CN104021518A - Reversible data hiding method for high dynamic range image - Google Patents

Abstract

The invention discloses a reversible data hiding method for a high dynamic range image. The reversible data hiding method comprises a data embedding process and a data extracting process, wherein the data extracting process is inverse operation of the data embedding process, and the data embedding process comprises the following steps of converting a rgb image into an RGBE image, and dividing to-be-embedded data into three parts in sequence for embedding the three parts into rgb three channels corresponding to the RGBE image; dividing all pixels of a certain embedding channel of the rgb three channels into two types: S1 and S2, dividing the corresponding embedded data into two types corresponding to S1 and S2, wherein in an embedding process, one type of data is embedded to S1, and then the same process is repeated for embedding the other type of data to S2. The reversible data hiding method disclosed by the invention achieves a purpose of embedding more data by sacrificing least image quality. Compared with the existing reversible HDR (high dynamic range) image hiding algorithm, the hiding algorithm disclosed by the invention improves hiding capacity simultaneously when guaranteeing reversible hiding, and guarantees a better visual effect.

Description

The reversible data concealing method of high dynamic range images

Technical field

The present invention relates to the reversible data concealing method of high dynamic range images, being specifically related to the hiding carrier of data is high-dynamics image, and carrier transmits by overt channel (as network), reaches the hidden transmission of embedding information.

Background technology

Many users are by the simple and quick pass-along message of network.Yet malicious attacker can illegally be obtained transmission of information by network, this causes user to have to seek better information security transmission method.The same with cryptography method, Information hiding has reached the object of safe transfer information equally.Information hiding is embedded into key message not in the media file of interest (as image, text, audio frequency, video), makes assailant be difficult to determine the existence of embedding information.Conventionally, we become initial carrier file the media file that does not embed information, and the file embedding after key message is called embedding file.

Now, in numerous different carrier media files, image is commonly used to carry out data embedding.View data hidden algorithm is often by visual quality and the embedding capacity of embedding image are assessed.View data hidden algorithm should, the in the situation that of embedding image quality loss minimum, embed more data.If embedding image can entirely true recovery original image after extracting data, claim that this embedded mode is hiding reversible data.This mode can be used in the occasion of embedding and the information of embedding no less important, as medical science and military affairs.

In early days, Fridrich[1] etc. a kind of hiding reversible data algorithm based on Lossless Image Compression proposed.This algorithm is for we provide important thinking, and we can carry out data by redundant information variable in original and hide.Tian[2 afterwards] proposed a kind of by the poor method of carrying out information embedding of calculating pixel.Compared with work before, method-difference expansion of Tian has better utilized redundancy feature, can embed more information.Thodi and Rodriguez[3] by predicated error, expand (PEE) replacement difference expansion and utilize histogram transformation, improved the algorithm of Tian.Current most of algorithm is if [4-5] is mainly based on PEE and histogram transformation.

The research main object of the hiding reversible data algorithm of present stage is gray level image, and most of in daily life what use is coloured image.Research element body is in this respect [6-8] now, and they are mainly by utilizing mutual relationship between RGB triple channel further to improve hiding reversible data algorithm.But along with making rapid progress of camera work, a kind of coloured image form that can record real scene is accepted by people gradually, and the hiding reversible data algorithm to this new picture format, need to further study.

In recent years, people had great interest to high dynamic range (HDR) image.The dynamic range of scene is the ratio of its brightest part and the darkest part.Ferwerda etc. have summed up the vision parameter of brightness range in some true environments.Our vision system has very large color and strength range.For example, the moonlight ratio at the sunlight at high noon in summer and the night of full moon can reach 1,000 ten thousand.Be different from traditional low dynamic range echograms (LDR), HDR recording image the monochrome information of scene, make picture retain the luminance detail in real scene and dark portion details.For example the sun is bright more a lot of than bulb, but in LDR image, due to low dynamic range, they can be shown as white.But significantly different at the sun shown in HDR image table and bulb brightness, can well make a distinction.Fig. 1 (a)-Fig. 1 (b) shows the vision difference of LDR and HDR image.HDR image is clear has recorded the shade of leaf in vestige on bright snowfield and the woods.And on LDR image, a lot of details are lost, snowfield is too bright and tree shade is dim.Due to the advantage of HDR, it comes into vogue gradually in digital photography, computer graphics, electronic game, medical image, remote sensing image.

Yet we notice that most of hiding reversible data algorithm of present stage is mainly for LDR, although HDR expectation is used to substitute LDR and become new graphics standard.Only have a small part about the work of hiding reversible data directly about HDR.Cheng and Wand[9] algorithm propose to carry out data by edge and hide, although this algorithm has obtained larger embedded quantity, embedding information can not be recovered.Yu and Wang[10] for the HDR image storage format feature of brilliant RGBE form coding, a kind of lossless data hiding algorithm is proposed.This algorithm is after embedding data, and embedding image is identical with the demonstration of cover image, but this algorithm embedded quantity is generally 0.29 every pixel of 0.12 – (bpp).This algorithm is not suitable for a lot of scenes, because its embedded quantity is very little.

[1]J.Fridrich,M.Goljan,R.Du,Lossless data embedding—new paradigm in digital watermarking,EURASIP Journal on Applied Signal Processing2002(2)(2002)185–196.

[2]J.Tian,Reversible data embedding using a difference expansion,IEEE Transactions on Circuits and Systems for Video Technology13(8)(2003)890–896

[3]D.M.Thodi,J.J.Rodriguez,Expansion embedding techniques for reversible watermarking,IEEE Transactions on Image Processing16(3)(2007)721–730.

