process mainly DWT domain watermark pretreatment (including the watermark information, the training sequence), DWT domain watermark embedding and watermark embedding DFT domain formwork three parts.

1)DWT域水印(信息水印、训练序列)的预处理:直接序列扩频编码、交织。 1) Pretreatment DWT domain watermark (watermark information, the training sequence): direct sequence spread spectrum coding, interleaving.

本发明将一些通信理论中常用的技术(如直接序列扩频调制和交织),引入图象水印方法中以增强隐藏信息的稳健性和秘密性。 The present invention is commonly used in some communication theory techniques (e.g., direct sequence spread spectrum modulation and interleaving), image watermarking method is introduced to increase the robustness of the secret and hidden information.

假设原始图象大小是512×512。 Assuming that the original image size is 512 × 512. 应用长度为N1的PN码序列m={mj;j=1,...,N1}对要嵌入的信息b{bi;i=1,...,L}(其中bj∈{0,1})进行扩频编码调制。 Application of a length of N1 is the PN code sequence m = {mj; j = 1, ..., N1} of the information to be embedded b {bi; i = 1, ..., L} (wherein bj∈ {0,1 }) spreading code modulation. “1”调制为m(双极性序列,mj∈{-1,1})的正相序列,即{+1×mi;j=1,...,N1},“0”调制为m的反相序列,即{-1×mj;j=1,...,N1}。 "1" is a normal phase modulation sequence m (bipolar sequence, mj∈ {-1,1}), i.e., {+ 1 × mi; j = 1, ..., N1}, "0" is modulated to m reverse phase sequence, i.e., {-1 × mj; j = 1, ..., N1}. 15位的PN码序列由一个密钥通过PN码序列发生器产生。 15-bit PN code sequence generated by a PN code sequence generator by a key. 这样可得到待嵌入的二进制水印数据W:bi→DSSScodingWi{wij;wij∈{-1,+1}1≤j<N1,1≤i<L}]]>训练序列是信息水印能否实现平移同步的关键,为尽量使其少受到图象裁剪的影响,应将它嵌入于图象中需要重点保护部分对应的低频子带部位或低频子带的中心行和中心列。 Thus obtaining be embedded binary watermark data W: bi & RightArrow; DSSScodingWi {wij; wij & Element; {- 1, + 1} 1 & le; j & lt; N1,1 & le; i & lt; L}]]> training sequence is a watermark information can be achieved translation the key synchronization, it is possible cropped image little affected, it should be embedded in the image need protection center line and the central portion or the lower sub-band portion corresponding to the lower sub-band column. 如图2所示嵌入于低频子带的32行和32列。 2, line 32 embedded in the lower sub-band and 32. 而在低频子带的其余部分中嵌入经过交织(采用二维交织技术或其他交织技术)后的二进制水印数据W。 While the remaining portion is embedded in the interleaved in the lower sub-band (or other two-dimensional interleaving interleaving technique) binary watermark data W.

在一个64×64二维矩阵的32行和32列位置上存放127位训练序列,其余位置顺序存放交织后二进制水印数据W,将得到的二维矩阵按行扫描变成一个一维数组,记为X。 Stored in a 64 × 64 32 rows and 32 two-dimensional matrix position 127 the training sequence, the rest position of the interleaved sequence stored binary watermark data W, obtained by the two-dimensional matrix into a one-dimensional line scan array, denoted It is X.

2)DWT域水印(信息水印、训练序列)的嵌入与检测方法DWT域水印的嵌入原始图象f(x,y)进行三级DWT分解,把低频子带LL3系数按行扫描变成一维数组,记为C。 2) DWT domain watermark (watermark information, the training sequence) is embedded in the DWT domain watermark detection method of embedding the original image f (x, y) for three DWT decomposition, the low frequency subband LL3 coefficients into one-dimensional scanning line by array, denoted by C. 按公式(1),我们把二进制数据X加到低频系数C上,得到新的低频系数c′: According to formula (1), we applied the binary data X low frequency coefficients C, giving the new low-frequency coefficients c ': 其中0≤i<4096,C(i)、C′(i)、xi分别为C、C′、X的第i个元素。 Where 0≤i <4096, C (i), C '(i), xi are C, C', X i-th element. α表示嵌入强度,在满足不可见性的前提下,尽可能选择最大的整数值。 α represents the embedding strength, the premise of meeting invisibility selecting the maximum integer value as possible. 将嵌入水印后的小波系数进行IDWT得到嵌入DWT域水印的图象f′(x,y)。 The wavelet coefficients obtained IDWT watermark embedding the domain watermark embedding DWT image f '(x, y).

DWT域水印的检测 把已经同步的图象g*(x,y)DWT分解后的低频子带LL3系数按行扫描变成一维数组,记为C*。 DWT domain watermark detection image has been synchronized to the g * (x, y) DWT decomposition coefficients are low frequency subband LL3 scanning line into one-dimensional array, referred to as C *. 抽取出来的二进制数据记为X*={xi-},抽取公式如下:xi*=+1,(C*(i)moda)&GreaterEqual;a2xi*=-1,otherwise---(2)]]>其中0≤i<4096,α为嵌入强度。 Extracted binary data denoted as X * = {xi-}, following decimation equation: xi * = + 1, (C * (i) moda) & GreaterEqual; a2xi * = - 1, otherwise --- (2)]] > where 0≤i <4096, α is the embedding strength. 将抽取的二进制数据X*进行反交织(交织的逆过程)恢复嵌入的二进制数据序列W*。 The extracted binary data X * inverse interleaving (the inverse process of interleaving) to recover the embedded binary data sequence W *. 然后W*按15位比特进行分段,每段与15bits的序列m进行相关,若相关值大于0,则判决嵌入信息比特为“1”,否则判决嵌入信息比特为“0”。 W * is then segmented by 15 bits, each related sequence m 15bits, if the correlation value is greater than 0, then decision information embedding bit is "1", the embedded information or the decision bit is "0." 解扩之后就得到恢复的嵌入信息。 After despreading get embedded information recovery.

3)DFT域模板的嵌入与检测在嵌入DWT域水印后图象f′(x,y)的DFT域嵌入一个模板用作水印图象变形后的同步信息。 3) fitted with a DFT domain template detected is embedded in the DWT domain watermark image f '(x, y) is embedded in a DFT domain template as a synchronization information after the watermarked image modification.

