CN101901470A - Image-tampering detection and recovery method based on energy-domain semi-fragile watermarking - Google Patents

Abstract

The invention relates to an image-tampering detection and recovery method based on an energy-domain semi-fragile watermarking. The method comprises the following steps: on the basis of image partitioning, a brightness difference value is extracted as an initial watermarking, on the basis of carrying convolutional encoding and scrambling on the initial watermarking, watermarking information is embedded into an image block after Walsh-Hadamard transform, and at last, Walsh-Hadamard reverse transform is carried out on the image to finish the embedment of the image watermarking to obtain a picture containing the watermarking; the image watermarking is extracted at a watermarking extraction and detection end through partitioning the image embedded with the watermarking and carrying out the Walsh-Hadamard transform, bit reverse scrambling, the convolutional encoding and other operations, and the preliminary detection and accurate detection of tampering are preformed through comparing the image brightness and other information with the watermarking information to locate a tampered region; and the tampered region is recovered by Viterbi decoders. The invention can resist attack of JPEG, JPEG2000 and other lossy compression and high-proportion lossy compression, and the blind detection of the watermarking and the recovery of the tampered image are achieved without additional information in the detection process.

Description

Distorted image based on energy-domain semi-fragile watermarking detects and restoration methods

Technical field

What the present invention relates to is the distorted image detection and the restoration methods in a kind of digital watermarking field, specifically is that a kind of distorted image based on energy-domain semi-fragile watermarking detects and restoration methods.

Background technology

The problem that the widespread use of digital picture has brought its content authenticity to differentiate.Particularly, make convenient to the edit-modify of digital picture along with the appearance of various high-quality and high-precision image processing equipment and a large amount of digital imaging processing software.Currently mainly adopt the digital watermark technology of half fragility at the evaluation of distorted image, this technology promptly has robustness to legal distortion, again illegal distortion is had susceptibility, and has certain station-keeping ability and former figure recovery capability.

Through literature search, Yu and Lu are at article " Mean quantization blind watermarking for imageauthentication[C] " (IEEE International Conference on Image Processing, VancouverBC, Canada.2000, proposed a digital picture authentication method 3:706-709), be specially: the weighted mean value by wavelet coefficient that the mean quantization technology is quantized is as the position of embed watermark.This algorithm is all carrying out modeling because the wavelet coefficient index word that distortion causes is distorted and attached to malice, and both have the Gaussian distribution of little variance and big variance respectively.This article thinks that the variation of wavelet coefficient obeys little variance Gaussian distribution, image is carried out the wavelet coefficient that malicious attack causes change the variance that often has greatly, and often have less variance by the index variation that accidentalia causes image fault to cause, thereby malice distorted to distort with non-malice to be made a distinction.This technology need both can detection of malicious not distorted by original image in authentication, can tolerate the subsidiary acceptable distortion that causes of compression again.But this method can not be well from being recovered original image by the image of distorting.

Find by retrieval again, (" embedding of high robust Large Volume Data ") (ICASSP 2001 at article " Robust high capacity data embedding[C] " for Lan and Tewfik, Utab.April 2001) (calendar year 2001 international acoustics, voice and signal Processing meeting) a kind of semi-fragile watermarking technology of middle proposition based on JPEG (JPEG (joint photographic experts group)) coding method, its method is: each 8 * 8 image block to original image carry out DCT (discrete cosine transform) conversion earlier, then the signal of each image block is arranged in order, be scanned into vector space with Hilbert, contrast JPEG quantization table resolves into littler subvector to vector again, this subvector is vertically arranged formation Hadamard matrix, the DCT coefficient that adopts each piece the Zig-zag scanning method to choose carries out the parity quantification, DCT coefficient inverse scan after the modulation is obtained the image block of embed watermark, and the combining image piece forms the image that contains watermark.Finish authentication by the quantization parameter that compares testing image with the parity situation of conforming to of former figure quantization parameter at last.This method is responsive to normal image processing operations reaction.But this method can not be used for blind Detecting, and can not recover original image from the image of being distorted.

Summary of the invention

The present invention is directed to the prior art above shortcomings, provide a kind of distorted image of the convolution error code based on the energy territory to detect and restoration methods, can resist lossy compression method such as JPEG and JPEG2000 and a high proportion of lossy compression method and attack; With the position that the mode accurate localization image of piecemeal is distorted, recover original image pixels information; With the influence of controllable mode control to picture quality; When detecting, do not need extra information, can realize the blind Detecting of watermark and the recovery of tampered image.