[4]L.Luo,Z.Chen,M.Chen,X.Zeng,and Z.Xiong,―Reversible image watermarking using interpolation technique,‖Information Forensics and Security,IEEE Transactions on,vol.5,no.1,pp.187–193,2010.

[5]V.Sachnev,H.J.Kim,J.Nam,S.Suresh,Y.Shi,Reversible watermarking algorithm using sorting and prediction,IEEE Transactions on Circuits and Systems for Video Technology19(2009)989–999.

[6]J.Li,X.Li,and B.Yang,―Reversible data hiding scheme for color image based on prediction-error expansion and cross-channel correlation,‖Signal Processing,vol.93,no.9,pp.2748–2758,2013.[7]C.-L.Tsai,K.-C.Fan,T.C.Chuang,C.-D.Chung,Lossless data hiding of color images using pixel decomposition and phase difference,Journal of Information Science and Engineering23(5)(2007)1481–1498

[8]H.Yang,K.Hwang,Reversible data hiding for color BMP image based on block difference histogram,in:Proceedings of the Fourth International Conference on Ubi-media Computing (U-Media),2011,pp.257–260.

[9]Y.-M.Cheng and C.-M.Wang,―A novel approach to steganography in high-dynamic-range images.,‖IEEE MultiMedia,vol.16,no.3,pp.70–80,2009.

80,2009.[10]C.-M.Yu,K.-C.Wu,and C.-M.Wang,―A distortion-free data hiding scheme for high dynamic range images,‖Displays,vol.32,no.5,pp.225–236,2011.

Summary of the invention

For solving the deficiency of prior art existence, the invention discloses a kind of reversible data concealing method of high dynamic range images, adopt the mode of predicated error expansion, the application is embedded into information in the HDR image of RGBE form.Our research finds that in RGBE passage, E can reflect image edge information.Jointing edge information, applies to HDR image the hiding reversible data algorithm based on PEE, has significantly improved pixel forecasting accuracy.In addition the pixel proposing after a kind of improvement, embeds ordering strategy.This algorithm can reach 2.2-2.85bpp for HDR image embedded quantity, and our algorithm belongs to blind Detecting, i.e. information extraction does not need original HDR image.Experimental result shows that this algorithm has reached the object of sacrificing minimum picture quality embedding more data.

For achieving the above object, concrete scheme of the present invention is as follows:

The reversible data concealing method of high dynamic range images, comprises the leaching process of data telescopiny and data, and the leaching process of data is the inverse operation of data telescopiny, and data embed and comprise the following steps:

Rgb image is converted to RGBE image, wish embedding data is divided into three parts according to the order of sequence, for embedding tri-passages of corresponding RGB of RGBE image;

All pixels of a certain embedding passage of tri-passages of RGB are divided into two classes: S ₁and S ₂, accordingly embedding data is divided into and S ₁and S ₂two corresponding classes, in telescopiny, embed S class data wherein ₁, then repeat identical process and embed another kind of data to S ₂in.

Data embed and specifically comprise the following steps:

Step 1: the last significance bit that is positioned at border and goes forward to set certain class pixel of number is recorded as to LSB, and is 0 last position of such pixel;

Step 2: E obtains marginal information according to reference channel, according to marginal information to certain class pixel { I ₁, I ₂... I _ncalculate parameters sortnig C, according to parameters sortnig C, certain class pixel is carried out to ascending order arrangement and obtain and the corresponding predicated error { P of each pixel after calculating sequence ₁, P ₂..., P _n;

Step 3: adopt the histogrammic mode of predicated error to obtain threshold value initial value (T _l, T _r), to pixel after sorting carry out successively data embedding, each pixel I _i(i=1,2 ...) data obtain the predicated error P ' of embedding data after embedding _iwith pixel value I ' _i, structure location map LM;

Step 4: remove the pixel of embedding data, detect residue sequence predicated error { P _e+1, P _e+2..., P _nin can be used to the quantity of embedding data, if lazy weight that can embedding data is to embed location map and LSB, reduce T _lor increase T _r, and get back to step 3; Otherwise, record site map length l and carry out next step;

Step 5: embed location map information and LSB, in this step, predicated error adopts { P _e+1, P _e+2..., P _n; T _l(5bits), T _r(5bits), l (18bits), e (20bits) becomes scale-of-two, finally it is become respectively to last position of a setting value pixel in step 1.

The step of constructing location map LM in described step 3 is:

(3-1) when the pixel value of embedding data a locating information LM (j)=1 is set, j=j+1, j is the position of locating information in location map, and this pixel is reduced to I _i, do not change this pixel value;

(3-2) as the pixel value I ' of embedding data _i∈ [0,255] still wherein, $P_i^{\&Prime;}=\left\{\begin{array}{cc}2P^{\&prime;}+b,& \mathrm{if}{P}^{\&prime;}\&Element;[T_l,T_r),\\ P^{\&prime;}+T_r,& \mathrm{if}{P}^{\&prime;}\&GreaterEqual;T_r,\\ P^{\&prime;}+T_l,& \mathrm{if}{P}^{\&prime;}<T_l.\end{array},\right.$ Wherein, when P ' < 0, b=0 is when P '>=0, and b=1, arranges locating information LM (j)=0, j=j+1, and to keep this pixel point value be I ' _i;

(3-3), when data can embed in pixel, it is e that record embeds completing place, otherwise, reduce T _lor increase T _r.