模板的嵌入分如下四个步骤:a)将f′(x,y)(512×512)用均值填充四周扩展至1024×1024。 Embedding sub-template following four steps of: a) f '(x, y) (512 × 512) extended to 1024 × 1024 average around filling.

b)作DFT变换,取傅立叶系数幅度分量。 b) for the DFT, taken Fourier coefficient amplitude component. 在中频区域(归一化频率为0.20~0.30)嵌入28个模板点,均匀地分布在DFT域倾角为θ1和θ2的两条直线,每条线上14个点,图2是嵌入模板的示意图,图中只画出上半平面14个模板点的情况,下半平面关于原点对称也嵌入14个模板点。 In the intermediate region (normalized frequency of 0.20 to 0.30) is embedded in the template 28 points, uniformly distributed in the DFT domain inclination θ1 and θ2 of the two straight lines, each line point 14, FIG. 2 is a schematic diagram of the embedded template , shown only in FIG. 14 where the half-plane template points, the half-plane symmetric about the origin point of the template 14 is also embedded. 直线的倾角和模板点的极径由一个密钥伪随机产生。 Electrode diameter randomly generated by a pseudo inclination and a key point of the line template.

c)增大模板点处傅立叶系数的模值,使之成为局部区域(可采用半径为R的圆形窗口,如图3所示)的极大值。 c) increasing the value of the modulus of the Fourier coefficients at the point of the template, making local region (employed radius R of the circular window, as shown in FIG. 3) is maximum. 改变量以不可见为标准,一般取极大值为局部平均值加上几倍到十几倍左右的方差。 The amount of change is not visible to the standard, and generally the maximum value of the mean plus local variance times to about ten times.

d)计算傅立叶反变换(IDFT)得到最终的水印图象f″(x,y)。 d) calculating the inverse Fourier transform (IDFT) to obtain the final watermarked image f "(x, y).

图象在空间域受到的线性变换将在DFT域产生相应的线性变换,所以通过模板点位置的变换关系就可以确定图象所经历的几何形变。 The corresponding linear image generated in the DFT domain transformation in the spatial domain by a linear transformation, so that the image can be determined geometric distortion experienced by the template transformation between point position. 如果一个方形图象在空间域发生了以下的线性变换:xy&RightArrow;Bxy---(3)]]>那么相当于在DFT域做了如下的线性变换:uv&RightArrow;(B-1)Tuv---(4)]]>对于一条模板线上的模板点,经历线性变换后,它们还是同在一条过原点直线上。 If a square image in the spatial domain occurs following linear transformation: xy & RightArrow; Bxy --- (3)]]> in the DFT domain corresponds then made the following linear transformation: uv & RightArrow; (B-1) Tuv-- - (4)]]> a template for template-point line, after subjected to a linear transformation, they are still the same in a straight line through the origin. 新模板点的坐标(如极径r′)与原模板点的坐标(如极径r)存在一定的关系(如r′=Kr,K为某一常数),这可用于搜索过程的快速匹配判断。 'There is a certain relationship (e.g., the original template r coordinate point (e.g., the polar radius r) = Kr, K is a constant) coordinates of the new point templates (e.g., the polar radius r)', which can be used for fast matching search procedure judgment.

模板检测的步骤如下:a)对待测图象g(x,y)作Barlette滤波。 The step of detecting the template as follows: a) treat prediction image g (x, y) for filtering Barlette.

b)同嵌入模板时一样,将滤波后的待测图象扩展至1024×1024。 b) embedding the same template, extended to test the filtered image 1024 × 1024.

c)作DFT变换。 c) for the DFT. 以一个半径为R′(R′<R,R为嵌入时的窗口半径)的圆形窗口(作为局部区域)在傅立叶系数幅度矩阵的上半平面中搜索,提取所有局部极大值点。 To a radius R '(R' <R, R is the radius of the window when embedding) the circular window (as a local region) searched in the upper half plane of the Fourier coefficients of the amplitude matrix, extract all the local maxima. 把DFT系数幅度矩阵上半平面以原点为顶点划分为Nb(Nb=360或180或其他值)个扇形区域,每个扇形的顶角均为0.5°或1°。 The amplitude matrix of DFT coefficients in the upper half plane origin vertex divided into Nb (Nb = 360 or 180, or other value) a fan-shaped area, the apex angle of each sector are 0.5 ° or 1 °. 再按角度将所有局部极大值点分别归入各个扇形区域。 Then all the local maxima of the angle points are included in each sector region.

d)找到与两条摸板线对应的可能的摸板点集合。 d) Locate the line corresponding to the two formwork possible Moban point set.

在每个扇形区域中,在Kmin<K<Kmax范围内搜索这样的K值:它使得此扇区中至少有Nm个局部极大值点满足|rli-KrTj′|<threshold,其中Nm为一个预先规定的数,rli是扇区i中局部极值点的极径(i=l...Nb),rTj′是原模板线j(j=1,2)上摸板点的极径,threshold>0为一阈值。 In each sector area, the K value Kmin <K <Kmax within this search range: in this sector so that it has at least two local maxima Nm satisfies | rli-KrTj '| <threshold, where Nm is a a predetermined number, the sector i RLI is the polar radius local extreme points (i = l ... Nb), rTj 'is the diameter of the original template electrode line j (j = 1,2) point formwork, threshold> 0 is a threshold value. 实验中我们取Nm=5,threshold=0.002,Kmin=0.5和Kmax=2.0(对应于空间域上的缩放参数为2~0.5)。 Experiment we take Nm = 5, threshold = 0.002, Kmin = 0.5 and Kmax = 2.0 (corresponding to a spatial-domain scaling parameters from 2 to 0.5). 如果找到这样的K值,我们就把相应的局部极值点坐标记录下来。 If you find such a K value, we put the appropriate local extreme point coordinates recorded.

e)通过上述步骤,得到可能的匹配线的集合,称为“准匹配线”,线上的局部极值点称为“准匹配点”,坐标记为(xij,yij)。 e) by the above procedure, the set of possible match line, referred to as "quasi-match line" local extreme points on a line called a "quasi-matching point", as coordinate notation (xij, yij). 图象上半平面相应的原始模板点的坐标记为(xij′,yij′),其中i∈{1,2}表示第i条模板(匹配)线,j∈{1,2,Λ,7}表示第j个模板(匹配)点。 Sitting on an image corresponding to the original template mark point half-plane (xij ', yij'), where i∈ {1,2} denotes the i-th template (matching) line, j∈ {1,2, Λ, 7 } denotes the j-th template (matching) point. 从对应于模板线1的准匹配点集中取出一个集合和对应于模板线2的准匹配点集中取出另一个集合。 From the line corresponding to the registration template matching point 1 set out a concentration corresponding to the quasi-two line template matching point another set of extracted concentrate. 根据这两个集合的点与模板点间的对应关系计算得到的一个可能的变换矩阵A。 Calculated based on the correspondence between these two points and a template set of points of possible transformation matrix A. 寻找平均误差MAE(Mean Absolute Error)最小的A。 Looking average error MAE (Mean Absolute Error) minimal A.

