CN102819842A - Displacement JPEG (joint photographic experts group) double-compression tampering blind detection method based on condition symbiotic probability matrix - Google Patents

Abstract

The invention relates to a displacement JPEG (joint photographic experts group) double-compression tampering blind detection method based on a condition symbiotic probability matrix. The method comprises the steps: firstly, carrying out differential and thresholding treatment of horizontal, vertical, main diagonal and secondary diagonal directions on an amplitude matrix of a discrete cosine transform coefficient quantified by a JPEG; then modeling the four thresholding differential matrixes by the condition symbiotic probability matrix; selecting an element of the condition symbiotic probability matrix as characteristic data, and carrying out dimension reduction treatment on the characteristic data by principal component analysis; and finally, judging whether an image block is subjected to displacement JPEG double-compression through a support vector machine technology. An experiment shows that the method has obvious advantages compared with the existing algorithm when the image block is small, so that the practicability of the displacement JPEG double-compression tampering detection is greatly improved, and good foundation is established for further development of the digital image forensics field.

Description

Blind checking method is distorted in two compressions based on the displacement JPEG of condition symbiosis probability matrix

Technical field

The present invention relates to a kind of distorted image detection method, specifically is that blind checking method is distorted in the two compressions of a kind of displacement JPEG based on condition symbiosis probability matrix.

Background technology

To obtain equipment and intelligent image process software fast-developing along with Internet technology, high-performance image, distorts the content of piece image and do not stay any vestige and become ery easy.In fact, digital picture authenticity problem is very severe, and sometimes malice distorts even possibly bring jural dispute to us.Therefore, the authenticity problem of digital picture has crucial meaning in fields such as judicial expertise and copyright protections.For the digital image evidence collecting technology, mainly can be divided into two big types, i.e. active and passive evidence collecting method.Initiatively evidence collecting method is that digital watermarking or digital signature are embedded digital picture, identifies the authenticity of image through the change that detects watermark or signature.But because the cost of great number, present most of image acquisition equipments do not have the function of watermark or signature embedding.Comparatively speaking, do not needed the prior imformation of watermark and image by the motion video forensic technologies, this evidence obtaining mode has very big development potentiality.

As everyone knows, JPEG is an International Standard of image compression that is widely used.Therefore, the authenticity of evaluation jpeg format image becomes particularly important in the digital image evidence collecting field.The copy-paste operation is a kind of common means in the jpeg image content tampering; The zone of being distorted has experienced twice inconsistent JPEG compression of grid dividing usually; This phenomenon is called the displacement two compressions of JPEG (shifted double JPEG compression, SD-JPEG compression) and distorts in the jpeg image content tampering.Weiqi Luo etc. are at document " A novel method for detecting cropped and recompressed image block [C] " (IEEE International Conference on Acoustic; Speech and Signal Processing; Honolulu; Hawaii; USA:IEEE has proposed that a kind of (blocking artifact characteristics matrix, algorithm BACM) detect the image blocks that the two compressions of displacement JPEG are distorted based on the blocking effect eigenmatrix in 2007:217-220); BACM has represented the symmetric shape of a rule for the image block of the single compression of JPEG, and the two fail in compressions of displacement JPEG this symmetry.Zhenhua Qu etc. are at document " A convolutive mixing model for shifted double JPEG compression with application to passive image authentication [C] " (IEEE International Conference on Acoustic; Speech and Signal Processing; Las Vegas; Nevada, USA:IEEE, 2008:1661-1664) the two compression problem of JPEG that will be shifted are regarded a noise convolution mixture model as; By independent component analysis (independent components analysis; ICA) theoretical thought, (independent value map, symmetry IVM) identifies whether image block experiences the two compressions of displacement JPEG to investigate the independent values mapping graph.When image block was bigger, these two kinds of methods can both obtain reasonable performance.Yet when image block was smaller, their detection accuracy was still lower.

Summary of the invention

The object of the invention is the deficiency to prior art, provides the two compressions of a kind of displacement JPEG based on condition symbiosis probability matrix to distort blind checking method.When this method image block is smaller, has clear superiority than existing method.