Extraction to data, specifically comprises the following steps:

Step11: find all certain class pixels, get the last significance bit of front setting value certain the class pixel point value on boundary position, obtain successively T _l, T _r, l, e, and last position 0 of these pixel values;

Step22: certain class pixel is calculated to parameters sortnig C, arrange all pixels by ascending order, obtain corresponding with identical in telescopiny sequence

Step33: from subsequence position LSB and l position location map information are set in middle extraction, and recover the partial pixel value of original image;

Step44: subsequence inverted order is extracted embedding information, recovers original image by location map information simultaneously;

Step55: finally the last significance bit of pixel in Step11 is replaced with to LSB.

Described rgb image is converted to RGBE image as shown in formula (1):

The described judgement that is positioned at certain colour vegetarian refreshments on border adopts two boundary methods or the threshold value method of difference;

Described two boundary methods, if determine that image is on region, the point that does not meet this condition is on border.

For borderline point, for obtaining marginal information, defined parameters D _e:

D _ebe to be determined by the minimum difference between boundary direction upper estimate and actual value, with following formula, determine level, vertical, diagonal angle, opposes the difference of angular direction upper estimate and actual value;

$D_h=|\frac{I_e^4+I_e^5}{2}-I_e|---\left(4\right)$

$D_v=|\frac{I_e^2+I_e^7}{2}-I_e|---\left(5\right)$

$D_d=|\frac{I_e^1+I_e^8}{2}-I_e|---\left(6\right)$

$D_{\mathrm{ad}}=|\frac{I_e^3+I_e^6}{2}-I_e|---\left(7\right)$

Wherein, D _hthe difference of horizontal direction upper estimate and actual value, D _vthe difference of vertical direction upper estimate and actual value, D _ddifference to angular direction upper estimate and actual value, D _adoppose the difference of angular direction upper estimate and actual value;

Getting wherein minimum value is D _e,

D _e＝min{D _h,D _v,D _d,D _ad} (8)

In conjunction with 8 pixels of this pixel periphery, can obtain actual value and estimated value minimum difference D on four boundary directions _eanother expression-form, by the D calculating before _ethe anti-directional information of releasing

$D_e|\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_e^k{I}_e^k-I_e|---\left(3\right)$

by 80 or 0.5 array, become, represent edge directional information, represent that pixel I is at parameatal 8 pixel values of E, I _ecurrent pixel point is in the value of E passage.

If D _e=D _h, λ _ebe expressed as (0,0,0,1/2,1/2,0,0,0).

The described threshold value method of difference, removes above-mentioned D _eoutward also must defined parameters D _mand threshold tau, wherein a D _mrepresent the value of current pixel point in E passage and its 8 distances that are worth mean values around:

$D_m|\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m{I}_e^k-I_e|---\left(9\right)$

λ wherein _m=1/8, represent that k adjacent pixel, in the value of E passage, use D _m-D _eas the condition on specification area and border, if D _m-D _ebe greater than setting threshold τ, judge that this pixel is on border, otherwise judge that this point is on region.

If be in regional location, D _mshould be very little, D in smooth domain equally _ea very little numerical value, but on edge, D on boundary direction _mvalue can be larger, and D _elikely a very little value, so can use D _m-D _econdition as specification area and border.

The histogrammic mode of described employing predicated error obtains threshold value initial value (T _l, T _r), concrete formula is:

Num (pixel prediction error ∈ [T _l, T _r]) > η * size (data) (24)

Wherein, size (data) represents to need the length of embedding data data, Num (pixel prediction error ∈ [T _l, T _r]) represent that predicated error P is in [T _l, T _r] between pixel number.

In described step 2, calculate (Ψ, Ω, Φ), for sampled point, at current path computation value C=Ψ+Ω+Φ, Ψ has represented the smoothness of content, is defined as follows:

$\mathrm{\&psi;}=\frac{1}{4}\underset{k}{\mathrm{\&Sigma;}}{[I^k-\stackrel{-}{I}]}^2,k\&Element;\left\{\mathrm{2,4,5,7}\right\}---\left(19\right)$

Wherein represent I ²(on), I ⁴(left side), I ⁵(right side), I ⁷(under) four neighboring pixels are worth mean value.For the pixel being on boundary direction, due to the effect of index passage, may cause can be very large by the local variance of above formula calculating, for fear of this problem, in conjunction with this pixel directional information of living in, calculates local variance;

Wherein, I ¹(upper left), I ³(upper right), I ⁶(lower-left), I ⁸(bottom right),

Here define a numeric representation and more may produce the possibility of overflow problem near 0 or 255 pixel value:

Q(x)＝max(Ν(I),Ν(255-I)) (21)

I is current pixel value, and Ν is (0,1) normal function, and because this pixel value in telescopiny can change, in order still to obtain correct sequence in leaching process, we replace I to use m=(I with the mean value of this pixel four points around ²+ I ⁴+ I ⁵+ I ⁷calculate)/4, gets Ω=λ Q (m) here, λ=2000;

While the image restoring after embedding data being rgb form by formula (2), the difference between each pixel value and original value is drawn by formula (22):

ε＝|P/256|×2 ^E-128 (22)

Wherein, formula (2) is:

r＝((R+0.5)/256)×2 ^E-128

g＝((G+0.5)/256)×2 ^E-128 (2)

b＝((B+0.5)/256)×2 ^E-128

By above formula (22), can be found out: in the situation that predicated error is certain, when E value is larger, resulting pixel value changes larger, i.e. graphical quality loss is larger, therefore the application introduces parameter Φ, makes the graphical quality loss reduction after embedding:

Φ＝ξ×2 ^E-128 (23)

Here λ and ξ=3 are used for adjusting Ψ, Ω, the weight relationship between Φ.