MAE=1nummatches||Ax11y11MMx11y11x21y21MMx21y21T-x11&prime;y11&prime;MMx11&prime;y11&prime;x21&prime;y21&prime;MMx21&prime;y21&prime;T||---(5)]]>其中模板点为(xij′,yij′)和“准匹配点”为(xij,yij),nummatches是匹配点个数,运算符‖Λ‖中是一个2行的误差矩阵。 MAE = Ax11y11MMx11y11x21y21MMx21y21T-x11 & prime 1nummatches ||; y11 & prime; MMx11 & prime; y11 & prime; x21 & prime; y21 & prime; MMx21 & prime; y21 & prime; T || --- (5)]]> wherein the template points (xij ', yij') and the "quasi-matching point "is (xij, yij), nummatches is the number of matching points, ‖Λ‖ operator error matrix is ​​a 2 line.

f)将对应于模板线1的准匹配点加上180°,重复e),由最小的MAE值确定最后的频域变换矩阵A。 f) corresponding to the quasi-line template matching point 1 plus 180 °, repeating e), to determine the final frequency domain by the transform matrix A. The minimum value of MAE 由式(3)和(4)可得空域变换矩阵B=AT。 By the formula (3) and (4) available spatial transformation matrix B = AT.

应用抽取的训练序列S与原始训练序列T的相关系数来确定图象是否同步及图象的平移同步参数。 Application of the training sequence extracted correlation coefficient S and T of the original training sequence to determine whether the image of the image synchronization and translational synchronization parameter.

在水印检测时,首先要检测水印是否同步。 When watermark detection, first to detect whether the watermark synchronization. 若不同步,则必须重同步水印后才能进行水印检测。 Without synchronization, it must re-synchronization watermark can be detected watermark. 若同步,则直接提取低频子带LL3子带隐藏数据和解码出信息。 If synchronization is extracted directly frequency subband LL3 sub-band data and the decoded information is hidden.

检测水印是否同步:将待测图象g(x,y)重定大小为512×512,然后对其进行3级DWT分解,从LL3子带的32行与32列中提取训练序列S,计算它与原始训练序列T的相关系数PT,S(0)=1127&Sigma;n=1127(TnSn),]]>看是否≥threshl。 Watermark detecting whether synchronization: the test image g (x, y) to resize 512 × 512, and then subjected to stage 3 DWT decomposition, the line 32 and the sub-band LL3 from 32 to extract the training sequence S, which is calculated PT correlation with the original training sequence T, S (0) = 1127 & Sigma; n = 1127 (TnSn),]]> to see if ≥threshl. 若是,我们认为S是真正的训练序列,并且水印是同步的,可以直接进行DWT域水印的抽取和译码。 If so, we believe that S is a real training sequence, and the watermark is synchronous, can be extracted and decoded watermark DWT domain directly. 若<threshl,则认为水印是不同步的,必须先经过重同步才能进行DWT域水印的抽取和译码。 If <threshl, believes that the watermark is not synchronized, it must go through in order to be re-synchronized to extract and decode DWT domain watermark. threshl一般可取0.56(由实验确定的值)。 threshl generally preferable 0.56 (determined by experimental values). 出现虚警即出现伪同步的概率可由计算得Pfp=12127&Sigma;k=127-e127C127k=8.59&times;10-9]]>其中e=round(127×(1-threshl)/2)。 I.e. the occurrence probability of false alarm occurs pseudo sync may be calculated Pfp = 12127 & Sigma; k = 127-e127C127k = 8.59 & times; 10-9]]> where e = round (127 × (1-threshl) / 2). Round表示四舍五入取整。 Round denotes rounding rounding.

重同步的第一步是:恢复原始几何形状。 The first step is to re-sync: restore the original geometry. 从待测图象g(x,y)中检测出嵌入的模板水印,并将之与原始的模板进行对比获得图象所经受的仿射变换矩阵B。 Detecting the test image g (x, y) of the template watermark embedded, and compared with the original template image affine transformation matrix obtained is subjected B. 获得仿射变换矩阵B后,将待测图象g(x,y)进行图象几何逆变换恢复成M×N大小的图象g′(x,y)(图5b),然后再填充0成512×512大小的图象I(x,y),被裁剪的部分以0填充,g′(x,y)在图象I(x,y)中心(图5c)。 After obtaining the affine transformation matrix B, and the test image g (x, y) for image geometric restored to an inverse transform of size M × N image g '(x, y) (FIG. 5b), and then refilling 0 into a size of 512 × 512 image I (x, y), the cropped area to fill the 0, g '(x, y) in the image I (x, y) center (FIG. 5c).

重同步的第二步是:平移同步。 The second step is to re-synchronization: synchronous translation. 即用抽取的训练序列S与T的相关系数来确定图象的平移同步参数。 I.e. using training sequences extracted correlation coefficient S and T to determine the translation image synchronization parameter.

平移同步可采用的一种办法是,将图象I(x,y)作如下的所有可能的平移:It(x,y)=I(x-xt)mod512,(y-yt)mod512);{-12(512-M)&le;x1&lt;12(512-M);-12(512-N)&le;y1&lt;12(512-N)}---(6)]]>每次平移后的图象,作DWT分解,从LL3子带的32行与32列中提取训练序列S。 One approach can be synchronized translation is the image I (x, y) for all possible translational follows: It (x, y) = I (x-xt) mod512, (y-yt) mod512); {-12 (512-M) & le; x1 & lt; 12 (512-M); - 12 (512-N) & le; y1 & lt; 12 (512-N)} --- (6)]]> each translating aft image, as DWT decomposition, the line 32 and the sub-band LL3 from 32 to extract the training sequence S. 根据提取的训练序列与原始训练序列的相关系数最大可确定平移参数(xt,yt)。 Determining the maximum translation parameter (xt, yt) in accordance with the correlation coefficients of the training sequence extracted from the original training sequence.

本发明提出的另外一种方法是将图象I(x,y)做最多8×8=64次平移即可,从而可大大减小计算量。 Another method proposed by the present invention is the image I (x, y) made up to 8 × 8 = 64 times can be translated, which can greatly reduce computation. 根据DWT的时频局部化性质,LL3子带的每个系数都对应于图象的一个局部。 Localized nature of the time-frequency DWT according to each coefficient LL3 sub-band corresponds to a partial image. 可以证明(我们的实验也证明了这点),若在DWT时采用紧支小波滤波器和采用周期延拓方式(若采用其他延拓方式,则除了图象边界外,也满足下列关系),图象I(x,y)平移8×xt1行和8×yt1列(xt1,yt1为整数),得到平移图象It(x,y):It(x,y)=I((x-8×xt1)mod512,(y-8×yt1))mod512) (7)则图象三级DWT分解后的LL3子带也平移xt1行和yt1列:LL3t(x,y)=LL3((x-xt1)mod 64,(y-yt1)mod64) (8) Can prove (our experiments also proved this point), the use of a filter and the use of compactly supported wavelets periodic extension DWT manner when (if other extension mode, in addition to the outer boundary of the image, but also satisfy the following relation), the image I (x, y) translating 8 × xt1 8 × yt1 rows and columns (xt1, yt1 is an integer), to give the translation image It (x, y): It (x, y) = I ((x-8 × xt1) mod512, (y-8 × yt1)) mod512) (7) the image LL3 subband decomposition three rows and DWT also translate xt1 yt1 columns: LL3t (x, y) = LL3 ((x- xt1) mod 64, (y-yt1) mod64) (8)

