CN110246076A - The high dynamic range images water mark method decomposed based on Tucker - Google Patents

Abstract

The invention discloses a kind of high dynamic range images water mark methods decomposed based on Tucker, it includes watermark insertion and watermark extracting two parts, in watermark telescopiny, high dynamic range host image is expressed as to the form of three rank tensors, then high dynamic range host image is handled using Tucker3 decomposition, and using the fisrt feature figure of obtained core tensor as the insertion carrier of watermark information, the fisrt feature figure for being not only because core tensor covers the main energetic of a panel height dynamic range host image, but also the watermark information of insertion can be made as the inverse transformation of decomposition is diffused into the R of high dynamic range watermarking images, G, in tri- channels B, to effectively improve robustness；Original high dynamic range host image is not needed in watermark extraction process, realizes the Blind extracting of watermark information；In addition, the method for the present invention will not damage the internal structure of high dynamic range images, preferable effect is also achieved in terms of invisibility and embedding capacity.

Description

The high dynamic range images water mark method decomposed based on Tucker

Technical field

The present invention relates to a kind of image watermark methods, more particularly, to a kind of high dynamic range figure decomposed based on Tucker As water mark method.

Background technique

As high dynamic range (High Dynamic Range, HDR) imaging technique is in consumer electronics, making video, void The extensive use in the fields such as quasi- reality, remote sensing and medicine detection, the Copyright Protection of corresponding high dynamic range images Also just of increasing concern.Different, the high dynamic range with traditional low-dynamic range (Low Dynamic Range, LDR) image Image is usually to be indicated and stored with real-coded GA, and its dynamic range generally can reach or more than 9 quantity Grade, this can not only improve the accuracy of scene brightness information to a certain extent, but also can bring color more abundant Details and shade.However, since existing low-dynamic range shows that equipment is typically only capable to the dynamic model of 3 orders of magnitude of display It encloses, therefore in order to showing the detailed information abundant of high dynamic range images on general display devices, generally requires pair High dynamic range images carry out tone mapping (Tone Mapping, TM) pretreatment, and for the copyright of high dynamic range images For protection, this is undoubtedly a kind of distinctive again and inevitably attacks form.In addition, by being reflected to existing global tone The analysis for penetrating operator finds that tone mapping is a nonlinear mapping process, specific manifestation are as follows: when the height containing watermark information Dynamic image can be with image pixel number in the watermark information for after tone mapping is handled, being embedded in low brightness area The amplification of value and increase, and the watermark information for being embedded in high-brightness region can then reduce with the compression of image pixel numerical value It even loses, so that the highlight regions in image show poor robustness, i.e., is showed in the different luminance areas of image Different robustness.Although traditional low dynamic range echograms digital watermark is increasingly mature, it is contemplated that high dynamic range The nonlinear characteristic of the characteristics of enclosing image and tone mapping attack, if traditional low dynamic range echograms digital watermark is simple High dynamic range images field is transplanted on ground, then is difficult to show preferable performance.

Currently, the research for high dynamic range images digital watermark is broadly divided into following two categories:

One kind is fragile watermarking, and what such methods were more focused on is invisibility and embedding capacity, often using high dynamic The storage format of state range image carries out the insertion of watermark information in airspace.Such as: Yu, Wang, Chang are to utilize RGBE per capita Index channel E under storage format instructs the lossless insertion of watermark information；Cheng, Li et al. people are respectively in connection with least significant bit (Least Significant Bit, LSB) algorithm is deposited in the RGBE storage format or LogLuv (TIFF) of high dynamic range images Store up the insertion that watermark information is carried out on format；Lin et al. is then to utilize high dynamic range images under OpenEXR storage format 10 mantissa bits convey secret information.

Another kind of is multipurpose watermarking, and such methods consider the features such as the structure of high dynamic range images itself mostly, The insertion of watermark information is carried out in the transform domain of image.Such as: Guerrini et al. devises one based on brightness, texture and side The perception exposure mask of edge information, and existed using quantization index modulation (Quantization Index Modulation, QIM) technology The low frequency sub-band in the domain wavelet transform (Discrete Wavelet Transformation, DWT) characterized by kurtosis value into The insertion of row watermark information, which attacks 7 kinds of tone mappings and 15 panel height dynamic images are tested, and has one Fixed robustness and invisibility, but average error rate is up to 29%；Xue et al. successively proposes two schemes, one is first The processing of tone mapping based on μ-Law is carried out to obtain corresponding low dynamic range echograms to high dynamic range images, is then existed The domain DWT of low dynamic range echograms carries out the insertion of watermark information, locates in advance the second is first carrying out logarithm to high dynamic range images Reason, and filter to obtain levels of detail using two-sided filter, the insertion of watermark information is then carried out in the domain DWT of levels of detail, this two Although kind of a scheme embodies preferable invisibility, only few high dynamic range images are tested, and accidentally Code rate is still higher；Wu et al. is then in the domain discrete cosine transform (Discrete Cosine Transform, DCT) to known Tone mapping treated image carries out the insertion of watermark information, however the algorithm is only to tone predetermined in telescopiny Mapping attack and its relevant parameter show preferable performance, this has significant limitation in practical applications； Solachidis et al. is sequentially embedded watermark information in low dynamic range echograms of the domain DWT to different depth of exposure, this multiple Though the thought of insertion obtains preferable robustness, being decreased obviously for invisibility is also resulted in simultaneously；The same year, Solachidis et al. combines human visual system (Human Visual System, HVS) characteristic to devise in the domain DWT again One based on just can perceptual distortion (Just Noticeable Difference, JND) and contrast sensitivity function (Contrast Sensitivity Function, CSF) perception exposure mask, and it is strong to select suitably to be embedded in region and reasonable insertion with this Degree, which shows preferable robustness when fighting 7 kinds of tone mapping attacks, but embedded quantity only has 128bit； Maiorana et al. is then first high dynamic range images successively to have been carried out with logarithm, DWT and RDCT processing, and then utilize quantization Index modulation technology is embedded in watermark information, which attacks 6 kinds of tone mappings and 15 panel height dynamic images are surveyed Examination has preferable invisibility, but average error rate is still about 20%；Anbarjafari et al. first to host image according to It is secondary carried out piecemeal, DWT, chirp z transform (Chirp-Z Transformation, CZT), QR (ORTHOGONAL TRIANGULAR) decompose with And the pretreatment such as singular value decomposition (Singular value decomposition, SVD) is to obtain the singular value of host image Then matrix S carries out singular value decomposition to original watermark information again and obtains S_wAnd additive insertion principle is combined to complete watermark information Insertion, which shows preferable robustness when fighting the attack of 14 kinds of tone mappings, but its watermark extraction process is Non-blind；Bai et al. proposes a kind of Robust Digital Watermarking Algorithm based on spatial activity, and the algorithm is first with non-lower sampling Contourlet transform (Non-subsampled Contourlet Transform, NSCT) and singular value decomposition extract phenogram The structural information of picture is embedded in carrier as watermark, and combines the proposed active conceptual design of high dynamic range images watermark one Kind classification embedment strength mechanism and mixing perception exposure mask optimize the robustness and invisibility of algorithm, which is fighting Existing 27 kinds of tone mappings show preferable robustness when attacking.Although in conclusion existing high dynamic range figure As Robust Digital Watermarking Algorithm has achieved certain effect when fighting a small number of tone mappings attack, but this kind of algorithm is not abundant Consider following two problem: first, DCT, DWT, redundant wavelet transformation (Redundant Discrete Wavelet Transform, RDWT), these traditional data processing methods such as NSCT and SVD be to convert vector for high dimensional data to carry out table Show, however this dimension for not only resulting in data sample is excessively high, and can also destroy the internal structure of data itself simultaneously；Its Two, tone mapping is a nonlinear mapping process, i.e., watermarking images can be in different brightness after being attacked by tone mapping Region shows inconsistent robustness.Therefore, tone mapping attack can be effectively antagonized by studying one kind, and in image procossing mistake It will not be to the high robust high dynamic range images blind watermark method that the internal structure of image damages, to high dynamic range in journey It encloses particularly important for the copyright protection of image.

Summary of the invention

Technical problem to be solved by the invention is to provide a kind of high dynamic range images watermarks decomposed based on Tucker Method will not damage the internal structure of high dynamic range images, have when fighting tone mapping attack preferable Robustness also achieves preferable effect in terms of invisibility and embedding capacity, and watermark extracting is Blind extracting.