Described predicated error { P ₁, P ₂..., P _n, if on border, considering pixel value, this pixel can not surpass 255, adopt formula (10) to predict estimation to this point.

$\hat{I}=\min (\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_e^k{I}_c^k,255)---\left(10\right)$

Wherein, edge directional information, when prepass pixel neighboring pixel, min () gets wherein minimum value

If this pixel, on region, does not now need boundary information, employing pixel around four point values is predicted estimation, calculates suc as formula (11):

${\hat{I}}_c=\min (\frac{I_c^2+I_c^4+I_c^5+I_c^7}{4},255)---\left(11\right)$

Wherein, for the estimated value when prepass pixel place.

The exponential quantity I of each pixel and surrounding pixel thereof _ewith perhaps have difference, need to be this pixel pixel index around when calculating be converted to the I identical with self _e, in calculating process, use newly obtains replace corresponding

$I_{\mathrm{cnew}}^k=I_c^k\&times;2^{I_e^k-I_e}---\left(12\right)$

Obtaining predicated error is:

The pixel value of described embedding data it is as follows that concrete predicated error is adjusted mode

$P^{\&prime;}=\left\{\begin{array}{cc}2P+b,& \mathrm{ifP}\&Element;[T_l,T_r),\\ P+T_r,& \mathrm{ifP}\&GreaterEqual;T_r,\\ P+T_l,& \mathrm{ifP}<T_l.\end{array}---\left(14\right)\right.$

Wherein, the predicated error after P ' embedding data, T _l, T _rfor threshold value, predicated error is in pixel embedding data wherein.

Here b is a data embedding, and the result after current pixel embeds through data becomes:

Described data extraction procedure, extracts end in data, and by formula (16), (17) obtain original predicated error and the data of embedding:

Finally can recover original pixel value

Beneficial effect of the present invention:

For HDR image, the hiding reversible data algorithm that the application proposes has reached to sacrifice minimum picture quality the object that embeds more data.This hidden algorithm guarantee reversible hiding in, more existing reversible HDR image concealing algorithm is hidden capacity and is improved, and has guaranteed good visual effect.

Accompanying drawing explanation

Fig. 1 (a) LDR image;

The HDR image that Fig. 1 (b) is corresponding with Fig. 1 (a);

Fig. 2 pixel I and surrounding pixel distribution schematic diagram thereof;

Fig. 3 classify of image element schematic diagram;

Fig. 4 (a) original HDR image border and area dividing figure;

The outline map that Fig. 4 (b) two boundary representations go out;

The image boundary figure that Fig. 4 (c) candy operator obtains;

Fig. 5 data telescopiny;

The original RGBE form of Fig. 6 (a) Nave image;

Fig. 6 (b) region and border distribution plan;

Image after Fig. 6 (c) embedding data;

Image and original image difference after Fig. 6 (d) embedding data;

The original RGBE form of Fig. 7 (a) Dani image;

Fig. 7 (b) region and border distribution plan;

Image after Fig. 7 (c) embedding data;

Image and original image difference after Fig. 7 (d) embedding data.

Embodiment:

Below in conjunction with accompanying drawing, the present invention is described in detail:

Each pixel of LDR image represents with the integer type of 24bit, with RGB pattern, and R, G, tri-passages of B represent with 8bit respectively.In other words, LDR image can only represent a less dynamic range.For each pixel of HDR image, need the floating number of 96bit to represent, each passage of corresponding RGB need to represent with 32bit.In order to reduce the storage space of HDR image, the scholars such as Greg Ward have proposed brilliant RGBE form, and the mode of replacing by exponent mantissa makes each pixel take the integer data format of 32bit, color channel R wherein, and G, B and utilities index passage E respectively account for 8.

Without loss of generality, suppose that P (r, g, b) represents 96 HDR images of relocatable, (R, G, B, E) represents 32 RGBE images, and between two images, transformational relation is as follows:

Rgb image is converted to RGBE image as shown in formula (1):

By inverse operation, can be by the RGBE image restoring of 32 96 floating-point rgb images as shown in formula (2):

r＝((R+0.5)/256)×2 ^E-128

g＝((G+0.5)/256)×2 ^E-128 (2)

b＝((B+0.5)/256)×2 ^E-128

Pixel classification: define a certain pixel I is being I in prepass _c, in E passage, be I _e.Fig. 2 represents I and surrounding pixel distribution schematic diagram thereof.

For E passage, numerically represent the index information of pixel value, in image, determined the monochrome information of this pixel.Theoretical according to chromatism, the brightness calibration of occurring in nature is gradual change.In general, pixel intensity on the same area approach and color comparatively level and smooth, on E passage, be reflected as numerical value identical or close; Contrary, the pixel on zone boundary, the numerical value change of E passage is larger, and this sampling point is because brightness differs larger, visually eye-catching and separate larger with surrounding pixel, is presented as intuitively texture information, and such pixel is referred to as border.Research by the E passage to HDRE image shows that E passage has comprised image outline information.For this feature, make further research in this application.

In tradition hiding reversible data algorithm, a crucial step is accurately to estimate current a certain pixel point value.HDR image under brilliant form, adopts conventional pixel prediction mode: rhombus algorithm (center pixel value is determined by four adjacent pixels of its space) is invalid for HDR image.Merely RGB Color Channel is estimated, although can obtain one at the reasonable numerical value of current location, add the monochrome information of index passage, this finally has a significant impact picture quality.The second, due to HDR image pixel numerical value scope large (effect of index passage E), for the pixel being on boundary direction, estimate at very large error.As Fig. 3 pixel 1, take R passage as example, by rhombus algorithm, calculate predicted value and be difference value 261-196=56, obviously this is not that we wish to obtain.The 3rd, pixel is more similar at boundary direction (especially gradient direction).So we improve the accuracy of estimated value by boundary information.