其中LL3(x,y)与LL3t(x,y)分别为图象I(x,y)和It(x,y)LL3子带系数。 Wherein LL3 (x, y) and LL3t (x, y) are the image I (x, y) and It (x, y) LL3 subband coefficients. LL3子带的平移导致嵌入的训练序列也发生平移。 LL3 sub-band translation of the embedded training sequences results in translation also occur. 应用公式7和8给出的性质,我们可以只对I(x,y)做最多8×8次平移:It(x,y)=I((x-xt)mod 512,(y-yt)mod 512);{-4≤xt,yt<4 (9)每平移一次,做DWT分解,获得LL3子带LL3t(x,y)。 Properties are given using the formula 7 and 8, we can only I (x, y) make up 8 × 8 times translation: It (x, y) = I ((x-xt) mod 512, (y-yt) mod 512); {- 4≤xt, yt <4 (9) each time a translation, do DWT decomposition, sub-band LL3 obtained LL3t (x, y). 将LL3t(x,y)作平移:LL3t′(x,y)=LL3t((x-xt1)mod64,(y-yt1)mod64);{-T1≤xt1<T1;-T2≤yt1<T2(10)上式中,T1=round(0.5×(512-M)/8),T2=round(0.5×(512-N)/8)。 The LL3t (x, y) translational: LL3t '(x, y) = LL3t ((x-xt1) mod64, (y-yt1) mod64); {- T1≤xt1 <T1; -T2≤yt1 <T2 ( 10) in the above formula, T1 = round (0.5 × (512-M) / 8), T2 = round (0.5 × (512-N) / 8). 每次平移从LL3t′;(x,y)的32行和32列抽取训练序列S,根据与原始训练序列T间的最大相关值来确定平移参数。 Every translation from the LL3t '; (x, y) of the 32 rows 32 and extracting a training sequence S, the parameters to determine the translation between the maximum correlation value and the original training sequence T. 最多64次平移搜索后即可确定图象的平移参数(8×xt+xt1,8×yt+yt1),从而获得平移校准后的图象g*(x,y)。 Translation parameter can be determined up to 64 times after the image translation search (8 × xt + xt1,8 × yt + yt1), to thereby obtain the shifted calibration image g * (x, y).

本发明具有以下优点:1)本发明提出的DFT-DWT复合域的数字水印盲检测方法,在同时对抗常规信号处理方面和仿射变换方面都达到了较强的稳健性(表1)。 The present invention has the following advantages: 1) The method of watermarking blind detection complex DFT-DWT domain proposed by the present invention, at the same time against the conventional signal processing and have reached the affine transformation is robust (Table 1). 在压缩因子为15的JPEG压缩(JPEG_15)时,能实现无差错检测,能抵抗国际通用水印测试平台StirMark 3.1中除Rand Bending外的其他几何变换,如对rotation(auto crop,auto scale)、jitter、scaling、shearing、general linear transform等都能实现无差错检测,并能抵抗JPEG压缩和几何变换的组合攻击,如同时抵抗JPEG_50压缩、旋转、缩放、裁剪、平移等组合攻击。 When the compression factor of JPEG 15 compression (JPEG_15), to achieve error-free detection, resistant internationally watermark test platform StirMark 3.1, in addition to Rand Bending other geometric transformation, such as rotation (auto crop, auto scale), jitter , scaling, shearing, general linear transform, etc. can be error-free detection, and resistant to JPEG compression and combo attacks geometric transformations, such as simultaneous resistance JPEG_50 compression, rotation, scaling, cropping, translation and other combinations of attack.

2)本发明提出的水印方法可隐藏264比特以上的信息,水印图象相对于原始图象的PSNR在40dB以上,水印的不可见性较好。 2) watermarking method proposed by the invention may be more than 264 bits of information hidden, invisible watermark image with respect to PSNR is better than 40dB in the original image, the watermark.

3)本发明中的图象几何校准技术,精确度高,并可避免在图象中嵌入一个可视标记。 3) the geometry of the image calibration technique of the present invention, high precision, and can avoid embedding a visual marker in the image.

表1 本发明提出方法用水印测试平台StirMark 3.1进行稳健性测试的结果。 Table 1 results of the present invention provides a method for testing the robustness of the watermark test platform StirMark 3.1.

附图说明 BRIEF DESCRIPTION

:图1是水印嵌入框图。 : FIG. 1 is a block diagram of a watermark embedding.

图2是训练序列及其嵌入于LL3子带的位置。 FIG 2 is a training sequence embedded in its position subbands LL3.

图3是在DFT变换幅度谱上半平面嵌入14个模板点的位置示意图。 FIG 3 is a schematic view of a half-plane position of the template 14 is embedded in the point DFT of the amplitude spectrum.

图4是嵌入了DWT-DFT域水印的图象。 FIG 4 is an image embedded watermark DWT-DFT domain.

图5是本发明方法对抗JPEG压缩和RST(旋转,裁剪,缩放和平移组合攻击)联合攻击的稳健性测试。 FIG 5 is a method of the present invention against JPEG compression and RST (rotate, crop, zoom and pan combinations attacks) testing the robustness of the joint attacks.

图1中,1是需要隐藏的比特信息,2是隐藏信息经过直接序列扩频,3是交织,4是原始图象,5是DWT,6是数据嵌入在图象DWT子带系数中,7是IDWT,8是嵌入了DWT域水印的图象f′(x,y),9是将f′(x,y)进行DFT变换,10是在DFT域中嵌入摸板水印。 In FIG 1, a 1 bit information to be hidden, the hidden information is 2 via direct sequence spread spectrum, interleaving is 3, 4 is the original image is DWT 5, is embedded in the image data. 6 DWT subband coefficients, 7 It is IDWT, 8 DWT domain watermark is embedded image f '(x, y), 9 is f' (x, y) for the DFT, formwork 10 embedded watermark in the DFT domain. 11是IDFT。 11 IDFT. 12是嵌入了DWT-DFT域复合水印的图象f″(x,y)。13是嵌入的训练序列。 12 is embedded in a watermarked DWT-DFT domain complex image f "(x, y) .13 embedded training sequence.

图2中,T1...T127是127位训练序列在LL3子带中的某一行和某一列的位置。 In FIG. 2, T1 ... T127 training sequence 127 is a row in the LL3 sub-band and the position of a column. Row 32表示32行,Column 32表示32列。 Row 32 represents a line 32, Column 32 represents 32.