The technical scheme of the invention to solve the technical problem is: a kind of high dynamic range decomposed based on Tucker Enclose image watermark method, it is characterised in that including watermark insertion and watermark extracting two parts；

The specific steps of the watermark embedding section are as follows:

Step 1_1: I is enabled_hostIndicate the high dynamic range host image of watermark information to be embedded, I_hostFor RGB color figure Picture, by I_hostR Color Channel, G Color Channel and B color channel correspondence be denoted as I_{host_r}、I_{host_g}And I_{host_b}；Wherein, I_host、 I_{host_r}、I_{host_g}And I_{host_b}Width be M_hostAnd height is N_host；

Step 1_2: by I_hostIt is expressed as the form of three rank tensors, is denoted as A_host；Then Tucker3 decomposition algorithm pair is utilized A_hostTensor resolution is carried out, A is obtained_hostCore tensor, be denoted as B_host；And by B_hostThe 1st channel as I_hostFirst Characteristic pattern is denoted as B_{host_1}；By B_hostThe 2nd channel as I_hostSecond feature image, be denoted as B_{host_2}；By B_hostThe 3rd A channel is as I_hostThird feature image, be denoted as B_{host_3}；Wherein, A_hostAnd B_hostSize be M_host×N_host× 3, B_{host_1}、B_{host_2}And B_{host_3}Width be M_hostAnd height is N_host；

Step 1_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Not into The filling of row pixel, and by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Again correspondence is denoted as B'_{host_1}、I'_{host_r}、I'_{host_g}With I'_{host_b}；

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b} Respective downside carries out the filling of pixel, fills N_block-mod(N_host,N_block) row, and by B_{host_1}Filling through pixel The image obtained afterwards is denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b} Respective right side carries out the filling of pixel, fills N_block-mod(M_host,N_block) column, and by B_{host_1}Filling through pixel The image obtained afterwards is denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}With I_{host_b}Respective right side and downside carry out the filling of pixel, fill N_block-mod(M_host,N_block) column and N_block-mod (N_host,N_block) row, and by B_{host_1}The image obtained after the filling of pixel is denoted as B' again_{host_1}, by I_{host_r}Through pixel The image obtained after the filling of point is denoted as I' again_{host_r}, by I_{host_g}The image obtained after the filling of pixel is denoted as again I'_{host_g}, by I_{host_b}The image obtained after the filling of pixel is denoted as I' again_{host_b}；

It is above-mentioned, N_blockFor odd number, N_block> 1, B'_{host_1}、I'_{host_r}、I'_{host_g}And I'_{host_b}Width be M'_hostAnd Height is N'_host,SymbolFor the fortune that rounds up Operator number, mod () indicate modulo operation symbol；

Step 1_4: by B'_{host_1}It is embedded in carrier as watermark, by B'_{host_1}It is divided intoIt is a not overlap Size be N_block×N_blockImage block, will be to B'_{host_1}Coordinate position is in all image blocks obtained after division The image block of (i, j) is denoted as B'_{host_1}(i,j)；Wherein,

Step 1_5: by I'_{host_r}、I'_{host_g}And I'_{host_b}To should be used as I'_hostR Color Channel, G Color Channel and B Color Channel；Then I' is obtained_hostWidth beAnd height isBrightness exposure mask, be denoted as Mask_lum, and will Mask_lumIt is saved as code key Key1；Then according to Mask_lumIn the pixel value of each pixel judge whether B'_{host_1}In same coordinate position image block in be embedded in watermark information, for B'_{host_1}(i, j), if Mask_lum(i, j)=0 Then determine in B'_{host_1}It is not embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as not process block；If Mask_lum(i, J)=1 judgement is in B'_{host_1}It is embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as to be embedded piece of watermark；Its In, I'_hostWidth be M'_hostAnd height is N'_host, Mask_lum(i, j) indicates Mask_lumMiddle coordinate position is the picture of (i, j) The pixel value of vegetarian refreshments；

Step 1_6: in B'_{host_1}In to be embedded piece of each watermark in be embedded in watermark information, if B'_{host_1}(i, j) is water To be embedded piece is printed, then in B'_{host_1}The process of insertion watermark information in (i, j) are as follows: by B'_{host_1}The central pixel point of (i, j) Pixel value is denoted as G_t(i, j), by B'_{host_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point Row vector is constituted, and is denoted as G_{host_1,i,j}, by B'_{host_1}The predicted pixel values of the central pixel point of (i, j) are denoted as G_p(i,j)；It will B'_{host_1}It is defined as watermark embedded block after (i, j) insertion watermark information, and is denoted as B^w _{host_1}(i, j), by B^w _{host_1}In (i, j) The pixel value of imago vegetarian refreshments is denoted asWhen watermark information to be embedded is 1, if | G_t(i, j) | > | G_p(i,j)|× (1+T) is then enabledIf | G_t(i,j)|≤|G_p(i, j) | × (1+T) is then enabledWhen watermark information to be embedded is 0, if | G_t(i, j) | < | G_p (i, j) | × (1-T) is then enabledIf | G_t(i,j)|≥|G_p(i, j) | × (1-T) is then enabledWherein, G_{host_1,i,j}Dimension be 1 × (N_block×N_block- 1), G_p(i, j) is to utilize local correlations model Γ and G_{host_1,i,j}It is calculated,M is positive integer, and the initial value of m is 1,1≤m≤N_block×N_block- 1, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) element element value, the dimension of Γ is (N_block×N_block- 1) it × 1, being saved Γ as code key Key2, Γ (m, 1) indicates to be designated as the AR coefficient of (m, 1) under in Γ, symbol " | | " be Take absolute value oeprator, and T is watermark embedment strength,In "=" be assignment, sign () indicate Take the function of symbol；

Step 1_7: to B'_{host_1}In all not process blocks and all watermark embedded blocks carry out image merged block, reconstruct To the image containing watermark information, it is denoted as B'_{host_1_w}；Then according to the inverse process of step 1_3, to B'_{host_1_w}It is handled, Make the width M of the image obtained after processing_hostAnd height is N_host, the image obtained after processing is denoted as B_{host_1_w}, and make For the fisrt feature figure containing watermark information；Recycle the inverse transformation of Tucker3 decomposition algorithm to B_{host_1_w}、B_{host_2}And B_{host_3} It is handled, obtains watermarking images, be denoted as I_w；Wherein, I_wWidth be M_hostAnd height is N_host；

Step 1_8: watermark built-in end sends code key Key1 and code key Key2 and gives watermark extracting end；

The specific steps of the watermark extracting part are as follows:

Step 2_1: the watermarking images containing watermark information are read, I' is denoted as_w；Wherein, I'_wFor RGB color image, I'_w's Width is M_hostAnd height is N_host, I'_wTo be not affected by the watermarking images of any attack or being the water after tone mapping is attacked Watermark image；

Step 2_2: by I'_wIt is expressed as the form of three rank tensors, is denoted as A_w；Then using Tucker3 decomposition algorithm to A_wInto Row tensor resolution obtains A_wCore tensor, be denoted as B_w；And by B_wThe 1st channel as I'_wFisrt feature figure, be denoted as B_{w_1}；By B_wThe 2nd channel as I'_wSecond feature image, be denoted as B_{w_2}；By B_wThe 3rd channel as I'_wThird Characteristic image is denoted as B_{w_3}；Wherein, A_wAnd B_wSize be M_host×N_host× 3, B_{w_1}、B_{w_2}And B_{w_3}Width be M_host And height is N_host；

Step 2_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Without the filling of pixel, and will B_{w_1}Again correspondence is denoted as B'_{w_1}；

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Downside carry out pixel and fill out It fills, fills N_block-mod(N_host,N_block) row, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Right side carry out pixel and fill out It fills, fills N_block-mod(M_host,N_block) column, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Right side and downside carry out N is filled in the filling of pixel_block-mod(M_host,N_block) column and N_block-mod(N_host,N_block) row, and by B_{w_1}Through pixel The image obtained after the filling of point is denoted as B' again_{w_1}；

It is above-mentioned, N_blockFor odd number, N_block> 1, B'_{w_1}Width be M'_hostAnd height is N'_host,SymbolFor the oeprator that rounds up, mod () table Show modulo operation symbol；

Step 2_4: by B'_{w_1}As watermark extracting carrier, by B'_{w_1}It is divided intoA ruler not overlapped Very little size is N_block×N_blockImage block, will be to B'_{w_1}Coordinate position is (i, j) in all image blocks obtained after division Image block is denoted as B'_{w_1}(i,j)；Wherein,

Step 2_5: B' is judged according to the pixel value of each pixel in code key Key1_{w_1}In same coordinate position Whether contain watermark information in image block, for B'_{w_1}(i, j), if coordinate position is the pixel of (i, j) in code key Key1 Pixel value is 0 and determines B'_{w_1}Watermark information is not contained in (i, j)；If coordinate position is the pixel of (i, j) in code key Key1 Pixel value be 1 determine B'_{w_1}Contain watermark information in (i, j), and by B'_{w_1}(i, j) is defined as to be extracted piece of watermark；

Step 2_6: from B'_{w_1}In to be extracted piece of each watermark in extract watermark information, if B'_{w_1}(i, j) waits for for watermark Block is extracted, then from B'_{w_1}The process of watermark information is extracted in (i, j) are as follows: by B'_{w_1}The pixel value of the central pixel point of (i, j) is remembered For G_t' (i, j), by B'_{w_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point constitute row to Amount, and it is denoted as G'_{w_1,i,j}；Then code key Key2 and G' are utilized_{w_1,i,j}Calculate B'_{w_1}The prediction pixel of the central pixel point of (i, j) Value, is denoted as G_p' (i, j),Further according to G_t' (i, j) and G_p' (i, j) from B'_{w_1}Watermark information is extracted in (i, j), if | G_t' (i, j) | > | G_p' (i, j) | setting up the watermark information then extracted is 1, if | G_t'(i,j)|≤|G_p' (i, j) | setting up the watermark information then extracted is 0；Wherein, G'_{w_1,i,j}Dimension be 1 × (N_block× N_block- 1), G'_{w_1,i,j}(1, m) G' is indicated_{w_1,i,j}In under be designated as (1, m) element element value.

In the step 1_3, to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective downside carries out the filling of pixel Process are as follows: by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}As processing image；Each of the last line of image will be handled N of the pixel value of pixel as filling_block-mod(N_host,N_block) pixel of the pixel of same row in every row in row Value；

To B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective right side carries out the process of the filling of pixel are as follows: by B_{host_1}、 I_{host_r}、I_{host_g}And I_{host_b}As processing image；Using the pixel value of each pixel in last column for handling image as The N of filling_block-mod(M_host,N_block) in each column in column with the pixel value of the pixel of a line.