The present invention is achieved by the following technical solutions, at first on the basis of image block, by extracting luminance difference as original watermark, this original watermark is being done on the basis of convolutional encoding and scramble, watermark information is embedded in the image block behind Walsh-Hadamard transform, at last image is done Walsh-inverse Hadamard transform and embed, obtain moisture impression sheet to finish image watermark;

In watermark extracting and test side by image block to embed watermark, through the operations such as unrest and convolutional encoding that are inverted of Walsh-Hadamard transform, bit, extract image watermark, distort Preliminary detection and distort accurate detection, positioning tampering zone by the comparison of information such as brightness of image and watermark information;

Recover the tampered region by Viterbi decoding.

The process that described image watermark embeds is as follows:

The watermark carrier picture is carried out piecemeal with 8 * 8 size, and the absolute difference that calculates brightness between the adjacent image piece is represented with two binary bits as watermark content, with convolutional code (12,1,2) resulting whole watermark content is carried out convolutional encoding; This convolutional encoding is formed matrix, and utilization diagonal way repeatedly scans and forms the scramble result, to obtain the embedding ratio paricular value;

It is 4 * 4 piecemeal that the watermark carrier picture is carried out size, respectively each piece is carried out Walsh-Hadamard transform, and transformation for mula is F=H _M* f * H _N, wherein transformation matrix is:

${H^{\′}}_8=\frac{1}{2\sqrt{2}}\left[\begin{array}{cccccccc}1& 1& 1& 1& 1& 1& 1& 1\\ 1& 1& 1& 1& -1& -1& -1& -1\\ 1& 1& -1& -1& -1& -1& 1& 1\\ 1& 1& -1& -1& 1& 1& -1& -1\\ 1& -1& -1& 1& 1& -1& -1& 1\\ 1& -1& -1& 1& -1& 1& 1& -1\\ 1& -1& 1& -1& -1& 1& -1& 1\\ 1& -1& 1& -1& 1& -1& 1& -1\end{array}\right]$

And the DC component after the extraction conversion, last according to the DC component that embeds bit quantization Walsh-Hadamard transform, 4 * 4 sized images pieces are carried out Walsh-Hadamard transform again, transformation for mula is f=H _M* F * H _N, successively to each 4 * 4 sized images piece embed watermark, all finished embedding up to all pieces in order to last method;

Wherein:

Convolutional code (12,1,2) representative: convolutional encoding, it is a coding that is produced by the shift register of linearity, finite state, this shift register is made of the algebraic function maker of 12 grades (every grade 1 bits) and 2 linearities.Promptly this scrambler is with 12 groups, and the bit group of every group of 1 bit produces the coding result of two bits by displacement;

F: with the original data signal of matrix form statement;

H _M: Walsh-hadamard matrix, wherein the M value is 8;

H _N: Walsh-hadamard matrix, wherein the N value is 8;

F: original signal is through the matrix signal behind Walsh-Hadamard transform;

H ' ₈: the Walsh-Hadamard transform matrix H during N=8 _N, i.e. Walsh-the hadamard matrix of this method employing.

Above-mentioned quantization method is as follows:

Interval in the positive coordinate district of X-axis with Δ q is divided between several region, and each interval end points is all represented a binary bits, the bit difference of representing between the adjacent end points: the end points of the Δ q of odd-multiple is represented bit 1; And the end points of the Δ q of even-multiple is represented bit 0; After setting up quantized interval figure, between i* Δ q and (i+1) value to be quantified between the * Δ q, relatively embed the bit of bit and i* Δ q representative, the two is identical, replaces the value of being quantized with i* Δ q, and the two is different, replaces the value of being quantized with (i+1) * Δ q.

Wherein:

I: the expression positive integer, desirable 1,2,3....I, wherein, I represents that institute divides quantification interval number amount;

Δ q: be quantized interval, be about to a numerical value interval and be divided into I part that wherein the length of each part is Δ q.