图4、5中给出的是使用标准图象Lena和Baboon测试的一些结果。 Figures 4, 5 is given to using the standard test image Lena and Baboon some results.

图4中,a)Lena水印图象(PSNR=40.1dB);b)Baboon水印图象(PSNR=39.6dB)。 In FIG. 4, a) Lena watermarked image (PSNR = 40.1dB); b) Baboon watermarked image (PSNR = 39.6dB).

图5中,a)Lena水印图象受到JPEG压缩和RST联合攻击后的图象g(x,y)。 FIG. 5, a) Lena watermarked image subjected to image g (x, y) after the JPEG compression and the Joint Strike RST.

b)图a经过较正仿射变换后的图象g′(x,y)。 b) after a more positive affine FIG g converted image '(x, y). 图象大小为504×504。 Picture size of 504 × 504.

c)图b周围补0到512×512大小I(x,y)。 c) Panel b to around zero with the size of 512 × 512 I (x, y).

d)图c经过平移校正后的图象g*(x,y)图象大小为512×512,填充部分为图象g(x,y)的均值。 d) FIG g c through the shifted image correction * (x, y) picture size of 512 × 512, mean filling portion of the image g (x, y) of. 隐藏的264比特仍然可以无差错检测出来。 Hidden 264-bit error-free can still be detected.

具体实施方式 Detailed ways

实施例:计算机、打印机和高速传输设备的发展使得在网络上进行图象和视频信号的传输变得非常方便。 Example: the development of computers, printers and high-speed transmission so that the transmission device for image and video signals over the network has become very convenient. 但电子图象、视频等的传输和储存所面临的一个严重问题是它们的复制品与原件完全一样,因而版权所有者不愿意以这种方式传播他们的材料。 But a serious problem of electronic images, video transmission and storage are facing is that they are exactly the same as the original replica, so the copyright owner in this way do not want to spread their material. 由于因特网在商业领域的应用日益广泛,因此急切需要一种可以对电子数据进行保护的手段。 Since the application of the Internet in the business world increasingly widespread, and thus may be an urgent need for a means to protect electronic data. 数字水印是携带所有者版权信息的一组辨别手段。 Digital watermarking is a set of tools to identify the copyright owner to carry information. 数字水印被永久地嵌入到多媒体数据中用于版权保护并检查数据是否被破坏。 Digital watermarking is permanently embedded in the multimedia data for copyright protection and checking whether the data is corrupted. 但目前的数字水印技术大多在图象经过一些很小的几何变形后就检测不出来。 But the current digital watermarking technology mostly in the picture after some minor after the geometric distortion can not be detected. 因此本发明的最大优点在于使嵌入的数字水印能同时抵抗几何变形和图象压缩编码的攻击等。 Therefore, the maximum advantage of the invention is to make the embedded digital watermark can simultaneously resist geometric deformation and image compression coding attacks.

下面给出抗几何变换的数字水印技术在标准图象Lena和Baboon使用和测试得到的一些结果。 Anti geometric transformations are given below digital watermarking technology standard images Lena and Baboon use and testing of some of the results obtained.

我们在Lena和Baboon(均为512×512×8bits)的图象上分别嵌入一个包含44个字符(264比特)的信息水印、匹配模板和训练序列,PSNR值分别为40.1dB和39.6dB(图4)。 We embed watermark information respectively containing 44 characters (264 bits) in the Lena image and the Baboon (both 512 × 512 × 8bits), the matching template and the training sequence, the PSNR values ​​of 40.1dB and 39.6dB (FIG. 4). 测试中我们取L=264,N1=15,α=56,threshl=0.56,threshold=0.002,Nb=180,Nm=5,Kmin=0.5和Kmax=2.0,以及在对图象离散小波变换中使用Daubechies 9/7双正交小波滤波器,训练序列嵌入在LL3子带的32行和32列。 Tests we take L = 264, N1 = 15, α = 56, threshl = 0.56, threshold = 0.002, Nb = 180, Nm = 5, Kmin = 0.5 and Kmax = 2.0, and using the image in the discrete wavelet transform biorthogonal Daubechies 9/7 wavelet filter, the training sequence embedded in the LL3 sub-band line 32 and 32. 本发明水印嵌入方法在1.7G的P4计算机(windows平台,VC++语言)上需要小于3秒时间,而检测方法需要1~13秒左右。 Watermark embedding method of the present invention is 1.7G P4 computer needs time less than 3 seconds (windows platform, VC ++ language) and the detection method requires about 1 to 13 seconds. 可以看出,计算量并不太大。 As can be seen, the amount of calculation is not too large.

表1是信息水印抵抗StirMark3.1攻击的情况。 Table 1 is the resistance situation StirMark3.1 information watermark attacks. 其中BER(Bit Error Rate)是指误比特率。 Wherein the BER (Bit Error Rate) refers to the bit error rate. 水印对StirMark3.1中缩放、旋转+裁剪、旋转+裁剪+缩放、一般线性变换、剪切、jitter、改变长宽比等几何攻击都能实现无差错检测。 StirMark3.1 watermark scaling, rotation + cropping, rotating, cropping + + Scale, general linear transformation, shear, Jitter, an attacker can change the aspect ratio of the geometric-free error detection. 在StirMark3.1中cropping_25、JPEG_15攻击时能实现无差错检测。 In StirMark3.1 in cropping_25, JPEG_15 to achieve error-free detection of attack. 对Gaussian滤波、锐化,水印字符串也可以无差错率或以极低的差错率提取出来。 Of the Gaussian filter, sharpening, the string may be no or a very low error rate, error rate extracted.

另外,我们还测试了在水印图象受到任意的旋转+剪切+缩放+平移(RST)的情况下水印的稳健性。 In addition, we also tested the watermark image in the rotation by any cut + scaling + robustness of the watermark at translational + case (RST) is. 如图5所示。 As shown in FIG. 图5(a)是Lena水印图象受到JPEG压缩和RST联合攻击后的图象g(x,y)。 FIG 5 (a) is subjected to image Lena watermarked image g (x, y) after the JPEG compression and the Joint Strike RST. 图5(b)是图5(a)经过较正仿射变换后恢复其原始形状后的图象g′(x,y)。 FIG. 5 (b) are diagrams 5 (a) more positive after the affine transformation to restore its original shape after the image g '(x, y). 图象大小为504×504。 Picture size of 504 × 504. 图5(c)是在图象g′(x,y)周围填补0到512×512大小的图象I(x,y),补0后原图象g′(x,y)位于的图象I(x,y)的中央。 FIG. 5 (c) is the image g '(x, y) 0 to fill the 512 × 512 size image I (x, y) around the original image after the meeting 0 g' (x, y) is located in FIG. as I (x, y) of the center. 图5(d)是图象I(x,y)经过平移校正后的图象g*(x,y),图象大小为512×512。 FIG. 5 (d) is an image I (x, y) after correction of the shifted image g * (x, y), the image size is 512 × 512. 水印检测时,填充部分以图象g(x,y)的均值填充,而不是以0值填充,这样做的结果是可以改善检测性能。 When the watermark detection, the filling section to the image g (x, y) mean filling, filling value instead of 0, the result of this is that detection performance can be improved. 测试结果是隐藏的264比特仍然可以无差错检测出来。 The test results are hidden 264-bit error-free can still be detected.