In the step 1_5, Mask_lumAcquisition process are as follows:

Step 1_5a: I' is calculated_{host_r}Amendment areflexia figure, be denoted as MSF_r, by MSF_rMiddle coordinate position is the picture of (x, y) The pixel value of vegetarian refreshments is denoted as MSF_r(x, y),Equally, it calculates I'_{host_g}Amendment areflexia figure, be denoted as MSF_g, by MSF_gMiddle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_g (x, y),Calculate I'_{host_b}Amendment areflexia figure, be denoted as MSF_b, By MSF_bMiddle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_b(x, y),Wherein, 1≤x≤M'_host, 1≤y≤N'_host, I'_{host_r}(x,y) Indicate I'_{host_r}Middle coordinate position is the pixel value of the pixel of (x, y), I'_min(x, y)=min (I'_{host_r}(x,y),I'_{host_g} (x,y),I'_{host_b}(x, y)), min () is to be minimized function, I'_{host_g}(x, y) indicates I'_{host_g}Middle coordinate position be (x, Y) pixel value of pixel, I'_{host_b}(x, y) indicates I'_{host_b}Middle coordinate position is the pixel value of the pixel of (x, y),

Step 1_5b: I' is calculated_hostWith amendment areflexia figure between differential chart, be denoted as d, by coordinate position in d be (x, Y) pixel value of pixel is denoted as d (x, y),

Step 1_5c: binary conversion treatment is carried out to d, the image obtained after binary conversion treatment is denoted as d_otsu, by d_otsuMiddle seat The pixel value that mark is set to the pixel of (x, y) is denoted as d_otsu(x, y),Wherein, th For by OTSU method to I'_hostThe threshold value handled；

Step 1_5d: by d_otsuIt is divided intoA size not overlapped is N_block×N_blockFigure It, will be to d as block_otsuCoordinate position is that the image block of (i, j) is denoted as d in all image blocks obtained after division_otsu(i,j)；Then Obtain d_otsuWidth beAnd height isDown-sampled images, be denoted as d_{otsu_sub}, by d_{otsu_sub}Middle coordinate position Pixel value for the pixel of (i, j) is denoted as d_{otsu_sub}(i, j), if d_otsuThe pixel value of all pixels point in (i, j) it is equal Value is 1, then enables d_{otsu_sub}(i, j)=1；If d_otsuThe mean value of the pixel value of all pixels point in (i, j) is not 1, then enables d_{otsu_sub}(i, j)=0；

Step 1_5e: to d_{otsu_sub}Carry out morphology opening and closing processing, using obtained image after morphology opening and closing processing as Mask_lum。

In the step 1_6, G_pThe acquisition process of (i, j) are as follows:

Step 1_6a1: in B'_{host_1}The upper side and lower side fill respectively k row pixel value be 0 pixel, in B'_{host_1}'s Left and right side fills the pixel that k column pixel value is 0 respectively, and the image obtained after filling is denoted as B "_{host_1}；Wherein, k is Positive integer,B”_{host_1}Width be M'_host+ 2k and height are N'_host+2k；

Step 1_6a2: being N with size_block×N_blockSquare be sliding window, exist by step-length for 1 pixel B”_{host_1}Middle sliding, by B "_{host_1}It is divided into M'_host×N'_hostThe size of a overlapping is N_block×N_blockImage block, By B "_{host_1}In s-th of image block be denoted as B "_{host_1_s}；Wherein, s is positive integer, and the initial value of s is 1,1≤s≤M'_host× N'_host；

Step 1_6a3: Autoregressive Prediction Method is utilized, B " is established_{host_1}In each image block central pixel point picture Element value and the N centered on the pixel_block×N_blockOffice between the pixel value of all vicinity points in contiguous range Portion's correlation, for B "_{host_1_s}, by B "_{host_1_s}The pixel value of central pixel point be denoted as G_s, by B "_{host_1_s}In remove its center The pixel value arranged in sequence of all pixels point outside pixel constitutes row vector, and is denoted as G_non,s；By G_sWith with B "_{host_1_s}In N centered on imago vegetarian refreshments_block×N_blockLocal correlations between the pixel value of all vicinity points in contiguous range Description are as follows:Wherein, G_non,sDimension be 1 × (N_block×N_block- 1), m For positive integer, the initial value of m is 1,1≤m≤N_block×N_block- 1, G_non,s(1, m) G is indicated_non,sIn under be designated as the member of (1, m) The element value of element, Γ (m, 1) indicate that the AR coefficient that (m, 1) is designated as under in Γ, Γ (m, 1) reflect G_sWith G_non,sBetween (1, m) Correlation, ε_sIndicate B "_{host_1_s}Corresponding error term, ε_sValue close to 0；

Step 1_6a4: by B "_{host_1}In all image blocks central pixel point pixel value arranged in sequence constitute column to Amount, is denoted as G_{host_1_t}；By B "_{host_1}In all image blocks respectively in all pixels point in addition to its central pixel point pixel The row vector that value arranged in sequence is constituted is arranged successively composition matrix, and is denoted as G_{host_1},It will B”_{host_1}In all image blocks central pixel point predicted pixel values arranged in sequence constitute column vector, be denoted as G_{host_1_p}；Its In, G_{host_1_t}Dimension be (M'_host×N'_host) × 1, G_{host_1}Dimension be (M'_host×N'_host)×(N_block×N_block- 1), G_nos,1Indicate B "_{host_1}In the 1st image block B "_{host_1_1}In all pixels point in addition to its central pixel point pixel It is worth the row vector that arranged in sequence is constituted, G_non,sIndicate B "_{host_1}In s-th of image block B "_{host_1_s}In remove its central pixel point The row vector that the pixel value arranged in sequence of outer all pixels point is constituted,Indicate B "_{host_1}In M'_host× N'_hostA image blockIn the pixel value arranged in sequence of all pixels point in addition to its central pixel point constitute Row vector, G_{host_1_p}Dimension be (M'_host×N'_host)×1；

Step 1_6a5: to make G_{host_1_p}With G_{host_1_t}Difference minimize, gone out using Least Square Method optimal Model parameter is denoted as Γ, Γ=((G_{host_1})^T×G_{host_1})^-1×(G_{host_1})^T×G_{host_1_t}, and using Γ as local correlations Model；Wherein, the dimension of Γ is (N_block×N_block- 1) × 1, (G_{host_1})^TFor G_{host_1}Transposition, ((G_{host_1})^T× G_{host_1}) -1 be (G_{host_1})^T×G_{host_1}It is inverse；

Step 1_6a6: according to Γ and G_{host_1,i,j}Calculate G_p(i, j),Wherein, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) Element element value.

In the step 1_6, the determination process of T are as follows:

Step 1_6b1: setting 0.5 for the initial value of T, and the decrement step size of T is set as 0.01；Then according to step 1_1 To the process of step 1_7, I when acquisition takes each T value_hostCorresponding high dynamic range watermarking images obtain 50 panel heights dynamic altogether The set that 50 panel height dynamic range watermarking images are constituted is denoted as { I by range watermarking images_w1,I_w2,…,I_w50}；Wherein, I_w1Table Show I when taking T value is 0.5_hostCorresponding high dynamic range watermarking images, I_w2Expression take T value be 0.49 when I_hostIt is corresponding high dynamic State range watermarking images, I_w50Expression take T value be 0.01 when I_hostCorresponding high dynamic range watermarking images；

Step 1_6b2: { I is calculated_w1,I_w2,…,I_w50In every panel height dynamic range watermarking images invisibility index And the extraction bit error rate of watermark information after 5 kinds of different tone mapping attacks is fought, by { I_w1,I_w2,…,I_w50In q The extraction error code of watermark information after the invisibility index and confrontation α kind tone mapping attack of panel height dynamic range watermarking images Rate correspondence is denoted as VDP_qAnd BER_q,α(%)；Wherein, q and α is positive integer, and the initial value of q and α are 1,1≤q≤50,1≤α ≤ 5, VDP_qAnd BER_q,αThe value range of (%) is [0,100]；

Step 1_6b3: f is enabled_max=max (f₁,f₂,…,f₅₀)；Then by f_maxEnd value of the corresponding T value as T；Its In, f_max、f₁、f₂、f₅₀It is the intermediate variable of introducing, max () is to be maximized function, VDP₁Indicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermark The invisibility index of image, VDP₂Indicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images it is invisible Property index, VDP₅₀Indicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic range watermarking images invisibility index, BER_1,αIndicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermarking images confrontation α kind tone mapping attack after water The extraction bit error rate of official seal breath, BER_2,αIndicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images fight α The extraction bit error rate of watermark information, BER after kind tone mapping attack_50,αIndicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic Range watermarking images fight the extraction bit error rate of watermark information after the attack of α kind tone mapping.

Compared with the prior art, the advantages of the present invention are as follows:

1) the method for the present invention can not only completely indicate high dimensional data, but also can in view of decomposing using Tucker To efficiently solve the characteristics of internal structure is destroyed in traditional data processing method, the method for the present invention moves the height of width colour State range host image is expressed as the form of three rank tensors, then using Tucker3 decompose to high dynamic range host image into Row processing, and using the fisrt feature figure of obtained core tensor as the insertion carrier of watermark information, it is not only because core The fisrt feature figure of tensor covers the main energetic of a panel height dynamic range host image, but also can make the watermark of insertion Information is diffused into tri- channels R, G, B of high dynamic range watermarking images with the inverse transformation of decomposition, to effectively improve The robustness of the method for the present invention.

2) the method for the present invention establishes local correlations mould using AR prediction technique on the fisrt feature figure of core tensor Type, and transmitted using the local correlations model as code key, this is not only efficiently solved because traditional prediction method is predicted The problem of inaccurate and bring high dynamic range watermarking images quality declines, improves invisibility, and also improve this The safety of inventive method.

3) the method for the present invention is directed to the nonlinear characteristic of tone mapping attack, with reference to overexposure region in low dynamic range echograms The extracting method in domain, devises the brightness exposure mask an of low complex degree, to advanced optimize the robustness of the method for the present invention.