The process in positioning tampering of the present invention zone is as follows:

At first moisture impression sheet to be detected is carried out piecemeal with 4 * 4 size, each image block is carried out Walsh-Hadamard transform, and extract its DC component, then this DC component is carried out inverse quantization, extract the embed watermark bit by inverse quantization, and with the unrest that is inverted of the bit of all extractions, after passing through Viterbi decoding then, decode the original watermark content, again this watermark content is carried out convolutional encoding, the result of this convolutional encoding and the watermark bit comparison of extracting from Walsh-Hadamard transform are distorted Preliminary detection and distort accurate detection, thus the positioning tampering zone.

Wherein, the described Preliminary detection of distorting is as follows with the process of distorting accurate detection:

Distort Preliminary detection: the result of the watermark content convolutional encoding extracted and the watermark bit of extracting are compared, mark the image block of the position that comparison back bit do not conform to from Walsh-Hadamard transform;

Distort accurate detection: moisture impression sheet is carried out piecemeal with 8 * 8 size, differential coding is carried out in the brightness of adjacent image piece, and coding result and Viterbi decoding result are compared, mark comparison back bit image block inequality;

To distort at last in the Preliminary detection testing result with distort accurate testing result and superpose, export the last testing result of distorting.

Above-mentioned quantification method is as follows:

Produce quantized interval with the method between the dividing regions identical with quantization method, relatively watermark data of Qian Ruing and quantized interval, if the watermark data that embeds is during between i* Δ q with (i+1) between the * Δ q, the watermark data that calculate to embed respectively and two end points apart from length, the bit of getting nearer endpoint value representative is the bit that will extract.

Wherein:

I: expression positive integer, desirable 1,2,3 ... I, wherein, I represent divide and quantize the interval number amount, this value is determined by Δ q;

Δ q: be quantized interval, be about to a numerical value interval and be divided into I part that wherein the length of each part is Δ q.

Recovery of the present invention tampered region method is as follows:

For detecting the image of being distorted, according to distorting Preliminary detection and the result who distorts accurate detection, the zone of being distorted of orienting, utilize moisture impression sheet to obtain the initial value of brightness of image, the former figure luminance difference score value that utilizes Viterbi decoding to decode then, adding up by iteration recovers monochrome information by the former pixel of tampered position.

Distorting Preliminary detection and distorting that accurate detection and recovery distort of image of the present invention is to use the C++ programming language to carry out on general purpose PC.

The present invention is in conjunction with the anti-tamper method of traditional semi-fragile watermarking image, the characteristics and the convolutional encoding speed that utilize that some energy territory conversion such as Walsh-Hadamard transform speed is fast, complexity is low, concentration of energy are high after the conversion are fast, the characteristics that antijamming capability is strong, the employing convolutional code error code of novelty realizes that watermark distorts detection, can very effective opposing live various lossy compression method, and can recover the image of being distorted by original image and watermark information.

Description of drawings

Fig. 1 is a process flow diagram of the present invention.

Fig. 2 is that watermark of the present invention embeds process flow diagram.

Fig. 3 distorts Preliminary detection and distorts accurate testing process figure for of the present invention.

Fig. 4 is a recovery tampered image process flow diagram of the present invention.

Fig. 5 is applied to the design sketch of 512 * 512 gray scales (former figure is colored, and the printed drawings of submitting the is a black and white) picture of aircraft by name for the present invention, wherein:

A. be to distort picture;

B. be the Preliminary detection result;

C. be accurate testing result;

D. be restoration result.

Fig. 6 is applied to the design sketch of lossy compression method challenge trial for the present invention, wherein:

A1 is an original image;

A2 is a tampered image;

B1 is the testing result of distorting of JPEG85% quality;

B2 is the restoration result of JPEG85% quality;

C1 is the testing result of distorting of JPEG65% quality;

C2 is the restoration result of JPEG85% quality;

D1 is the testing result of distorting of JPEG45% quality;

D2 is the restoration result of JPEG45% quality;

E1 is the testing result of distorting of JPEG25% quality;

E2 is the restoration result of JPEG25% quality;

F1 is the testing result of distorting of JPEG2000 ratio of compression 10: 1;

F2 is 10: 1 a restoration result of JPEG2000 ratio of compression;

G1 is the testing result of distorting of JPEG2000 ratio of compression 15: 1;

G2 is 15: 1 a restoration result of JPEG2000 ratio of compression;

H1 is the testing result of distorting of JPEG2000 ratio of compression 20: 1;

H2 is 20: 1 a restoration result of JPEG2000 ratio of compression.