Claims (2)

1.一种图象几何校准和保护数字图象的方法,其特征是该方法首先将信息水印经扩频调制和交织后与一个训练序列一起嵌入到图象DWT域中,再将一个匹配模板嵌入到图象DFT域,最后由嵌入于图象DFT域的匹配模板和嵌入于图象DWT域的训练序列来实现扩频调制和交织后嵌入于图象DWT域中的信息水印的重同步检测,具体做法是:1)水印嵌入:水印嵌入过程主要有包括信息水印、训练序列的DWT域水印的预处理、DWT域水印的嵌入和DFT域摸板水印的嵌入三部分;i)DWT域水印的预处理:直接序列扩频编码、交织;应用长度为N1的PN码序列m={mj;j=1,...,N1}对要嵌入的信息b{bi;i=1,...,L}其中bi∈{0,1},进行扩频编码调制;“1”调制为m的正相序列,即+1×mj;j=1,...,N1,“0”调制为m的反相序列,即{-1×mj;j=1,...,N1},这样可得到待嵌入的二进制水印数据W;在 An image geometry and the calibration method of protecting digital image, wherein the method is embedded into the first watermark information by interleaving and spread spectrum modulation with a training sequence with DWT domain image, then a matching template re-embedded in image DFT domain, and finally embedded in the DFT domain matching template image and the embedded watermark information is embedded in the DWT domain image after the training sequence of images is achieved on DWT interleaving and spread spectrum modulation synchronous detection , specifically: 1) the watermark embedding: watermark embedding process mainly comprises pretreatment DWT domain watermark watermark information, training sequence, and the embedded watermark embedding DFT domain formwork DWT domain watermark three parts; I) DWT domain watermark pretreatment: direct sequence spread spectrum coding, interleaving; application length of the PN code sequence m N1 = {mj; j = 1, ..., N1} of the information to be embedded b {bi; i = 1, .. ., L} where bi∈ {0,1}, spreading coded modulation; "1" is a normal phase modulation sequence m, i.e., + 1 × mj; j = 1, ..., N1, "0" modulation m is the phase-sequence, i.e., {-1 × mj; j = 1, ..., N1}, thus obtaining the data to be embedded watermark W is binary; in 个与图象DWT低频子带相同大小的二维矩阵的中心行和中心列上或者图象中需要重点保护部分对应的低频子带部位存放由密钥产生的伪随机训练序列T,其余位置顺序存放交织后的二进制水印数据W,将得到的二维矩阵按行扫描变成一个一维数组,记为X;ii)DWT域水印的嵌入原始图象f(x,y)进行三级DWT分解,把低频子带LL3系数按行扫描变成一维数组,记为C;按下列公式,我们把二进制数据X加到低频系数C上,得到新的低频系数C′: The center line and a two-dimensional matrix image DWT lower sub-band of the same size and position of the center column or the remaining portion of protection desired sequence portion corresponding to the lower sub-band pseudo-random training sequence stored T generated by the key, the image after storing interleaved binary watermark data W, obtained by the two-dimensional matrix into a one-dimensional line scan array, referred to as X; ii) DWT domain watermark is embedded in the original image f (x, y) for three DWT decomposition , the lower sub-band LL3 coefficients are scanned into a one-dimensional array of rows, referred to are C; according to the following formula, we added to the binary data X low frequency coefficients C, giving the new low-frequency coefficients C ': 其中C(i)、C′(i)、xi分别为C、C′、X的第i个元素;α表示水印嵌入强度;将嵌入水印后的小波系数进行IDWT得到嵌入DWT域水印的图象f′(x,y);iii)DFT域模板的嵌入:在嵌入DWT域水印后图象f′(x,y)的DFT域,沿过原点的两条直线增大一些中频点处傅立叶系数的模值,使这些点成为局部区域的极大值;改变量以不可见为标准,一般取极大值为局部平均值加上几倍到十几倍左右的方差;这些嵌入的沿两条直线分布的局部极大值点构成一个模板,用作水印图象变形后的同步信息;这些模板点的位置可由一个密钥控制产生;这两条过原点的直线称为模板线;2)水印的检测:检测过程如下:a)在水印检测时,首先要应用训练序列检测水印是否同步;若不同步,则必须先将测试图象经过重同步得到同步图象g*(x,y),重同步包括DFT域摸板水印检测、逆仿射变换、应用训 Wherein C (i), C '(i), xi are C, C', X i-th element; [alpha] represents a watermark embedding strength; wavelet coefficients obtained IDWT watermark embedding the domain watermark embedded image DWT f '(x, y); iii) embedded in the DFT domain template: DWT domain watermark is embedded in the image f' (x, y) in the DFT domain, along two straight lines through the origin of the increased number of Fourier coefficients at intermediate points modulus value, so that these points become the maximum value of the local area; invisibly change amount as the standard, and generally the maximum value of the mean plus local variance times to about ten times; these embedded along two the local maximum point of the line profile constituting a template, the synchronization information as a watermark image after deformation; the location of these points by the template generating a control key; straight line through the origin of these two lines is called a template; 2) watermark detection: detection process is as follows: a) during watermark detection, whether to apply the first training sequence to detect the watermark synchronization; if the synchronization, the first test images must be synchronized after resynchronization image g * (x, y), DFT domain formwork comprising resynchronization watermark detection, the inverse affine transformation, application training 序列平移同步;若同步,直接做下一步;b)对同步图象g*(x,y)作DWT域水印检测,得到了实际隐藏的数据;重同步的第一步是:恢复原始几何形状;从待测图象g(x,y)中检测出嵌入的模板水印,并将之与原始的嵌入模板进行对比获得图象所经受的仿射变换矩阵B;假设原始图象大小为M1×M1,模板水印检测的步骤如下:a)对待测图象g(x,y)作Barlette滤波;b)同嵌入模板时一样,将滤波后的待测图象扩展至1024×1024;c)作DFT变换;以一个半径为R′的圆形窗口,在傅立叶系数幅度矩阵的上半平面中搜索,提取所有局部极大值点;把DFT系数幅度矩阵上半平面以原点为顶点划分为Nb个扇形区域,每个扇形的顶角均为0.5°或1°;再按角度将所有局部极大值点分别归入各个扇形区域;d)找到与两条模板线对应的可能的模板点集合;在每个扇形区域中,在Kmin<K<Kmax范围内搜 Translation synchronization sequence; if the synchronization and directly in the next step; b) synchronization image g * (x, y) for the DWT domain watermark detection to obtain the actual hidden data; resynchronize the first step: to restore the original geometry ; detecting the test image g (x, y) in the template watermark embedded, and compared with the original template embedded image obtained is subjected to affine transformation matrix B; assumes that the original image size of M1 × step M1, the template watermark detection as follows: a) treat prediction image g (x, y) for filtering Barlette; b) embedding the same time as the template, the filtered image to be tested extended to 1024 × 1024; c) for the DFT; to a radius R 'of the circular window, the search in the upper half plane of the Fourier coefficients of the amplitude matrix, extract all the local maxima; matrix of the DFT coefficient magnitude as the upper half plane origin to the vertices is divided into a Nb fan-shaped area, the apex angle of each sector are 0.5 ° or 1 °; then all the local maxima of the angle points are included in each sector region; D) may be found in point two template sets corresponding to the template line; in each sector area, in the search Kmin <<Kmax range K 这样的K值:它使得此扇区中至少有Nm个局部极大值点满足|rli-KrTj&prime;|&lt;threshold,]]>其中Nm为一个预先规定的数,rli是扇区i中局部极值点的极径(i=1...Nb),rTj′是原模板线j(j=1,2)上摸板点的极径,threshold>0为一阈值。 