4) the method for the present invention is for balanced invisibility and robustness, in order to be directed to different high dynamic range host images Embedment strength is adaptive selected, proposes a simple watermark embedment strength calculative strategy.

5) the method for the present invention does not need original high dynamic range host image in watermark information extraction process, realizes The Blind extracting of watermark information.

6) the method for the present invention is tested, obtains embedding capacity height, and the bit error rate is low.

Detailed description of the invention

Fig. 1 is that the overall of the watermark embedding section of the method for the present invention realizes block diagram；

Fig. 2 is that the overall of the watermark extracting part of the method for the present invention realizes block diagram；

Fig. 3 is the evaluation result of the 30 width watermarking images that are obtained using the method for the present invention about invisibility；

Fig. 4 is the watermark extracting error code of the 30 width watermarking images that are obtained using the method for the present invention when no tone mapping is attacked The average watermark extracting bit error rate (%) when 27 kinds of tone mappings attacks of rate (%) and respectively confrontation；

Fig. 5 is that 27 kinds of tone mapping attacks are respectively acting on the water after the 30 width watermarking images obtained using the method for the present invention Print extracts the mean value (%) of the bit error rate.

Specific embodiment

The present invention will be described in further detail below with reference to the embodiments of the drawings.

A kind of high dynamic range images water mark method decomposed based on Tucker proposed by the present invention comprising watermark insertion With watermark extracting two parts.

As shown in Figure 1, the specific steps of the watermark embedding section are as follows:

Step 1_1: I is enabled_hostIndicate the high dynamic range host image of watermark information to be embedded, I_hostFor RGB color figure Picture, by I_hostR Color Channel, G Color Channel and B color channel correspondence be denoted as I_{host_r}、I_{host_g}And I_{host_b}；Wherein, I_host、 I_{host_r}、I_{host_g}And I_{host_b}Width be M_hostAnd height is N_host。

Step 1_2: by I_hostIt is expressed as the form of three rank tensors, is denoted as A_host；Then it is decomposed using existing Tucker3 Algorithm is to A_hostTensor resolution is carried out, A is obtained_hostCore tensor, be denoted as B_host；And by B_hostThe 1st channel as I_host Fisrt feature figure, be denoted as B_{host_1}；By B_hostThe 2nd channel as I_hostSecond feature image, be denoted as B_{host_2}；It will B_hostThe 3rd channel as I_hostThird feature image, be denoted as B_{host_3}；Wherein, A_hostAnd B_hostSize be M_host× N_host× 3, B_{host_1}、B_{host_2}And B_{host_3}Width be M_hostAnd height is N_host。

Step 1_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Not into The filling of row pixel, and by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Again correspondence is denoted as B'_{host_1}、I'_{host_r}、I'_{host_g}With I'_{host_b}。

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b} Respective downside carries out the filling of pixel, fills N_block-mod(N_host,N_block) row, and by B_{host_1}Filling through pixel The image obtained afterwards is denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}。

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b} Respective right side carries out the filling of pixel, fills N_block-mod(M_host,N_block) column, and by B_{host_1}Filling through pixel The image obtained afterwards is denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}。

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}With I_{host_b}Respective right side and downside carry out the filling of pixel, fill N_block-mod(M_host,N_block) column and N_block-mod (N_host,N_block) row, and by B_{host_1}The image obtained after the filling of pixel is denoted as B' again_{host_1}, by I_{host_r}Through pixel The image obtained after the filling of point is denoted as I' again_{host_r}, by I_{host_g}The image obtained after the filling of pixel is denoted as again I'_{host_g}, by I_{host_b}The image obtained after the filling of pixel is denoted as I' again_{host_b}。

It is above-mentioned, N_blockFor odd number, N_block> 1, N is taken in the present embodiment_block=5 pixels, B'_{host_1}、I'_{host_r}、 I'_{host_g}And I'_{host_b}Width be M'_hostAnd height is N'_host, SymbolFor the oeprator that rounds up, mod () indicates modulo operation symbol.

In this particular embodiment, in step 1_3, to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective downside carries out picture The process of the filling of vegetarian refreshments are as follows: by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}As processing image；By handle image last N of the pixel value of each pixel in row as filling_block-mod(N_host,N_block) pixel of same row in every row in row The pixel value of point, as last line full line replicate completely.To B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective right side carries out The process of the filling of pixel are as follows: by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}As processing image；The last of image will be handled N of the pixel value of each pixel in one column as filling_block-mod(M_host,N_block) in each column in column with the picture of a line The pixel value of vegetarian refreshments, as last column permutation replicate completely.

Step 1_4: by B'_{host_1}It is embedded in carrier as watermark, by B'_{host_1}It is divided intoIt is a not overlap Size be N_block×N_blockImage block, will be to B'_{host_1}Coordinate position is in all image blocks obtained after division The image block of (i, j) is denoted as B'_{host_1}(i,j)；Wherein,

Step 1_5: by I'_{host_r}、I'_{host_g}And I'_{host_b}To should be used as I'_hostR Color Channel, G Color Channel and B Color Channel；Then I' is obtained_hostWidth beAnd height isBrightness exposure mask, be denoted as Mask_lum, and will Mask_lumIt is saved as code key Key1；Then according to Mask_lumIn the pixel value of each pixel judge whether B'_{host_1}In same coordinate position image block in be embedded in watermark information, for B'_{host_1}(i, j), if Mask_lum(i, j)=0 Then determine in B'_{host_1}It is not embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as not process block；If Mask_lum(i, J)=1 judgement is in B'_{host_1}It is embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as to be embedded piece of watermark；Its In, I'_hostWidth be M'_hostAnd height is N'_host, Mask_lum(i, j) indicates Mask_lumMiddle coordinate position is the picture of (i, j) The pixel value of vegetarian refreshments.

In this particular embodiment, in step 1_5, Mask_lumAcquisition process are as follows:

Step 1_5a: I' is calculated_{host_r}Amendment areflexia (Modified Specular Free, MSF) figure, be denoted as MSF_r, by MSF_rMiddle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_r(x, y),Equally, I' is calculated_{host_g}Amendment areflexia figure, be denoted as MSF_g, By MSF_gMiddle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_g(x, y),Calculate I'_{host_b}Amendment areflexia figure, be denoted as MSF_b, by MSF_b Middle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_b(x, y),Wherein, 1≤x≤M'_host, 1≤y≤N'_host, I'_{host_r}(x,y) Indicate I'_{host_r}Middle coordinate position is the pixel value of the pixel of (x, y), I'_min(x, y)=min (I'_{host_r}(x,y),I'_{host_g} (x,y),I'_{host_b}(x, y)), min () is to be minimized function, I'_{host_g}(x, y) indicates I'_{host_g}Middle coordinate position be (x, Y) pixel value of pixel, I'_{host_b}(x, y) indicates I'_{host_b}Middle coordinate position is the pixel value of the pixel of (x, y),

Step 1_5b: I' is calculated_hostWith amendment areflexia figure between differential chart, be denoted as d, by coordinate position in d be (x, Y) pixel value of pixel is denoted as d (x, y),

Step 1_5c: binary conversion treatment is carried out to d, the image obtained after binary conversion treatment is denoted as d_otsu, by d_otsuMiddle seat The pixel value that mark is set to the pixel of (x, y) is denoted as d_otsu(x, y),Wherein, th For by OTSU (Nobuyuki Otsu) method to I'_hostThe threshold value handled.

Step 1_5d: by d_otsuIt is divided intoA size not overlapped is N_block×N_blockFigure It, will be to d as block_otsuCoordinate position is that the image block of (i, j) is denoted as d in all image blocks obtained after division_otsu(i,j)；Then Obtain d_otsuWidth beAnd height isDown-sampled images, be denoted as d_{otsu_sub}, by d_{otsu_sub}Middle coordinate position Pixel value for the pixel of (i, j) is denoted as d_{otsu_sub}(i, j), if d_otsuThe pixel value of all pixels point in (i, j) it is equal Value is 1, then enables d_{otsu_sub}(i, j)=1；If d_otsuThe mean value of the pixel value of all pixels point in (i, j) is not 1, then enables d_{otsu_sub}(i, j)=0.

Step 1_5e: to d_{otsu_sub}Carry out morphology opening and closing processing, using obtained image after morphology opening and closing processing as Mask_lum。

Step 1_6: in B'_{host_1}In to be embedded piece of each watermark in be embedded in watermark information, if B'_{host_1}(i, j) is water To be embedded piece is printed, then in B'_{host_1}The process of insertion watermark information in (i, j) are as follows: by B'_{host_1}The central pixel point of (i, j) Pixel value is denoted as G_t(i, j), by B'_{host_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point Row vector is constituted, and is denoted as G_{host_1,i,j}, by B'_{host_1}The predicted pixel values of the central pixel point of (i, j) are denoted as G_p(i,j)；It will B'_{host_1}It is defined as watermark embedded block after (i, j) insertion watermark information, and is denoted as B^w _{host_1}(i, j), by B^w _{host_1}In (i, j) The pixel value of imago vegetarian refreshments is denoted asWhen watermark information to be embedded is 1, if | G_t(i, j) | > | G_p(i,j)|× (1+T) is then enabledIf | G_t(i,j)|≤|G_p(i, j) | × (1+T) is then enabledWhen watermark information to be embedded is 0, if | G_t(i, j) | < | G_p (i, j) | × (1-T) is then enabledIf | G_t(i,j)|≥|G_p(i, j) | × (1-T) is then enabledWherein, G_{host_1,i,j}Dimension be 1 × (N_block×N_block- 1), G_p(i, j) is to utilize local correlations model Γ and G_{host_1,i,j}It is calculated,M is positive integer, and the initial value of m is 1,1≤m≤N_block×N_block- 1, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) element element value, the dimension of Γ is (N_block×N_block- 1) it × 1, being saved Γ as code key Key2, Γ (m, 1) indicates to be designated as the AR coefficient of (m, 1) under in Γ, symbol " | | " be Take absolute value oeprator, and T is watermark embedment strength, in the present embodiment for different high dynamic range host images T be it is adaptively selected according to corresponding high dynamic range host image,In "=" be assignment Symbol, sign () indicate to take the function of symbol.