Embodiment

The present invention is elaborated with embodiment below in conjunction with accompanying drawing, embodiment is being to implement under the prerequisite with the technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Embodiment 1

Present embodiment adopts still image gray scale picture LENA as the watermark carrier picture, in this example,

The first step: referring to Fig. 2, watermark carrier picture LENA is carried out piecemeal with 8 * 8 size, the absolute difference that calculates brightness between the adjacent image piece is as watermark content.Represent with two binary bits.With convolutional code (12,1,2) resulting whole watermark content are carried out convolutional encoding.This convolutional encoding is formed matrix, and utilization diagonal way repeatedly scans and forms the scramble result, to obtain the embedding ratio paricular value.It is 4 * 4 piecemeal that the watermark carrier picture is carried out size, respectively each piece is carried out Walsh-Hadamard transform (Fig. 2 is called for short Walsh transform), and transformation for mula is F=H _M* f * H _N, wherein transformation matrix is

${H^{\′}}_8=\frac{1}{2\sqrt{2}}\left[\begin{array}{cccccccc}1& 1& 1& 1& 1& 1& 1& 1\\ 1& 1& 1& 1& -1& -1& -1& -1\\ 1& 1& -1& -1& -1& -1& 1& 1\\ 1& 1& -1& -1& 1& 1& -1& -1\\ 1& -1& -1& 1& 1& -1& -1& 1\\ 1& -1& -1& 1& -1& 1& 1& -1\\ 1& -1& 1& -1& -1& 1& -1& 1\\ 1& -1& 1& -1& 1& -1& 1& -1\end{array}\right]$

And the DC component after the extraction conversion.According to the DC component that embeds bit quantization Walsh-Hadamard transform, quantization method is as follows: between the interval dividing regions of the positive coordinate district of X-axis with Δ q=12, each interval end points is all represented a binary bits, the bit difference of representing between the adjacent end points.For example, the Δ q of odd-multiple (Δ q, 3 Δ q, 5 Δ q ...) end points represent bit 1; And the Δ q of even-multiple (2 Δ q, 4 Δ q, 6 Δ q ...) end points represent bit 0.After setting up quantized interval figure, if value to be quantified during between i* Δ q with (i+1) between the * Δ q, is seen the watermark bit that will embed this moment, if the bit that embeds is identical with the bit that i* Δ q represents, then replace the value of being quantized, otherwise replace the value of being quantized with (i+1) * Δ q with i* Δ q.Finish embedding operation by above quantizing process.At last, 4 * 4 sized images pieces are carried out Walsh-Hadamard transform again, transformation for mula is f=H _M* F * H _NSuccessively to each 4 * 4 sized images piece embed watermark, all finished embedding in order to last method up to all pieces.

At last image is done Walsh-inverse Hadamard transform and embed, obtain moisture impression sheet to finish image watermark.

After embedding is finished the picture that contains watermark is calculated its value according to the formula of Y-PSNR (PSNRP, i.e. Peak Signal Noise Ratio).Computing formula is as follows:

$\mathrm{PSNR}=10{\log }_{10}\frac{{255}^2}{\frac{1}{\mathrm{MN}}\underset{m=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}{[f(m,n)-g(m,n)]}^2}$

Make e (m, n)=f (m, n)-g (m, n)

$\mathrm{PSNR}=10{\log }_{10}\frac{{255}^2}{E\left(e{(m,n)}^2\right)}$

$=10{\log }_{10}\frac{{255}^2}{D\left(e(m,n)\right)+E{\left(e(m,n)\right)}^2}$

Wherein: E (.) represents average, and D (.) represents variance.For the LENA image behind the embed watermark, its Y-PSNR is 43.0948, shows that the watermark data of this method embedding is little to the influential effect of original image.

Wherein:

Convolutional code (12,1,2) representative: convolutional encoding, it is a coding that is produced by the shift register of linearity, finite state, this shift register is made of the algebraic function maker of 12 grades (every grade 1 bits) and 2 linearities.Promptly this scrambler is with 12 groups, and the bit group of every group of 1 bit produces the coding result of two bits by displacement;

F: with the original data signal of matrix form statement;

H _M: Walsh-hadamard matrix, wherein M=8;

H _N: Walsh-hadamard matrix, wherein M=8;

F: through the matrix signal behind Walsh-Hadamard transform;

H ' ₈: the Walsh-Hadamard transform matrix H during N=8 _N

I: positive integer, f ∈ (0～I), I be divide and quantize the interval number amount.