The technical scheme that realizes the object of the invention is:

There is correlativity between the element of two dimension JPEG matrix of coefficients.The range value that two-dimentional JPEG matrix of coefficients has been upset in the two compressions of displacement JPEG distributes, and weakened the correlativity between the coefficient, and its phase change is less.In order to reduce the influence of picture material, strengthen the two pinch effects of displacement JPEG, the magnitude matrix of the discrete cosine transform coefficient that at first JPEG is quantized is carried out level, the secondary diagonal angle four direction difference of vertical, main diagonal sum and thresholding and is handled; Then, adopt condition symbiosis probability matrix that the difference matrix of these four thresholdings is carried out modeling, the element of choosing condition symbiosis probability matrix is as characteristic, and with principal component analysis (PCA) characteristic carried out dimension-reduction treatment; Whether adjudicate image block through the two compressions of displacement JPEG through SVMs (SVM) training and detection at last.

Described characteristic is extracted, and may further comprise the steps:

1. after colored jpeg image piece being decompressed, obtain brightness (Y) and colourity (C _bAnd C _r) the two-dimentional JPEG matrix of coefficients X (referring to the two-dimensional matrix that discrete cosine transform coefficient amplitude that JPEG quantizes is formed) of component;

2. the two-dimentional JPEG matrix of coefficients X with the Y component does four direction difference (being the secondary diagonal angle of level, vertical, main diagonal sum) processing, obtains four two-dimentional JPEG coefficient difference sub matrixs:

E _h(i,j)=X(i,j)-X(i+1,j)

E _v(i,j)=X(i,j)-X(i,j+1)

E _d(i,j)=X(i,j)-X(i+1,j+1)

E _m(i,j)=X(i+1,j)-X(i,j+1)

E wherein _h, E _v, E _dAnd E _mThe two-dimentional JPEG coefficient difference sub matrix of representing level, the secondary diagonal of vertical, main diagonal sum respectively; Four two-dimentional JPEG coefficient difference sub matrix E that 3. will obtain _h, E _v, E _dAnd E _mCarrying out thresholding handles:

$E(i,j)=\left\{\begin{array}{cc}E(i,j)& \left|E(i,j)\right|<T\\ T& E(i,j)\&GreaterEqual;T\\ -T& E(i,j)\&le;-T\end{array}\right.$

T=2 wherein;

4. utilize condition symbiosis probability matrix model that four two-dimentional JPEG coefficient difference sub matrixs of thresholding are carried out modeling, the condition symbiosis probability matrix of level, the secondary diagonal of vertical, main diagonal sum is following:

$p\left\{(E_h(i+1,j)=m,E_h(i+2,j)=n)\right|E_h(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l,E_h(i+1,j)=m,E_h(i+2,j)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l)}$

$p\left\{(E_v(i,j+1)=m,E_v(i,j+2)=n)\right|E_v(i,j)=l\}=\frac{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l,E_v(i,j+1)=m,E_v(i,j+2)=n))}{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l)}$

$p\left\{(E_d(i+1,j+1)=m,E_d(i+2,j+2)=n)\right|E_d(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l,E_d(i+1,j+1)=m,E_d(i+2,j+2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l)}$

$p\left\{(E_m(i+1,j-1)=m,E_m(i+2,j-2)=n)\right|E_m(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l,E_m(i+1,j-1)=m,E_m(i+2,j-2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l)}$

Wherein, M and N are respectively the level and the vertical direction sizes of the two-dimentional JPEG coefficient difference sub matrix of thresholding.L, m, n ∈ Z;-T≤l, m, n≤T; And

therefore; Adopt four conditional probability entry of a matrix elements as characteristic, when T=2, its intrinsic dimensionality is 500 dimensions.