In this particular embodiment, in step 1_6, G_pThe acquisition process of (i, j) are as follows:

Step 1_6a1: in B'_{host_1}The upper side and lower side fill respectively k row pixel value be 0 pixel, in B'_{host_1}'s Left and right side fills the pixel that k column pixel value is 0 respectively, and the image obtained after filling is denoted as B "_{host_1}；Wherein, k is Positive integer,B”_{host_1}Width be M'_host+ 2k and height are N'_host+2k。

Step 1_6a2: being N with size_block×N_blockSquare be sliding window, exist by step-length for 1 pixel B”_{host_1}Middle sliding, by B "_{host_1}It is divided into M'_host×N'_hostThe size of a overlapping is N_block×N_blockImage block, By B "_{host_1}In s-th of image block be denoted as B "_{host_1_s}；Wherein, s is positive integer, and the initial value of s is 1,1≤s≤M'_host× N'_host。

Step 1_6a3: existing autoregression (Auto-Regressive, AR) prediction technique is utilized, B " is established_{host_1}In The pixel value of the central pixel point of each image block and the N centered on the pixel_block×N_blockIt is all in contiguous range Local correlations between the pixel value of vicinity points, for B "_{host_1_s}, by B "_{host_1_s}Central pixel point pixel value It is denoted as G_s, by B "_{host_1_s}In the pixel value arranged in sequence of all pixels point in addition to its central pixel point constitute row vector, and remember For G_non,s；By G_sWith with B "_{host_1_s}Central pixel point centered on N_block×N_blockAll neighborhood pixels in contiguous range Local correlations description between the pixel value of point are as follows:Wherein, G_non,s's Dimension is 1 × (N_block×N_block- 1), m is positive integer, and the initial value of m is 1,1≤m≤N_block×N_block- 1, G_non,s(1,m) Indicate G_non,sIn under be designated as (1, m) element element value, Γ (m, 1) indicates to be designated as the AR coefficient of (m, 1) under in Γ, Γ (m, 1) G is reflected_sWith G_non,sCorrelation between (1, m), ε_sIndicate B "_{host_1_s}Corresponding error term, ε_sValue close to 0, such as Take ε_s=0.0000001.

Step 1_6a4: by B "_{host_1}In all image blocks central pixel point pixel value arranged in sequence constitute column to Amount, is denoted as G_{host_1_t}；By B "_{host_1}In all image blocks respectively in all pixels point in addition to its central pixel point pixel The row vector that value arranged in sequence is constituted is arranged successively composition matrix, and is denoted as G_{host_1},It will B”_{host_1}In all image blocks central pixel point predicted pixel values arranged in sequence constitute column vector, be denoted as G_{host_1_p}；Its In, G_{host_1_t}Dimension be (M'_host×N'_host) × 1, G_{host_1}Dimension be (M'_host×N'_host)×(N_block×N_block- 1), G_nos,1Indicate B "_{host_1}In the 1st image block B "_{host_1_1}In all pixels point in addition to its central pixel point pixel It is worth the row vector that arranged in sequence is constituted, G_non,sIndicate B "_{host_1}In s-th of image block B "_{host_1_s}In remove its central pixel point The row vector that the pixel value arranged in sequence of outer all pixels point is constituted,Indicate B "_{host_1}In M'_host× N'_hostA image blockIn the pixel value arranged in sequence of all pixels point in addition to its central pixel point constitute Row vector, G_{host_1_p}Dimension be (M'_host×N'_host)×1。

Step 1_6a5: to make G_{host_1_p}With G_{host_1_t}Difference minimize, gone out using existing Least Square Method Optimal model parameter is denoted as Γ, Γ=((G_{host_1})^T×G_{host_1})^-1×(G_{host_1})^T×G_{host_1_t}, and using Γ as part Correlation models；Wherein, the dimension of Γ is (N_block×N_block- 1) × 1, (G_{host_1})^TFor G_{host_1}Transposition, ((G_{host_1})^T× G_{host_1})^-1For (G_{host_1})^T×G_{host_1}It is inverse.

Step 1_6a6: according to Γ and G_{host_1,i,j}Calculate G_p(i, j),Wherein, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) Element element value.

In this particular embodiment, in step 1_6, the determination process of T are as follows:

Step 1_6b1: setting 0.5 for the initial value of T, and the decrement step size of T is set as 0.01；Then according to step 1_1 To the process of step 1_7, I when acquisition takes each T value_hostCorresponding high dynamic range watermarking images obtain 50 panel heights dynamic altogether The set that 50 panel height dynamic range watermarking images are constituted is denoted as { I by range watermarking images_w1,I_w2,…,I_w50}；Wherein, I_w1Table Show I when taking T value is 0.5_hostCorresponding high dynamic range watermarking images, I_w2Expression take T value be 0.49 when I_hostIt is corresponding high dynamic State range watermarking images, I_w50Expression take T value be 0.01 when I_hostCorresponding high dynamic range watermarking images.

Step 1_6b2: { I is calculated_w1,I_w2,…,I_w50In every panel height dynamic range watermarking images invisibility index And the extraction bit error rate of watermark information after 5 kinds of different tone mapping attacks is fought, by { I_w1,I_w2,…,I_w50In q The extraction error code of watermark information after the invisibility index and confrontation α kind tone mapping attack of panel height dynamic range watermarking images Rate correspondence is denoted as VDP_qAnd BER_q,α(%)；Wherein, q and α is positive integer, and the initial value of q and α are 1,1≤q≤50,1≤α ≤ 5, VDP_qAnd BER_q,αThe value range of (%) is [0,100]；5 kinds of different tone mapping attacks point in the present embodiment It Wei not DragoTMO, DurandTMO, FattalTMO, PattanaikVisualTMO and ReinhardDevlinTMO.

Step 1_6b3: f is enabled_max=max (f₁,f₂,…,f₅₀)；Then by f_maxEnd value of the corresponding T value as T；Its In, f_max、f₁、f₂、f₅₀It is the intermediate variable of introducing, max () is to be maximized function, VDP₁Indicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermark The invisibility index of image, VDP₂Indicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images it is invisible Property index, VDP₅₀Indicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic range watermarking images invisibility index, BER_1,αIndicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermarking images confrontation α kind tone mapping attack after water The extraction bit error rate of official seal breath, BER_2,αIndicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images fight α The extraction bit error rate of watermark information, BER after kind tone mapping attack_50,αIndicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic Range watermarking images fight the extraction bit error rate of watermark information after the attack of α kind tone mapping.

Step 1_7: to B'_{host_1}In all not process blocks and all watermark embedded blocks carry out image merged block, reconstruct To the image containing watermark information, it is denoted as B'_{host_1_w}；Then according to the inverse process of step 1_3, to B'_{host_1_w}It is handled, Make the width M of the image obtained after processing_hostAnd height is N_host, the image obtained after processing is denoted as B_{host_1_w}, and make For the fisrt feature figure containing watermark information；Recycle the inverse transformation of existing Tucker3 decomposition algorithm to B_{host_1_w}、B_{host_2} And B_{host_3}It is handled, obtains watermarking images, be denoted as I_w；Wherein, I_wWidth be M_hostAnd height is N_host。

Step 1_8: watermark built-in end sends code key Key1 and code key Key2 and gives watermark extracting end.

As shown in Fig. 2, the specific steps of the watermark extracting part are as follows:

Step 2_1: the watermarking images containing watermark information are read, I' is denoted as_w；Wherein, I'_wFor RGB color image, I'_w's Width is M_hostAnd height is N_host, I'_wTo be not affected by the watermarking images of any attack or being the water after tone mapping is attacked Watermark image.

Step 2_2: by I'_wIt is expressed as the form of three rank tensors, is denoted as A_w；Then existing Tucker3 decomposition algorithm is utilized To A_wTensor resolution is carried out, A is obtained_wCore tensor, be denoted as B_w；And by B_wThe 1st channel as I'_wFisrt feature figure, It is denoted as B_{w_1}；By B_wThe 2nd channel as I'_wSecond feature image, be denoted as B_{w_2}；By B_wThe 3rd channel as I'_w's Third feature image, is denoted as B_{w_3}；Wherein, A_wAnd B_wSize be M_host×N_host× 3, B_{w_1}、B_{w_2}And B_{w_3}Width be M_hostAnd height is N_host。

Step 2_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Without the filling of pixel, and will B_{w_1}Again correspondence is denoted as B'_{w_1}。

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Downside carry out pixel and fill out It fills, fills N_block-mod(N_host,N_block) row, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}。

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Right side carry out pixel and fill out It fills, fills N_block-mod(M_host,N_block) column, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}。

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Right side and downside carry out N is filled in the filling of pixel_block-mod(M_host,N_block) column and N_block-mod(N_host,N_block) row, and by B_{w_1}Through pixel The image obtained after the filling of point is denoted as B' again_{w_1}。

It is above-mentioned, N_blockFor odd number, N_block> 1, N is taken in the present embodiment_block=5 pixels, B'_{w_1}Width be M'_hostAnd height is N'_host,SymbolIt is upward Rounding operation symbol, mod () indicate modulo operation symbol.