Second step: referring to Fig. 3, at first moisture impression sheet to be detected is carried out piecemeal with 4 * 4 size, each image block is carried out Walsh-Hadamard transform (Fig. 3 is called for short Walsh transform), and extract its DC component, and then this DC component is carried out inverse quantization, extract the embed watermark bit by inverse quantization, and with the unrest that is inverted of the bit of all extractions, after passing through Viterbi decoding then, decode the original watermark content, again this watermark content is carried out convolutional encoding.The watermark bit of utilizing the result of this convolutional encoding and extracting from Walsh-Hadamard transform can be distorted Preliminary detection and be distorted accurate detection.

Described inverse quantization is meant: the division methods identical with the watermark built-in end produces quantized interval, getting quantized interval is Δ q=12, the watermark data and the quantized interval that embed are compared, if the watermark data that embeds is during between i* Δ q with (i+1) between the * Δ q, the watermark data that calculate to embed respectively and two end points apart from length, the bit of getting nearer endpoint value representative is the bit that will extract.

Wherein:

I: expression positive integer, desirable 1,2,3 ... I, wherein, I represent divide and quantize the interval number amount, this value is determined by Δ q.

Δ q: be quantized interval, be about to a numerical value interval and be divided into I part that wherein the length of each part is Δ q.

The described Preliminary detection of distorting is meant: the result of the watermark content convolutional encoding of extraction and the watermark bit of extracting from Walsh-Hadamard transform are compared, if bit does not conform to, illustrate that then the image block of this position is distorted, and it is marked.

Described accurate detection is meant: moisture impression sheet is carried out piecemeal with 8 * 8 size, differential coding is carried out in brightness to the adjacent image piece, and coding result and Viterbi decoding result (watermark content) compared, if there is bit inequality, the image that this image block then is described is distorted, and it is marked.

The testing result that to distort at last in the Preliminary detection superposes with accurate testing result, exports the last testing result of distorting.

The 3rd step: referring to Fig. 4.According to distorting Preliminary detection and accurately distorting the result of detection, the zone of being distorted, utilize moisture impression sheet to obtain the initial value of brightness of image, the former figure luminance difference score value that utilizes Veterbi decoding to go out then, adding up by iteration recovers monochrome information by the former pixel of tampered position.

Distorting Preliminary detection and accurately distorting detection and recover to distort of image of the present invention is to use the C++ programming language to carry out on general purpose PC.

Embodiment 2

512 * 512 gray scales (colour) picture that above method is applied to aircraft by name is as the watermark carrier picture, and adds an airplane (shown in Fig. 5 .a) above the original image aircraft again.For accurate positioning image tampered position, the size of piecemeal is 4 * 4.Picture is applied to the above testing process of distorting.The result who distorts detection can clearly see the zone that image is distorted shown in Fig. 5 .b; Accurately the result who detects is shown in Fig. 5 .c, former figure restoration result such as Fig. 5 .d.

Embodiment 3

This method is applied to the LENA picture, makes the lossy compression method challenge trial.Experiment effect figure as shown in Figure 6, from actual effect, getting Δ q=17 lists respectively in all cases then, comprise: JPEG85%, 65%, 40%, 25% quality, the JPEG2000 ratio of compression is respectively 10: 1,15: 1,20: 1 o'clock the testing result of distorting, and the former figure of carrying out of tampered region recovered, its result is as shown in Figure 6.

More than typical case's test picture is used in experiment, from aspects such as the detection of distorting of the structure of watermark and embedding, still image and recoveries the present invention is detected respectively, thereby has proved the validity of this method.

Claims (5)

1. the distorted image based on energy-domain semi-fragile watermarking detects and restoration methods, it is characterized in that:

At first on the basis of image block, by extracting luminance difference as original watermark, this original watermark is being done on the basis of convolutional encoding and scramble, watermark information is embedded in the image block behind Walsh-Hadamard transform, at last image is done Walsh-inverse Hadamard transform and embed, obtain moisture impression sheet to finish image watermark;

In watermark extracting and test side by image block to embed watermark, through the operations such as unrest and convolutional encoding that are inverted of Walsh-Hadamard transform, bit, extract image watermark, distort Preliminary detection and distort accurate detection, positioning tampering zone by the comparison of information such as brightness of image and watermark information;

Recover the tampered region by Viterbi decoding.