In region, pixel is generally identical with the E value of neighboring pixel, and difference is embodied on EGB passage; And on border, because numerical value differs greatly, tetra-passages of RGBE have difference simultaneously, this is wherein E passage to differentia influence maximum.

In this application, reversible data embeds three color channel RGB successively, in reference channel E, under the help of texture information, for the pixel in a certain color channel (working as prepass), estimates to have improved accuracy.

Judge that pixel whether on border, can adopt two boundary methods, upper left two numerical value of E passage are determined.If determine that image is on region, the point that does not meet this condition is on border.The image outline figure that original HDR image as shown in Fig. 4 (a) obtains through this mode, as Fig. 4 (b), can find out that this is very similar to the image boundary information obtaining with Candy boundary operator shown in Fig. 4 (c).。

For borderline point, for obtaining marginal information, defined parameters D _e:

$D_e|\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_e^k{I}_e^k-I_e|---\left(3\right)$

D _eto be determined by the minimum difference between boundary direction upper estimate and actual value.We determine level with following formula, vertical, and the difference of angular direction upper estimate and actual value is opposed at diagonal angle.

$D_h=|\frac{I_e^4+I_e^5}{2}-I_e|---\left(4\right)$

$D_v=|\frac{I_e^2+I_e^7}{2}-I_e|---\left(5\right)$

$D_d=|\frac{I_e^1+I_e^8}{2}-I_e|---\left(6\right)$

$D_{\mathrm{ad}}=|\frac{I_e^3+I_e^6}{2}-I_e|---\left(7\right)$

Getting wherein minimum value is D _e,

D _e＝min{D _h,D _v,D _d,D _ad} (8)

λ _ebe one group and become with non-zero array by 0, its numerical value is determined by boundary information.If D _e=D _h, λ _ebe expressed as (0,0,0,1/2,1/2,0,0,0).

Also have a kind of border to determine mode, definition two number [D _m, D _e] an and threshold tau. calculate suc as formula (8) (9). D wherein _mrepresent the value of current pixel point in E passage and its 8 distances that are worth mean values around.

$D_m|\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m^k{I}_e^k-I_e|---\left(9\right)$

λ wherein _m=1/8, represent that k adjacent pixel is in the value of E passage.If be in regional location, D _mshould be very little.D in smooth domain equally _eit is a very little numerical value.But on edge, D on boundary direction _mvalue can be larger, and D _eit is likely a very little value.So we can use D _m-D _eas the condition on specification area and border, if D _m-D _ebe greater than setting threshold τ, judge that this pixel is on border, otherwise judge that this point is on region.

The processing of pixel: first the pixel when prepass is divided into the two class S that cross one another ₁, S ₂, at S ₁middle embedding data can not affect S ₂embedding.As Fig. 3, middle pixel belongs to S ₁, its predicted value is by S ₂and boundary information obtains.

Under the help of E channel boundary information, improved the accuracy to estimating when prepass pixel value.If this pixel on border, is considered pixel value and can not be surpassed 255, adopts formula (10) to predict estimation to this point.

$\hat{I}=\min (\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_e^k{I}_c^k,255)---\left(10\right)$

Wherein, edge directional information, when prepass pixel neighboring pixel.The R passage predicted value of Fig. 3 pixel 1 obtaining by this mode is (217+187)/2=202, predicated error 202-196=6, and this compares and has improved accuracy of estimation with rhombus algorithm.

If this pixel, on region, does not now need boundary information, employing pixel around four point values is predicted estimation, calculates suc as formula (11).

${\hat{I}}_c=\min (\frac{I_c^2+I_c^4+I_c^5+I_c^7}{4},255)---\left(11\right)$

At certain, in prepass is processed, should be noted that the exponential quantity I of each pixel and surrounding pixel thereof _ewith perhaps have difference, need to be this pixel pixel index around when calculating be converted to the I identical with self _e, in calculating process, use newly obtains replace corresponding

$I_{\mathrm{cnew}}^k=I_c^k\&times;2^{I_e^k-I_e}---\left(12\right)$

Obtaining predicated error is

In hiding reversible data algorithm, two threshold values are used to T _land T _r, T wherein _la negative, T _pit is a positive threshold value.They are divided into a center section and two threshold values outer part, wherein center section [T the difference histogram obtaining _l: T _r] can be used to carry out data embedding.Predicated error P ' after embedding data is obtained by formula (14)

$P^{\&prime;}=\left\{\begin{array}{cc}2P+b,& \mathrm{ifP}\&Element;[T_l,T_r),\\ P+T_r,& \mathrm{ifP}\&GreaterEqual;T_r,\\ P+T_l,& \mathrm{ifP}<T_l.\end{array}---\left(14\right)\right.$

Here b is a data embedding.Result after current pixel embeds through data becomes

Data fetch strategy is the inverse process of embedding strategy.In data, extract end, by formula (16), (17) obtain original predicated error and the data of embedding

Finally can recover original pixel value

Ordering strategy and location map, it is believed that data are embedded in the pixel value that has less predicated error minimum on image impact.For in the situation that embedding more data, reduce visual deformation as far as possible, embedding information is successively worthless line by line, pixel should sort according to its whether suitable embedding information, and then information is embedded wherein.In this article ordering strategy is improved, considered underflow/overflow problem and anamorphose loss problem.For sampled point at current path computation value C=Ψ+Ω+Φ.Ψ has represented the smoothness of content, is defined as follows

$\mathrm{\&psi;}=\frac{1}{4}\underset{k}{\mathrm{\&Sigma;}}{[I^k-\stackrel{-}{I}]}^2,k\&Element;\left\{\mathrm{2,4,5,7}\right\}---\left(19\right)$

A pixel has less local variance can illustrate that this point is on smooth domain.Yet for the pixel being on boundary direction, due to the effect of index passage, may cause can be very large by the local variance of above formula calculating, for fear of this problem, calculates local variance in conjunction with this pixel directional information of living in.