Step 2_4: by B'_{w_1}As watermark extracting carrier, by B'_{w_1}It is divided intoA ruler not overlapped Very little size is N_block×N_blockImage block, will be to B'_{w_1}Coordinate position is (i, j) in all image blocks obtained after division Image block is denoted as B'_{w_1}(i,j)；Wherein,

Step 2_5: B' is judged according to the pixel value of each pixel in code key Key1_{w_1}In same coordinate position Whether contain watermark information in image block, for B'_{w_1}(i, j), if coordinate position is the pixel of (i, j) in code key Key1 Pixel value is 0 and determines B'_{w_1}Watermark information is not contained in (i, j)；If coordinate position is the pixel of (i, j) in code key Key1 Pixel value be 1 determine B'_{w_1}Contain watermark information in (i, j), and by B'_{w_1}(i, j) is defined as to be extracted piece of watermark.

Step 2_6: from B'_{w_1}In to be extracted piece of each watermark in extract watermark information, if B'_{w_1}(i, j) waits for for watermark Block is extracted, then from B'_{w_1}The process of watermark information is extracted in (i, j) are as follows: by B'_{w_1}The pixel value of the central pixel point of (i, j) is remembered For G_t' (i, j), by B'_{w_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point constitute row to Amount, and it is denoted as G'_{w_1,i,j}；Then code key Key2 and G' are utilized_{w_1,i,j}Calculate B'_{w_1}The prediction pixel of the central pixel point of (i, j) Value, is denoted as G_p' (i, j),Further according to G_t' (i, j) and G_p' (i, j) from B'_{w_1}Watermark information is extracted in (i, j), if | G_t' (i, j) | > | G_p' (i, j) | setting up the watermark information then extracted is 1, if | G_t'(i,j)|≤|G_p' (i, j) | setting up the watermark information then extracted is 0；Wherein, G'_{w_1,i,j}Dimension be 1 × (N_block× N_block- 1), G'_{w_1,i,j}(1, m) G' is indicated_{w_1,i,j}In under be designated as (1, m) element element value.

The feasibility and validity of method in order to further illustrate the present invention, is tested.

In the present embodiment, 30 panel height dynamic range host images and 27 kinds of tone mapping operators has been selected to be surveyed altogether Examination.Wherein, 30 panel height dynamic range host images are respectively from [1] Greg Ward website；[2] website Image Gallery； [3] TMQI database；27 kinds of tone mapping attacks are selected from HDR Toolbox for Matlab kit.Table 1 gives 30 The details of panel height dynamic range host image；Table 2 gives the details of 27 kinds of tone mappings attack.

The details of 1 30 panel height dynamic range host image of table

The details of 2 27 kinds of tone mappings (TM) of table attack

Using the method for the present invention 30 panel height dynamic range host images are executed with the specific steps of watermark embedding section respectively And the specific steps of watermark extracting part.Here, using existing signal-to-noise ratio (Signal to Noise Ratio, SNR), Structural similarity (Structural Similarity Index Measurement, SSIM) and HDR-VDP-2.2 this three refer to It marks to assess the invisibility of the method for the present invention；And utilize the bit error rate (Bit Error Rate, BER) Lai Hengliang the method for the present invention Robustness；The watermark embedding capacity of the method for the present invention is indicated by embedded quantity (bit) and insertion rate (bpp).Wherein, except mistake Other than code rate, other several evaluation indexes are that the higher performance for representing water mark method of numerical value is better, and the bit error rate is then numerical value It is smaller that represent robustness higher.

The embedding capacity of the method for the present invention depends primarily on the size N of the image block in image processing process_block(in this reality Apply N in example_block5) and brightness exposure mask Mask it takes_lumThe distribution proportion of middle pixel value 0 and pixel value 1.What table 3 gave The embedding capacity of 30 width watermarking images, the average embedded quantity of this 30 width watermarking images are up to 47644bit, and averagely insertion rate is about 0.0290bpp, hence it is evident that be higher than existing most of high dynamic range images multipurpose watermarking methods.In addition, the method for the present invention is same Sample is effective to other image block sizes, and with image block size N_blockIncrease, embedding capacity can show becoming for decline Gesture, and invisibility then shows as being gradually increased.Table 4, which gives the high dynamic range host image that picture number is " 29 ", is scheming As block size N_blockEmbedded quantity (bit), insertion rate (bpp), SNR, SSIM, VDP when taking 3,5 and 7 respectively are attacked without tone mapping BER value (being indicated with BER0 (%)) when hitting and the average BER value in the case where 27 kinds of tone mappings are attacked (use BER1 (%) table Show).

The embedding capacity of 3 30 width watermarking images of table

4 picture number of table is correlation data of the high dynamic range host image of " 29 " under different images block size

For the invisibility and robustness of the method for the present invention, Fig. 3 gives 30 width watermarking images about invisibility Evaluation result, Fig. 4 give the watermark extracting bit error rate (%) and difference of the 30 width watermarking images when no tone mapping is attacked The average watermark extracting bit error rate (%) when 27 kinds of tone mapping attacks is fought, Fig. 5 gives every kind of tone mapping attack difference The mean value of the watermark extracting bit error rate after acting on 30 width watermarking images shares 27 kinds of attack types.Due to 30 selected panel heights Dynamic range host image includes different scene information and indoor and outdoor surroundings, therefore by observation Fig. 3, Fig. 4 and Fig. 5 Data can be found that different high dynamic range host images can show different robustness and invisibility, however this 30 width The average SNR of watermarking images is about 53.2654dB, average SSIM is 1.0000, average VDP may be up to 91.5311, and in nothing The average error rate of 30 panel height dynamic range host images only has 0.44% when tone mapping is attacked, and is fighting 27 kinds of colors respectively Average error rate when mapping attack is adjusted also to only have 7.13%, overall the method for the present invention shows higher robust Property and preferable invisibility.

In order to further prove the superiority of the method for the present invention, comprehensively considered practicability, the validity of existing method with And whether extraction process is the factors such as blind Detecting, " the High dynamic range proposed respectively with Guerrini et al. Image watermarking robust against tone-mapping the operators " (robust of confrontation tone mapping Property high dynamic range images watermarking algorithms), " the High-capacity watermarking of that proposes of Maiorana et al. What high dynamic range images " (the large capacity watermarking algorithms of high dynamic range images) and Bai et al. was proposed “Towards a tone mapping-robust watermarking algorithm for high dynamic range Image based on spatial activity " (the high dynamic range images Robust Digital Watermarking Algorithm based on spatial activity) It is compared in terms of the robustness of confrontation tone mapping attack and embedding capacity two, and guarantees to verify the method for the present invention institute The test image and attack type of choosing are consistent with selected by above-mentioned three kinds of existing methods, the method for the present invention and Guerrini etc. People, Bai et al. about the correlation data of embedded quantity (bit) and BER (%) as listed in table 5, the method for the present invention and Maiorana etc. People, Bai et al. are about the correlation data of embedded quantity (bit) and BER (%) as listed by table 6.

The correlation data of 5 the method for the present invention of table and Guerrini et al., Bai et al. about embedded quantity (bit) and BER (%)

The correlation data of 6 the method for the present invention of table and Maiorana et al., Bai et al. about embedded quantity (bit) and BER (%)

6 column datas of analytical table 5 and table, it can be seen that the embedding capacity of the method for the present invention be Guerrini et al., It is proposed 15 times or so of water mark method of Maiorana et al., and it is suitable with the water mark method that Bai et al. is proposed；About robust Property, although individual images show the bit error rate (%) when being attacked by tone mapping higher than existing method, most of colors Mapping is adjusted to attack the bit error rate (%) of lower watermark information extraction or significantly lower than Guerrini et al. and Maiorana et al. The water mark method of proposition.In addition, table 7 gives the method for the present invention and now there are three types of the comprehensive performance comparing results of water mark method.

The comparison of 7 water mark method comprehensive performance of table

The method of the present invention is to more high dynamic range hosts it can be seen from the comprehensive performance correlation data that table 7 is given Image and more tone mapping attack types are tested, and have higher embedding capacity and better robustness.

Claims (5)

1. a kind of high dynamic range images water mark method decomposed based on Tucker, it is characterised in that including watermark insertion and watermark Extract two parts；

The specific steps of the watermark embedding section are as follows:

Step 1_1: I is enabled_hostIndicate the high dynamic range host image of watermark information to be embedded, I_hostIt, will for RGB color image I_hostR Color Channel, G Color Channel and B color channel correspondence be denoted as I_{host_r}、I_{host_g}And I_{host_b}；Wherein, I_host、 I_{host_r}、I_{host_g}And I_{host_b}Width be M_hostAnd height is N_host；

Step 1_2: by I_hostIt is expressed as the form of three rank tensors, is denoted as A_host；Then using Tucker3 decomposition algorithm to A_host Tensor resolution is carried out, A is obtained_hostCore tensor, be denoted as B_host；And by B_hostThe 1st channel as I_hostFisrt feature Figure, is denoted as B_{host_1}；By B_hostThe 2nd channel as I_hostSecond feature image, be denoted as B_{host_2}；By B_hostThe 3rd it is logical Road is as I_hostThird feature image, be denoted as B_{host_3}；Wherein, A_hostAnd B_hostSize be M_host×N_host× 3, B_{host_1}、B_{host_2}And B_{host_3}Width be M_hostAnd height is N_host；

Step 1_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Without picture The filling of vegetarian refreshments, and by B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Again correspondence is denoted as B'_{host_1}、I'_{host_r}、I'_{host_g}With I'_{host_b}；