2. the distorted image based on energy-domain semi-fragile watermarking according to claim 1 detects and restoration methods, and it is characterized in that: the process that described image watermark embeds is as follows:

The watermark carrier picture is carried out piecemeal with 8 * 8 size, and the absolute difference that calculates brightness between the adjacent image piece is represented with two binary bits as watermark content, with convolutional code (12,1,2) resulting whole watermark content is carried out convolutional encoding; This convolutional encoding is formed matrix, and utilization diagonal way repeatedly scans and forms the scramble result, to obtain the embedding ratio paricular value;

It is 4 * 4 piecemeal that the watermark carrier picture is carried out size, respectively each piece is carried out Walsh-Hadamard transform, and transformation for mula is F=H _M* f * H _N, wherein transformation matrix is:

And the DC component after the extraction conversion, last according to the DC component that embeds bit quantization Walsh-Hadamard transform, 4 * 4 sized images pieces are carried out Walsh-Hadamard transform again, transformation for mula is f=H _M* F * H _N, successively to each 4 * 4 sized images piece embed watermark, all finished embedding up to all pieces in order to last method;

Wherein:

Convolutional code (12,1,2) representative: convolutional encoding, it is a coding that is produced by the shift register of linearity, finite state, this shift register is made of the algebraic function maker of 12 grades (every grade 1 bits) and 2 linearities.Promptly this scrambler is with 12 groups, and the bit group of every group of 1 bit produces the coding result of two bits by displacement;

F: with the original data signal of matrix form statement;

H _M: Walsh-hadamard matrix, wherein the M value is 8;

H _N: Walsh-hadamard matrix, wherein the N value is 8;

F: original signal is through the matrix signal behind Walsh-Hadamard transform;

H ' ₈: the Walsh-Hadamard transform matrix H during N=8 _N, i.e. Walsh-the hadamard matrix of this method employing.

3. the distorted image based on energy-domain semi-fragile watermarking according to claim 2 detects and restoration methods, and it is characterized in that: quantization method is as follows:

Interval in the positive coordinate district of X-axis with Δ q is divided between several region, and each interval end points is all represented a binary bits, the bit difference of representing between the adjacent end points: the end points of the Δ q of odd-multiple is represented bit 1; And the end points of the Δ q of even-multiple is represented bit 0; After setting up quantized interval figure, between i* Δ q and (i+1) value to be quantified between the * Δ q, relatively embed the bit of bit and i* Δ q representative, the two is identical, replaces the value of being quantized with i* Δ q, and the two is different, replaces the value of being quantized with (i+1) * Δ q;

Wherein:

I: the expression positive integer, desirable 1,2,3....I, wherein, I represents that institute divides quantification interval number amount;

Δ q: be quantized interval, be about to a numerical value interval and be divided into I part that wherein the length of each part is Δ q.

4. the distorted image based on energy-domain semi-fragile watermarking according to claim 1 detects and restoration methods, and it is characterized in that: the process in described positioning tampering zone is as follows:

At first moisture impression sheet to be detected is carried out piecemeal with 4 * 4 size, each image block is carried out Walsh-Hadamard transform, and extract its DC component, then this DC component is carried out inverse quantization, extract the embed watermark bit by inverse quantization, and with the unrest that is inverted of the bit of all extractions, after passing through Viterbi decoding then, decode the original watermark content, again this watermark content is carried out convolutional encoding, the result of this convolutional encoding and the watermark bit comparison of extracting from Walsh-Hadamard transform are distorted Preliminary detection and distort accurate detection, thus the positioning tampering zone.

5. detect and restoration methods according to claim 1 or 4 described distorted images based on energy-domain semi-fragile watermarking, it is characterized in that: the described Preliminary detection of distorting is as follows with the process of distorting accurate detection:

Distort Preliminary detection: the result of the watermark content convolutional encoding extracted and the watermark bit of extracting are compared, mark the image block of the position that comparison back bit do not conform to from Walsh-Hadamard transform;

Distort accurate detection: moisture impression sheet is carried out piecemeal with 8 * 8 size, differential coding is carried out in the brightness of adjacent image piece, and coding result and Viterbi decoding result are compared, mark comparison back bit image block inequality;

To distort at last in the Preliminary detection testing result with distort accurate testing result and superpose, export the last testing result of distorting.

Images