Such a K value: it makes this sector has at least Nm a local maximum point satisfies | rli-KrTj & prime; | & lt; threshold,]]> where Nm is a predetermined number, rli is the sector i topical extreme point of the polar radius (i = 1 ... Nb), rTj 'original template line j (j = 1,2) on the formwork polar radius points, threshold> 0 is a threshold value. 如果找到这样的K值,我们就把相应的局部极值点坐标记录下来;通过上述步骤,得到可能的匹配线的集合,称为“准匹配线”,线上的局部极值点称为“准匹配点”,坐标记为(xij,yij);图象上半平面相应的原始模板点的坐标记为(xij′,yij′),其中i∈{1,2}表示第i条模板匹配线,j ∈{1,2,Λ}表示第j个模板匹配点;从对应于模板线1的准匹配点集中取出一个集合和对应于模板线2的准匹配点集中取出另一个集合;根据这两个集合的点与原始模板点间的对应关系计算得到的一个可能的变换矩阵A;寻找平均误差MAE最小的4;MAE=1nummatches||Ax11y11MMx1jy1jx21y21MMx2jy2jT-x11&prime;y11&prime;MMx1j&prime;x1j&prime;x21&prime;y21&prime;MMx2j&prime;y2j&prime;T||]]>其中模板点为(xij′,yij′)和“准匹配点”为(xij,yij),nummatches是匹配点个数,运算符||Λ||中是一个2行的误差矩阵;f)将对应于模 If such a value of K, we put the corresponding local extrema point coordinate recorded; By the above procedure, the set of possible match line, referred to as "quasi-match line" local extreme points on a line called " quasi matching point ", as coordinate notation (xij, yij); sit indicia on the template image corresponding to the original point of the half plane (xij ', yij'), where i∈ {1,2} denotes the i-th template matching line, j ∈ {1,2, Λ} denotes the j-th template matching points; concentration corresponding to a set and another set of quasi-extraction template matching point corresponding to the line 2 from quasi-line template matching point extraction concentration 1; according to the correspondence between these two points with the set of original template point calculated a possible transformation matrix a; Looking average error MAE minimum 4; MAE = 1nummatches || Ax11y11MMx1jy1jx21y21MMx2jy2jT-x11 & prime; y11 & prime; MMx1j & prime; x1j & prime; x21 & prime; y21 & prime ; MMx2j & prime; y2j & prime; T ||]]> wherein the template points (xij ', yij') and the "quasi matching point" is (xij, yij), nummatches is the number of matching points, the operator || Lambda || of 2 is an error matrix rows; F) corresponding to the mold 线1的准匹配点加上180°,重复e),由最小的MAE值确定最后的频域变换矩阵A;可得空域变换矩阵B=AT;获得仿射变换矩阵B后,将待测图象g(x,y)进行图象几何逆变换恢复成M×N大小的图象g′(x,y),然后再填充0成M1×M1大小的图象I(x,y),被裁剪的部分以0填充,g(x,y)在图象I(x,y)中心;重同步的第二步是:平移同步,即用抽取的训练序列S与原始训练序列T的相关系数来确定图象的平移同步参数;平移同步方法是将图象I(x,y)做最多8×8=64次平移即可:It(x,y)=I((x-xt)mod N2,(y-yt)mod N2);{-4≤xt,yt<4}每平移一次,做DWT分解,获得LL3子带LL3t(x,y);将LL3t(x,y)作平移:LL3t&prime;(x,y)=LL3t((x-xt1)mod64,(y-yt1)mod64);]]>{-T1≤xt1<T1;-T2≤yt1<T2}上式中,T1=round(0.5×(M1-M)/8),T2=round(0.5×(M1-N)/8),每次平移从LL3′,(x,y)的中心行和中心列抽取训练序列S,根据与原始训练序列 1 line quasi matching point plus 180 °, repeating e), to determine the final frequency domain transformation matrix A by a minimum value of MAE; available spatial transformation matrix B = AT; After obtaining the affine transformation matrix B, and FIG tested as g (x, y) for image geometric restored to an inverse transform of size M × N image g '(x, y), and then filled into M1 × M1 0 size image I (x, y), is zero-padded portion cut, g (x, y) in the image I (x, y) center; resynchronization second step: synchronous translation, i.e., the correlation coefficient with the extracted training sequence S and the original training sequence T translating the image to determine the synchronization parameter; synchronization method is to pan the image I (x, y) made up to 8 × 8 = 64 times translate to: It (x, y) = I ((x-xt) mod N2 , (y-yt) mod N2); {- 4≤xt, yt <4} for each shift once, do DWT decomposition, to obtain LL3 sub-band LL3t (x, y); the LL3t (x, y) translational: LL3t & prime ; (x, y) = LL3t ((x-xt1) mod64, (y-yt1) mod64);]]> {- T1≤xt1 <T1; -T2≤yt1 <T2} in the above formula, T1 = round ( 0.5 × (M1-M) / 8), T2 = round (0.5 × (M1-N) / 8), each translation from the LL3 ', (x, y) of the central row and central column to extract the training sequence S, according to the original training sequence T间的最大相关值来确定平移参数,最多64次平移搜索后即可确定图象的平移参数(8×xt+xt1,8×yt+yt1),从而获得平移校准后的同步图象g*(x,y);DWT域水印的检测:把同步图象g*(x,y)DWT分解后的低频子带LL3系数按行扫描变成一维数组,记为C*;抽取出来的二进制数据记为X*={xi*},抽取公式如下:xi*=+1,(C*(i)mod&alpha;)&GreaterEqual;&alpha;2xi*=-1,otherwise]]>将抽取的二进制数据X*进行反交织(交织的逆过程)恢复嵌入的二进制数据序列W*;然后W*按N1位比特进行分段,每段与序列m进行相关,若相关值大于0,则判决嵌入信息比特为“1”,否则判决嵌入信息比特为“0”;解扩之后就得到恢复的嵌入信息。 Synchronization image g after the maximum correlation value to determine the translation between the parameters T, to determine the translation parameter image (8 × xt + xt1,8 × yt + yt1) after translation of searches up to 64 times to obtain a translation of calibration * binary extracted; and synchronizing image g * (x, y) decomposing the low frequency subband LL3 DWT coefficients are scanned into a one-dimensional array of rows, referred to as C *:; (x, y) of the domain watermark detection DWT data referred to as X * = {xi *}, extracting the following formula: xi * = + 1, (C * (i) mod & alpha;) & GreaterEqual; & alpha; 2xi * = - 1, otherwise]]> the extracted binary data X * anti-interleaving (the inverse interleaving process) embedded in a sequence of binary data recovery W *; and W * N1 press segmented bits, each sequence of m and correlating, if the correlation value is greater than 0, then decision bits embedded information to "1", the embedded information or the decision bit is "0"; obtained after despreading to recover the embedded information.
2.如权利要求1所述的一种图象几何校准和保护数字图象的方法,其特征是在图象DFT变换的幅度谱系数中嵌入由局部极大点构成的一个匹配模板,用于检测图象所经受仿射变换的变换矩阵并作逆变换恢复其原始几何形状;同时在图象DWT域嵌入一个训练序列用于图象平移校准。 2. An image geometry and the calibration method of protecting digital images according to claim 1, characterized in that embedded in a matching template consisting of a local maximum in the amplitude spectrum the number of images in the DFT, for the detected image is subjected to affine transformation, and the transformation matrix as the inverse transform to recover its original geometry; simultaneously embedding a training sequence for an image in the image and DWT translate calibration.
CNB021496145A 2002-12-12 2002-12-12 Watermark method using geometry calibrated anti-geometry conversion image CN1321393C (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CNB021496145A CN1321393C (en) 2002-12-12 2002-12-12 Watermark method using geometry calibrated anti-geometry conversion image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CNB021496145A CN1321393C (en) 2002-12-12 2002-12-12 Watermark method using geometry calibrated anti-geometry conversion image