Ω has solved overflow/underflow problem to a certain extent.The process change of embedding data raw value, wherein, near 0 or 255 pixel value, more may produce overflow problem.Here define the possibility that a numeric representation produces problems.

Q(x)＝max(Ν(I),Ν(255-I)) (21)

X is current pixel value, and Ν is (0,1) normal function.Because this pixel value in telescopiny can change, in order still to obtain correct sequence in leaching process, we replace I. to use m=(I with the mean value of this pixel four points around ²+ I ⁴+ I ⁵+ I ⁷calculate)/4.Here get Ω=λ Q (m), λ=2000.

Because this algorithm changes the pixel value of HDR format-pattern, according to the feature of HDR image, while the image restoring after embedding data being rgb form by formula (2), the difference between each pixel value and original value is drawn by formula (22).

ε＝|P/256|×2 ^E-128 (22)

As can be seen from the above equation: in the situation that predicated error is certain, when E value is larger, resulting pixel value changes larger, i.e. graphical quality loss is larger.Therefore introduce parameter Φ herein, make the graphical quality loss reduction after embedding.

Φ＝ξ×2 ^E-128 (23)

Here λ and ξ are used for adjusting Ψ, Ω, the weight relationship between Φ.In our experiment, we set λ=2000, ξ=3. pixel to be embedded was entered after ascending order arrangement, carry out one by one embedding data, we have not only embedded smooth region more data, have also reduced the possibility of some overflow/underflow problems generations and the loss of picture quality.

It should be noted that at S ₁in pixel embedding data can not affect S ₂, because they are separated from each other.Moreover, S ₁the definite of sequencing information is by S ₂information decision, therefore, original arrangement information still can obtain after embedding information.

Although the ordering strategy that the application proposes has reduced the generation of overflow/underflow problem to a certain extent, can not eliminate completely.In order to address this problem, to define a location map record here and can not those by formula, adjust the pixel of twice.Wherein those are adjusted the point once having problems and are recorded as 0, and the point that continuous setup goes wrong for twice is recorded as 1.

In addition, in formula (10), in conjunction with boundary information prediction S ₁in the pixel value of some point,, can use and belong to equally S ₁pixel (as I ¹and I ⁸), and these points are perhaps through having adjusted.Therefore, determining of location map must embed and carry out simultaneously with data.After data have embedded, then by same procedure, embed location map on the other hand, can produce so again new overflow/underflow problem.If construct again a location map, address this problem, so new problem also may produce until there will be no overflow/underflow problem to produce, and the quantity of information of location map can increase like this, and increases according to the pixel of formula change the picture quality variation causing.In order to address this problem, after data embedding is complete, we detect [T _l: T _r] in the remaining some quantity that embeds information whether meet the demand of location map and supplementary, if could not; would increase threshold value and re-start embedding.

Suitable threshold value: as noted earlier, the data certain for quantity embed in image, can have a lot of threshold values to follow the example of.But in order to obtain best picture quality, be necessary to find and select applicable threshold value for data bulk. find in this article suitable threshold value initial value, than approaching gradually appropriate threshold, in the situation that embedding data is more, reduced greatly operand.Adopt the histogrammic mode of predicated error can obtain appropriate threshold initial value.

Num (pixel prediction error ∈ [T _l, T _r]) > η * size (data) (24)

Owing to being in the predicated error of the pixel on frontier point, in telescopiny, can recalculate, in order to obtain appropriate threshold before embedding, all pixel prediction errors are all provided by formula (11)-(13) herein.But in telescopiny, the introducing of boundary information increases pixel that can embedding data, adjusts η=0.9 here with η.

As shown in Figure 5, complete data embed algorithm: first, a width HDR image is transformed into HDRE form, embedding data is divided into 3 parts according to the order of sequence, embeds respectively tri-passages of rgb.Then, all pixels of a certain embedding passage are divided into two classes, S ₁, S ₂(representing as the grey in figure and white), accordingly embedding data is divided into two classes.Afterwards in telescopiny, a part of data are wherein embedded to S ₁, then repeat identical process and embed another part data to S ₂in.