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respectively Downside carry out pixel filling, fill N_block-mod(N_host,N_block) row, and by B_{host_1}After the filling of pixel To image be denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respectively Right side carry out pixel filling, fill N_block-mod(M_host,N_block) column, and by B_{host_1}After the filling of pixel To image be denoted as B' again_{host_1}, by I_{host_r}The image obtained after the filling of pixel is denoted as I' again_{host_r}, will I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, by I_{host_b}The figure obtained after the filling of pixel As being denoted as I' again_{host_b}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respectively From right side and downside carry out the filling of pixel, fill N_block-mod(M_host,N_block) column and N_block-mod(N_host, N_block) row, and by B_{host_1}The image obtained after the filling of pixel is denoted as B' again_{host_1}, by I_{host_r}Filling out through pixel The image obtained after filling is denoted as I' again_{host_r}, by I_{host_g}The image obtained after the filling of pixel is denoted as I' again_{host_g}, By I_{host_b}The image obtained after the filling of pixel is denoted as I' again_{host_b}；

It is above-mentioned, N_blockFor odd number, N_block> 1, B'_{host_1}、I'_{host_r}、I'_{host_g}And I'_{host_b}Width be M'_hostAnd height It is N'_host,SymbolFor the operator that rounds up Number, mod () indicates modulo operation symbol；

Step 1_4: by B'_{host_1}It is embedded in carrier as watermark, by B'_{host_1}It is divided intoA ruler not overlapped Very little size is N_block×N_blockImage block, will be to B'_{host_1}Coordinate position is (i, j) in all image blocks obtained after division Image block be denoted as B'_{host_1}(i,j)；Wherein,

Step 1_5: by I'_{host_r}、I'_{host_g}And I'_{host_b}To should be used as I'_hostR Color Channel, G Color Channel and B color it is logical Road；Then I' is obtained_hostWidth beAnd height isBrightness exposure mask, be denoted as Mask_lum, and by Mask_lumMake It is saved for code key Key1；Then according to Mask_lumIn the pixel value of each pixel judge whether in B'_{host_1}In phase With watermark information is embedded in the image block of coordinate position, for B'_{host_1}(i, j), if Mask_lum(i, j)=0 item determines B'_{host_1}It is not embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as not process block；If Mask_lum(i, j)=1 item Determine in B'_{host_1}It is embedded in watermark information in (i, j), and by B'_{host_1}(i, j) is defined as to be embedded piece of watermark；Wherein, I'_host Width be M'_hostAnd height is N'_host, Mask_lum(i, j) indicates Mask_lumMiddle coordinate position is the picture of the pixel of (i, j) Element value；

Step 1_6: in B'_{host_1}In to be embedded piece of each watermark in be embedded in watermark information, if B'_{host_1}(i, j) waits for for watermark Embedded block, then in B'_{host_1}The process of insertion watermark information in (i, j) are as follows: by B'_{host_1}The pixel of the central pixel point of (i, j) Value is denoted as G_t(i, j), by B'_{host_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point is constituted Row vector, and it is denoted as G_{host_1,i,j}, by B'_{host_1}The predicted pixel values of the central pixel point of (i, j) are denoted as G_p(i,j)；It will B'_{host_1}It is defined as watermark embedded block after (i, j) insertion watermark information, and is denoted as B^w _{host_1}(i, j), by B^w _{host_1}In (i, j) The pixel value of imago vegetarian refreshments is denoted asWhen watermark information to be embedded is 1, if | G_t(i, j) | > | G_p(i,j)|× (1+T) then enables G_t ^w(i, j)=G_t(i, j), if | G_t(i,j)|≤|G_p(i, j) | × (1+T) is then enabledWhen watermark information to be embedded is 0, if | G_t(i, j) | < | G_p (i, j) | × (1-T) is then enabledIf | G_t(i,j)|≥|G_p(i, j) | × (1-T) is then enabledWherein, G_{host_1,i,j}Dimension be 1 × (N_block×N_block- 1), G_p(i, j) is to utilize local correlations model Γ and G_{host_1,i,j}It is calculated,M is positive integer, and the initial value of m is 1,1≤m≤N_block×N_block- 1, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) element element value, the dimension of Γ is (N_block×N_block- 1) it × 1, being saved Γ as code key Key2, Γ (m, 1) indicates to be designated as the AR coefficient of (m, 1) under in Γ, symbol " | | " be Take absolute value oeprator, and T is watermark embedment strength,In "=" be assignment, sign () indicate Take the function of symbol；

Step 1_7: to B'_{host_1}In all not process blocks and all watermark embedded blocks carry out image merged block, reconstruct contained There is the image of watermark information, is denoted as B'_{host_1_w}；Then according to the inverse process of step 1_3, to B'_{host_1_w}It is handled, makes to locate The width of the image obtained after reason is M_hostAnd height is N_host, the image obtained after processing is denoted as B_{host_1_w}, and as containing There is the fisrt feature figure of watermark information；Recycle the inverse transformation of Tucker3 decomposition algorithm to B_{host_1_w}、B_{host_2}And B_{host_3}It carries out Processing, obtains watermarking images, is denoted as I_w；Wherein, I_wWidth be M_hostAnd height is N_host；

Step 1_8: watermark built-in end sends code key Key1 and code key Key2 and gives watermark extracting end；

The specific steps of the watermark extracting part are as follows:

Step 2_1: the watermarking images containing watermark information are read, I' is denoted as_w；Wherein, I'_wFor RGB color image, I'_wWidth For M_hostAnd height is N_host, I'_wTo be not affected by the watermarking images of any attack or being the watermark figure after tone mapping is attacked Picture；

Step 2_2: by I'_wIt is expressed as the form of three rank tensors, is denoted as A_w；Then using Tucker3 decomposition algorithm to A_wIt is opened Amount is decomposed, and A is obtained_wCore tensor, be denoted as B_w；And by B_wThe 1st channel as I'_wFisrt feature figure, be denoted as B_{w_1}；It will B_wThe 2nd channel as I'_wSecond feature image, be denoted as B_{w_2}；By B_wThe 3rd channel as I'_wThird feature figure Picture is denoted as B_{w_3}；Wherein, A_wAnd B_wSize be M_host×N_host× 3, B_{w_1}、B_{w_2}And B_{w_3}Width be M_hostAnd height It is N_host；

Step 2_3: if M_hostAnd N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Without the filling of pixel, and by B_{w_1}Again Correspondence is denoted as B'_{w_1}；

If M_hostIt can be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Downside carry out pixel filling, fill out Fill N_block-mod(N_host,N_block) row, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt can be by N_blockDivide exactly, then to B_{w_1}Right side carry out pixel filling, fill out Fill N_block-mod(M_host,N_block) column, and by B_{w_1}The image obtained after the filling of pixel is denoted as B' again_{w_1}；

If M_hostIt cannot be by N_blockDivide exactly and N_hostIt cannot be by N_blockDivide exactly, then to B_{w_1}Right side and downside carry out pixel N is filled in the filling of point_block-mod(M_host,N_block) column and N_block-mod(N_host,N_block) row, and by B_{w_1}Through pixel The image obtained after filling is denoted as B' again_{w_1}；

It is above-mentioned, N_blockFor odd number, N_block> 1, B'_{w_1}Width be M'_hostAnd height is N'_host,SymbolFor the oeprator that rounds up, mod () table Show modulo operation symbol；

Step 2_4: by B'_{w_1}As watermark extracting carrier, by B'_{w_1}It is divided intoA size not overlapped is big Small is N_block×N_blockImage block, will be to B'_{w_1}Coordinate position is the image of (i, j) in all image blocks obtained after division Block is denoted as B'_{w_1}(i,j)；Wherein,

Step 2_5: B' is judged according to the pixel value of each pixel in code key Key1_{w_1}In same coordinate position image Whether contain watermark information in block, for B'_{w_1}(i, j), if coordinate position is the pixel of the pixel of (i, j) in code key Key1 Value is 0 and determines B'_{w_1}Watermark information is not contained in (i, j)；If coordinate position is the picture of the pixel of (i, j) in code key Key1 Plain value is 1 and determines B'_{w_1}Contain watermark information in (i, j), and by B'_{w_1}(i, j) is defined as to be extracted piece of watermark；

Step 2_6: from B'_{w_1}In to be extracted piece of each watermark in extract watermark information, if B'_{w_1}(i, j) is that watermark is to be extracted Block, then from B'_{w_1}The process of watermark information is extracted in (i, j) are as follows: by B'_{w_1}The pixel value of the central pixel point of (i, j) is denoted as G_t' (i, j), by B'_{w_1}The pixel value arranged in sequence of all pixels point in (i, j) in addition to its central pixel point constitutes row vector, and It is denoted as G'_{w_1,i,j}；Then code key Key2 and G' are utilized_{w_1,i,j}Calculate B'_{w_1}The predicted pixel values of the central pixel point of (i, j), note For G_p' (i, j),Further according to G_t' (i, j) and G_p' (i, j) from B'_{w_1} Watermark information is extracted in (i, j), if | G_t' (i, j) | > | G_p' (i, j) | setting up the watermark information then extracted is 1, if | G_t' (i,j)|≤|G_p' (i, j) | setting up the watermark information then extracted is 0；Wherein, G'_{w_1,i,j}Dimension be 1 × (N_block× N_block- 1), G'_{w_1,i,j}(1, m) G' is indicated_{w_1,i,j}In under be designated as (1, m) element element value.

2. the high dynamic range images water mark method according to claim 1 decomposed based on Tucker, it is characterised in that institute In the step 1_3 stated, to B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective downside carries out the process of the filling of pixel are as follows: will B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}As processing image；By the pixel of each pixel in the last line for handling image It is worth the N as filling_block-mod(N_host,N_block) pixel value of the pixel of same row in every row in row；

To B_{host_1}、I_{host_r}、I_{host_g}And I_{host_b}Respective right side carries out the process of the filling of pixel are as follows: by B_{host_1}、 I_{host_r}、I_{host_g}And I_{host_b}As processing image；Using the pixel value of each pixel in last column for handling image as The N of filling_block-mod(M_host,N_block) in each column in column with the pixel value of the pixel of a line.