Publications (2)

Publication Number Publication Date
CN1414778A CN1414778A (en) 2003-04-30
CN1321393C true CN1321393C (en) 2007-06-13

Family

ID=4751698

Family Applications (1)

Application Number Title Priority Date Filing Date
CNB021496145A CN1321393C (en) 2002-12-12 2002-12-12 Watermark method using geometry calibrated anti-geometry conversion image

Country Status (1)

Country Link
CN (1) CN1321393C (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100521781C (en) 2003-07-25 2009-07-29 皇家飞利浦电子股份有限公司 Method and device for generating and detecting fingerprints for synchronizing audio and video
US7266216B2 (en) * 2003-08-07 2007-09-04 International Business Machines Corporation Inserting and detecting watermarks in images derived from a source image
AU2004300115B2 (en) 2003-12-19 2010-11-25 Creative Technology Ltd Method and system to process a digital image
PL1684265T3 (en) 2005-01-21 2009-01-30 Unlimited Media Gmbh Method of embedding a digital watermark in a useful signal
CN101282457B (en) 2005-02-06 2010-06-02 陆健 False proof detection method for real time monitoring videosignal
CN103177413B (en) 2011-12-20 2016-04-13 深圳市腾讯计算机系统有限公司 The method that localization blind watermatking generates, detect and device
CN105023235B (en) * 2015-07-09 2018-06-12 哈尔滨工程大学 A kind of electronic chart water mark method based on spatial redundancy relationship
CN105072453B (en) * 2015-07-21 2018-07-24 河海大学 A kind of video watermark process of facing moving terminal

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001078010A (en) * 1999-09-02 2001-03-23 Hitachi Ltd Method for extracting electronic watermark information
CN1383528A (en) * 2000-06-23 2002-12-04 皇家菲利浦电子有限公司 Watermark embedding method and arrangement

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001078010A (en) * 1999-09-02 2001-03-23 Hitachi Ltd Method for extracting electronic watermark information
CN1383528A (en) * 2000-06-23 2002-12-04 皇家菲利浦电子有限公司 Watermark embedding method and arrangement

Also Published As

Publication number Publication date
CN1414778A (en) 2003-04-30

Similar Documents

Publication Publication Date Title
Petitcolas et al. Evaluation of copyright marking systems
Chang et al. SVD-based digital image watermarking scheme
US6031914A (en) Method and apparatus for embedding data, including watermarks, in human perceptible images
Swanson et al. Multimedia data-embedding and watermarking technologies
Pereira et al. Fast robust template matching for affine resistant image watermarks
Pereira et al. Robust template matching for affine resistant image watermarks
US7263203B2 (en) Digital watermarks
US6385329B1 (en) Wavelet domain watermarks
US7068809B2 (en) Segmentation in digital watermarking
Voloshynovskiy et al. Multibit digital watermarking robust against local nonlinear geometrical distortions
Deguillaume et al. Method for the estimation and recovering from general affine transforms in digital watermarking applications
US8144924B2 (en) Content objects with computer instructions steganographically encoded therein, and associated methods
Kashyap et al. Image watermarking using 3-level discrete wavelet transform (DWT)
US6226387B1 (en) Method and apparatus for scene-based video watermarking
US8036419B2 (en) Digital watermarks
Rastegar et al. Hybrid watermarking algorithm based on singular value decomposition and radon transform
US6373974B2 (en) Method for extracting multiresolution watermark images to determine rightful ownership
Kaur et al. Steganographic approach for hiding image in DCT domain
Lumini et al. A wavelet-based image watermarking scheme
US6625297B1 (en) Self-orienting watermarks
US7346776B2 (en) Authenticating media signals by adjusting frequency characteristics to reference values
US6904151B2 (en) Method for the estimation and recovering of general affine transform
EP1444653B1 (en) Digital watermarking method robust against local and global geometrical distortions and projective transforms
Barni et al. Improved wavelet-based watermarking through pixel-wise masking
Nikolaidis et al. Asymptotically optimal detection for additive watermarking in the DCT and DWT domains

Legal Events

Date Code Title Description
PB01 Publication
C06 Publication
SE01 Entry into force of request for substantive examination
C10 Entry into substantive examination
GR01 Patent grant
C14 Grant of patent or utility model
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20070613

Termination date: 20100112

C17 Cessation of patent right