The step of white pixel point hiding data is as follows:

Step1: find all white pixel points, and the last significance bit of limitrophe wherein 48 values is recorded as to LSB, the last active position 0 of this pixel;

Step2: E obtains marginal information according to reference channel, has two kinds of methods, two boundary methods and the threshold value method of difference;

Step3: according to boundary information, for white pixel point { I ₁, I ₂... I _nutilize formula (19-23) to calculate in turn (Ψ, Ω, Φ), and C is carried out to ascending order arrangement, obtain new embedding order and according to (11), (12), (13) obtain corresponding predicated error { P ₁, P ₂..., P _n}

Step4: obtain suitable threshold value initial value (T according to (24) _l, T _r);

Step5: to sequence carry out data embedding.For each pixel wherein in conjunction with boundary information, if this pixel predicated error P in region _i=P _i, on border, with formula (10), (12), (13) recalculate predicated error P _i

Utilize formula (14), (15) obtain the pixel value I ' of embedding data _i, by mode below, construct location map LM:

If ● we arrange a locating information LM (j)=1, j=j+1, and this pixel is reduced to do not change this pixel value;

If ● I ' _i∈ [0,255] still p " _ito be obtained by formula (14), wherein as P ' _iduring < 0, b=0 is as P ' _i>=0 o'clock, b=1.Locating information LM (j)=0 is set, j=j+1, and to keep this pixel value be I ' _i;

If data can embed in white pixel point, it is e that record embeds completing place, then carries out Step6; Otherwise, reduce T _lor increase T _r, repeat Step5;

Step6: detect residue sequence predicated error { P _e+1, P _e+2..., P _nin can embedding data quantity.In this step, can not be recorded as the part in location map by the pixel of twice of continuous setup, if lazy weight that can embedding data is to embed location map and LSB, reduce T _lor increase T _r, and get back to Step5; Otherwise, record site map length l and carry out next step.

Step7: embed location map information and LSB, it should be noted that in this step, predicated error adopts { P _e+1, P _e+2..., P _n.T _l(5bits), T _r(5bits), l (18bits), e (20bits) becomes scale-of-two, and it is become respectively to last position of extracting pixel in Step1.

Next the gray pixels point of a certain passage is embedded, its telescopiny is similar with white pixel point.

Data extraction procedure is the inverse operation of telescopiny, for a certain passage, first gray pixels point is carried out to data extraction.Leaching process is as follows:

Step11: find all gray pixels points, get borderline 48 pixel values and obtain last significance bit, obtain successively T _l, T _r, l, e, and the last active position 0 of these pixel values;

Step22: as embedded Step3, calculate C, arrange all gray pixels points by ascending order

Step33: from sequence 48 LSB of middle extraction and l position location map information.Utilize formula (16), (18) recover original image partial pixel value;

Step44: utilize formula (17) by sequence extract embedding information, recover original image mode identical with Step33 simultaneously;

In Step55:Step11, the last significance bit of pixel replaces with LSB.

Next the white pixel point of a certain passage is embedded, its telescopiny is similar with gray pixels point.

This experiment be take Matlab12a and is carried out emulation experiment, " 0 " that embedding data is produced by matlab, " 1 " random number as experiment porch.

Because HDR image comprises more monochrome information, the dynamic range of its expression is considerably beyond traditional LDR image, and this type of image can not directly be presented in conventional display apparatus.Yet through mapping process, be a kind of image processing techniques that HDR dynamic range of images Nonlinear Mapping is arrived to limited low-dynamic range, we can be presented at HDR image on 24 output display units of tradition.Here adopt the tone mapping function tonemap () that matlab carries to carry out this operation.

For weighing the picture quality after image embedding data, adopt Y-PSNR as criterion herein.Its definition is as shown in formula (25):

$\begin{array}{c}\mathrm{MSE}=\frac{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j-1}{\overset{N}{\mathrm{\&Sigma;}}}{(I_{\mathrm{ori}}(i,j)-I_{\mathrm{emb}}(i,j))}^2}{M\&times;N}\\ \mathrm{PSNR}=\frac{1}{4}\underset{c=1}{\overset{4}{\mathrm{\&Sigma;}}}10.{\mathrm{Log}}_{10}\left(\frac{R^2}{\mathrm{MSR}}\right)\end{array}---\left(25\right)$

In formula, I _orirepresent original RGBE image; I _embrepresent embedding data RGBE image afterwards; Each channel image size is M * N, totally 4 passages; R represents the max pixel value in original image, and c represents passage.

Fig. 6 (a)-Fig. 6 (d) and Fig. 7 (a)-Fig. 7 (d) have provided two test result figure.Wherein Fig. 6 (a)-Fig. 6 (d) experimental subjects is that size is the HDR image of 720*480.Data embedded quantity is 2bpp, after embedding information, image PSNR=43.43. Fig. 7 (a)-Fig. 7 (d) experimental subjects is the HDR image of 767*1023, embedding quantity of information is 2.2bpp, image PSNR=36.39. Fig. 6 (a) after embedding information, Fig. 6 (c) and Fig. 7 (a), the contrast of Fig. 7 (c) has shown that the HDR image original and through embedding data showing through tone mapping is very similar on direct feel.Fig. 6 (d) and Fig. 7 (d) figure shown after embedding data, original image and through embedding data image difference figure under same map condition, can find out embedding data after picture distortion little.

Have experimental result to find out, for HDR image, the hiding reversible data algorithm that this algorithm proposes has reached to sacrifice minimum picture quality the object that embeds more data.This hidden algorithm guarantee reversible hiding in, more existing reversible HDR image concealing algorithm is hidden capacity and is improved, and has guaranteed good visual effect.It should be noted that above all tests, the embedding information of recovery is without any mistake.

Claims (10)

1. the reversible data concealing method of high dynamic range images, is characterized in that, comprises the leaching process of data telescopiny and data, and the leaching process of data is the inverse operation of data telescopiny, and data embed and comprise the following steps:

Rgb image is converted to RGBE image, wish embedding data is divided into three parts according to the order of sequence, for embedding tri-passages of corresponding RGB of RGBE image;

All pixels of a certain embedding passage of tri-passages of RGB are divided into two classes: S ₁and S ₂, accordingly embedding data is divided into and S ₁and S ₂two corresponding classes, in telescopiny, embed S class data wherein ₁, then repeat identical process and embed another kind of data to S ₂in.