3. the high dynamic range images water mark method according to claim 1 or 2 decomposed based on Tucker, it is characterised in that In the step 1_5, Mask_lumAcquisition process are as follows:

Step 1_5a: I' is calculated_{host_r}Amendment areflexia figure, be denoted as MSF_r, by MSF_rMiddle coordinate position is the pixel of (x, y) Pixel value be denoted as MSF_r(x, y),Equally, I' is calculated_{host_g}'s Areflexia figure is corrected, MSF is denoted as_g, by MSF_gMiddle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_g(x, y),Calculate I'_{host_b}Amendment areflexia figure, be denoted as MSF_b, by MSF_b Middle coordinate position is that the pixel value of the pixel of (x, y) is denoted as MSF_b(x, y),Wherein, 1≤x≤M'_host, 1≤y≤N'_host, I'_{host_r}(x,y) Indicate I'_{host_r}Middle coordinate position is the pixel value of the pixel of (x, y), I'_min(x, y)=min (I'_{host_r}(x,y),I'_{host_g} (x,y),I'_{host_b}(x, y)), min () is to be minimized function, I'_{host_g}(x, y) indicates I'_{host_g}Middle coordinate position be (x, Y) pixel value of pixel, I'_{host_b}(x, y) indicates I'_{host_b}Middle coordinate position is the pixel value of the pixel of (x, y),

Step 1_5b: I' is calculated_hostDifferential chart between amendment areflexia figure, is denoted as d, is (x, y) by coordinate position in d The pixel value of pixel is denoted as d (x, y),

Step 1_5c: binary conversion treatment is carried out to d, the image obtained after binary conversion treatment is denoted as d_otsu, by d_otsuMiddle coordinate bit The pixel value for being set to the pixel of (x, y) is denoted as d_otsu(x, y),Wherein, th is logical OTSU method is crossed to I'_hostThe threshold value handled；

Step 1_5d: by d_otsuIt is divided intoA size not overlapped is N_block×N_blockImage Block, will be to d_otsuCoordinate position is that the image block of (i, j) is denoted as d in all image blocks obtained after division_otsu(i,j)；Then it obtains Take d_otsuWidth beAnd height isDown-sampled images, be denoted as d_{otsu_sub}, by d_{otsu_sub}Middle coordinate position is The pixel value of the pixel of (i, j) is denoted as d_{otsu_sub}(i, j), if d_otsuThe mean value of the pixel value of all pixels point in (i, j) It is 1, then enables d_{otsu_sub}(i, j)=1；If d_otsuThe mean value of the pixel value of all pixels point in (i, j) is not 1, then enables d_{otsu_sub}(i, j)=0；

Step 1_5e: to d_{otsu_sub}Carry out morphology opening and closing processing, using obtained image after morphology opening and closing processing as Mask_lum。

4. the high dynamic range images water mark method according to claim 1 decomposed based on Tucker, it is characterised in that institute In the step 1_6 stated, G_pThe acquisition process of (i, j) are as follows:

Step 1_6a1: in B'_{host_1}The upper side and lower side fill respectively k row pixel value be 0 pixel, in B'_{host_1}Left side It fills the pixel that k column pixel value is 0 respectively with right side, the image obtained after filling is denoted as B "_{host_1}；Wherein, k is positive whole Number,B”_{host_1}Width be M'_host+ 2k and height are N'_host+2k；

Step 1_6a2: being N with size_block×N_blockSquare be sliding window, exist by step-length for 1 pixel B”_{host_1}Middle sliding, by B "_{host_1}It is divided into M'_host×N'_hostThe size of a overlapping is N_block×N_blockImage block, By B "_{host_1}In s-th of image block be denoted as B "_{host_1_s}；Wherein, s is positive integer, and the initial value of s is 1,1≤s≤M'_host× N'_host；

Step 1_6a3: Autoregressive Prediction Method is utilized, B " is established_{host_1}In each image block central pixel point pixel value With the N centered on the pixel_block×N_blockLocal Phase between the pixel value of all vicinity points in contiguous range Guan Xing, for B "_{host_1_s}, by B "_{host_1_s}The pixel value of central pixel point be denoted as G_s, by B "_{host_1_s}In remove its center pixel The pixel value arranged in sequence of all pixels point outside point constitutes row vector, and is denoted as G_non,s；By G_sWith with B "_{host_1_s}Middle imago N centered on vegetarian refreshments_block×N_blockLocal correlations description between the pixel value of all vicinity points in contiguous range Are as follows:Wherein, G_non,sDimension be 1 × (N_block×N_block- 1), m is positive Integer, the initial value of m are 1,1≤m≤N_block×N_block- 1, G_non,s(1, m) G is indicated_non,sIn under be designated as the element of (1, m) Element value, Γ (m, 1) indicate that the AR coefficient that (m, 1) is designated as under in Γ, Γ (m, 1) reflect G_sWith G_non,sPhase between (1, m) Guan Xing, ε_sIndicate B "_{host_1_s}Corresponding error term, ε_sValue close to 0；

Step 1_6a4: by B "_{host_1}In all image blocks central pixel point pixel value arranged in sequence constitute column vector, note For G_{host_1_t}；By B "_{host_1}In all image blocks respectively in all pixels point in addition to its central pixel point pixel value sequentially The row vector being arranged to make up is arranged successively composition matrix, and is denoted as G_{host_1},By B "_{host_1}In All image blocks central pixel point predicted pixel values arranged in sequence constitute column vector, be denoted as G_{host_1_p}；Wherein, G_{host_1_t}Dimension be (M'_host×N'_host) × 1, G_{host_1}Dimension be (M'_host×N'_host)×(N_block×N_block- 1), G_nos,1Indicate B "_{host_1}In the 1st image block B "_{host_1_1}In the pixel value of all pixels point in addition to its central pixel point press The row vector that sequence is arranged to make up, G_non,sIndicate B "_{host_1}In s-th of image block B "_{host_1_s}In in addition to its central pixel point The row vector that the pixel value arranged in sequence of all pixels point is constituted,Indicate B "_{host_1}In M'_host×N'_host A image blockIn all pixels point in addition to its central pixel point the row that constitutes of pixel value arranged in sequence Vector, G_{host_1_p}Dimension be (M'_host×N'_host)×1；

Step 1_6a5: to make G_{host_1_p}With G_{host_1_t}Difference minimize, go out optimal model using Least Square Method and join Number, is denoted as Γ, Γ=((G_{host_1})^T×G_{host_1})^-1×(G_{host_1})^T×G_{host_1_t}, and using Γ as local correlations model； Wherein, the dimension of Γ is (N_block×N_block- 1) × 1, (G_{host_1})^TFor G_{host_1}Transposition, ((G_{host_1})^T×G_{host_1})^-1For (G_{host_1})^T×G_{host_1}It is inverse；

Step 1_6a6: according to Γ and G_{host_1,i,j}Calculate G_p(i, j), Wherein, G_{host_1,i,j}(1, m) G is indicated_{host_1,i,j}In under be designated as (1, m) element element value.

5. the high dynamic range images water mark method according to claim 1 or 4 decomposed based on Tucker, it is characterised in that In the step 1_6, the determination process of T are as follows:

Step 1_6b1: setting 0.5 for the initial value of T, and the decrement step size of T is set as 0.01；Then according to step 1_1 to step The process of rapid 1_7, I when acquisition takes each T value_hostCorresponding high dynamic range watermarking images obtain 50 panel height dynamic ranges altogether The set that 50 panel height dynamic range watermarking images are constituted is denoted as { I by watermarking images_w1,I_w2,…,I_w50}；Wherein, I_w1Expression takes T I when value is 0.5_hostCorresponding high dynamic range watermarking images, I_w2Expression take T value be 0.49 when I_hostCorresponding high dynamic range Watermarking images, I_w50Expression take T value be 0.01 when I_hostCorresponding high dynamic range watermarking images；

Step 1_6b2: { I is calculated_w1,I_w2,…,I_w50In every panel height dynamic range watermarking images invisibility index and The extraction bit error rate of watermark information after the different tone mapping attack of 5 kinds of confrontation, by { I_w1,I_w2,…,I_w50In q panel height The extraction bit error rate pair of watermark information after the invisibility index and confrontation α kind tone mapping attack of dynamic range watermarking images VDP should be denoted as_qAnd BER_q,α(%)；Wherein, q and α is positive integer, and the initial value of q and α are 1,1≤q≤50,1≤α≤5, VDP_qAnd BER_q,αThe value range of (%) is [0,100]；

Step 1_6b3: f is enabled_max=max (f₁,f₂,…,f₅₀)；Then by f_maxEnd value of the corresponding T value as T；Wherein, f_max、f₁、f₂、f₅₀It is the intermediate variable of introducing, max () is to be maximized function, VDP₁Indicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermark The invisibility index of image, VDP₂Indicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images it is invisible Property index, VDP₅₀Indicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic range watermarking images invisibility index, BER_1,αIndicate { I_w1,I_w2,…,I_w50In the 1st panel height dynamic range watermarking images confrontation α kind tone mapping attack after water The extraction bit error rate of official seal breath, BER_2,αIndicate { I_w1,I_w2,…,I_w50In the 2nd panel height dynamic range watermarking images fight α The extraction bit error rate of watermark information, BER after kind tone mapping attack_50,αIndicate { I_w1,I_w2,…,I_w50In the 50th panel height dynamic Range watermarking images fight the extraction bit error rate of watermark information after the attack of α kind tone mapping.