2. the reversible data concealing method of high dynamic range images as claimed in claim 1, is characterized in that, data embed and specifically comprise the following steps:

Step 1: the last significance bit that is positioned at border and goes forward to set certain class pixel of number is recorded as to LSB, and is 0 last position of such pixel;

Step 2: E obtains marginal information according to reference channel, according to marginal information to certain class pixel { I ₁, I ₂... I _ncalculate parameters sortnig C, according to parameters sortnig C, certain class pixel is carried out to ascending order arrangement, and calculate the rear corresponding predicated error { P of each pixel of sequence ₁, P ₂..., P _n;

Step 3: adopt the histogrammic mode of predicated error to obtain threshold value initial value (T _l, T _r), to pixel after sorting carry out successively data embedding, each pixel I _i(i=1,2 ...) data obtain the predicated error P ' of embedding data after embedding _iwith pixel value I ' _i, structure location map LM;

Step 4: remove the pixel of embedding data, detect residue sequence { I _e+1, I _e+2..., I _npredicated error { P _e+1, P _e+2..., P _nin can be used to the quantity of embedding data, if lazy weight that can embedding data is to embed location map and LSB, reduce T _lor increase T _r, and get back to step 3; Otherwise, record site map length l and carry out next step;

Step 5: embed location map information and LSB, in this step, predicated error adopts { P _e+1, P _e+2..., P _n; T _l(5bits), T _r(5bits), l (18bits), e (20bits) becomes scale-of-two, finally it is become respectively to last position of a setting value pixel in step 1.

3. the reversible data concealing method of high dynamic range images as claimed in claim 2, is characterized in that, the step of constructing location map LM in described step 3 is:

(3-1) when the pixel value of embedding data a locating information LM (j)=1 is set, j=j+1, j is the position of locating information in location map, and this pixel is reduced to I _i, do not change this pixel value;

(3-2) as the pixel value I ' of embedding data _i∈ [0,255] still wherein, $P_i^{\&Prime;}=\left\{\begin{array}{cc}2P^{\&prime;}+b,& \mathrm{if}{P}^{\&prime;}\&Element;[T_l,T_r),\\ P^{\&prime;}+T_r,& \mathrm{if}{P}^{\&prime;}\&GreaterEqual;T_r,\\ P^{\&prime;}+T_l,& \mathrm{if}{P}^{\&prime;}<T_l.\end{array},\right.$ Wherein, when P ' < 0, b=0 is when P '>=0, and b=1, arranges locating information LM (j)=0, j=j+1, and to keep this pixel point value be I ' _i;

(3-3), when data can embed in pixel, it is e that record embeds completing place, otherwise, reduce T _lor increase T _r.

4. the reversible data concealing method of high dynamic range images as claimed in claim 1, is characterized in that, the extraction of described data, specifically comprises the following steps:

Step11: find all certain class pixels, get the last significance bit of front setting value certain the class pixel point value on boundary position, obtain successively T _l, T _r, l, e, and last position 0 of these pixel values;

Step22: certain class pixel is calculated to parameters sortnig C, arrange all pixels by ascending order, obtain corresponding with identical in telescopiny sequence { I ₁, I ₂... I _n;

Step33: from { I ₁, I ₂... I _nsubsequence position LSB and l position location map information are set in middle extraction, and recover the partial pixel value of original image;

Step44:{I ₁, I ₂... I _nsubsequence inverted order is extracted embedding information, recovers original image by location map information simultaneously;

Step55: finally the last significance bit of pixel in Step11 is replaced with to LSB.

5. the reversible data concealing method of high dynamic range images as claimed in claim 1, is characterized in that, described rgb image is converted to RGBE image as shown in formula (1):

6. the reversible data concealing method of high dynamic range images as claimed in claim 2, is characterized in that, described in be positioned at certain class pixel on border judgement adopt two boundary methods or the threshold value method of difference.

7. the reversible data concealing method of high dynamic range images as claimed in claim 6, is characterized in that, described two boundary methods, if determine that image is on region, the point that does not meet this condition is on border.

8. the reversible data concealing method of high dynamic range images as claimed in claim 7, is characterized in that, the described threshold value method of difference also must defined parameters D _mand threshold tau, wherein a D _mrepresent the value of current pixel point in E passage and its 8 distances that are worth mean values around:

$D_m|\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m{I}_e^k-I_e|---\left(9\right)$

λ wherein _m=1/8, represent k adjacent pixel e value, use D _m-D _eas the condition on specification area and border, if D _m-D _ebe greater than setting threshold τ, judge that this pixel is on border, otherwise judge that this point is on region.

9. the high dynamic range images reversible data concealing method based on predicated error expansion as claimed in claim 2, is characterized in that, the histogrammic mode of described employing predicated error obtains threshold value initial value (T _l, T _r), concrete formula is:

Num (pixel prediction error ∈ [T _l, T _r]) > η * size (data) (24)

Wherein, size (data) represents to need the length of embedding data data, Num (pixel prediction error ∈ [T _l, T _r]) represent that predicated error P is in [T _l, T _r] between pixel number.

10. the reversible data concealing method of high dynamic range images as claimed in claim 4, is characterized in that, described data extraction procedure is extracted end in data, and by formula (16), (17) obtain original predicated error and the data of embedding:

Finally can recover original pixel value

Wherein, P ' is the predicated error after embedding data, T _la negative, T _pbe a positive threshold value, b is a data embedding.