Described characteristic dimension-reduction treatment may further comprise the steps:

Make that eigenmatrix is C _X, its size is m ₁* n ₁, proper vector of each line display of matrix, therefore, intrinsic dimensionality is m ₁The PCA dimensionality reduction is intended to find transformation matrix W, obtains new feature C _Y=W ^T(C _X-μ _X), make new feature C _YHave the composition of better classification capacity, may further comprise the steps:

1. calculate the covariance matrix ∑ _X:

${\mathrm{\&Sigma;}}_X=\frac{1}{n_1}\left((C_X-{\mathrm{\&mu;}}_X){(C_X-{\mathrm{\&mu;}}_X)}^T\right)$

Wherein, μ _XBe eigenvectors matrix C _XAverage, ∑ _XIt is symmetric matrix;

2. through finding the solution secular equation

det(λ _iI-∑ _X)＝0

Obtain eigenvalue _i(1≤i≤m ₁).The eigenwert of being asked constitutes the diagonal covariance matrix ∑ _Y

3. to each eigenvalue _i(1≤i≤m ₁), according to (λ _iThe I-∑ _X) w _i=0 finds the solution corresponding w _iThereby, obtaining proper vector, proper vector should be normalization and linear independence;

4. eigenwert is sorted, choose bigger preceding d (d＜m ₁) the pairing proper vector composition of individual eigenwert transformation matrix W;

5. new feature Matrix C _Y=W ^T(C _X-μ _X).

Described SVM training and detection may further comprise the steps:

1. detection, quality factor q F are distorted in two compressions for the displacement JPEG of image block ₁Generally be unknown, suppose QF ₁50,60,70,80, evenly choose quality factor q F among the 90} at random ₂Can from the jpeg image header file, calculate.According to quality factor q F ₁And QF ₂, use the size that comprises of known class to be QF as B * B, quality factor ₂Single compressed picture blocks of JPEG and quality factor be QF ₁And QF ₂The two compressed picture blocks composing training collection of displacement JPEG; Each image block to training set extracts above-mentioned condition symbiosis probability matrix characteristic; The characteristic that obtains is through normalization and PCA dimension-reduction treatment; And indicate classification under its image block (it is the single compressed picture blocks of JPEG to zone bit for-1 expression, the two compressed picture blocks of 1 expression displacement JPEG) with zone bit;

2. the kernel function of svm classifier device is selected radially basic kernel function for use, and the training set characteristic is sent into SVM, through grid search and cross validation, obtains cross validation accuracy rate the highest optimized parameter c and g, and trains SVMs with them;

3. extract the condition symbiosis probability matrix characteristic of test pattern piece, the characteristic that obtains uses the SVM that trained to judge its affiliated classification through normalization and PCA dimension-reduction treatment.

In the SVM training and detecting, the synthetic jpeg image of described image block is distorted detection, may further comprise the steps:

1. from the header file of synthetic jpeg image, calculate quality factor q F ₂

2. according to the matrix of B * B size and quality factor q F ₂, make QF ₁{ 50,60,70; 80, evenly choose at random among the 90}, utilize said method training SVM; The SVM that trains is obtained 1% wrong positive routine rate through the adjustment decision-making value, and (FP refers to that the image block of the single compression of JPEG is by the ratio that is categorized into the two image blocks that compress of displacement JPEG of mistake.）；

3. use the big or small sub-piece of B * B from left to right and from top to bottom to travel through given synthetic jpeg image with the interval of 32 pixels; Obtain the sub-piece of B to be detected * B size successively; Extract the characteristic of the sub-piece of each B to be detected * B size, and through normalization and PCA dimension-reduction treatment.Then; (image block of the two compressions of displacement JPEG is marked as " 1 " with the SVM that trains every sub-block to be carried out classification and marking successively; The image block of the single compression of JPEG is marked as " 0 "); Behind the intact all image blocks to be detected of mark, can obtain a two-value decision matrix (binary decision matrix, BDM);

4. using window size is that 3 * 3 median filter carries out filtering to above-mentioned two-value decision matrix; Further remove the single compressed picture blocks of isolated JPEG that some possibly be mistaken for the two compressions of displacement JPEG; Can obtain filtered two-value decision matrix (filtered binary decision matrix, FBDM);

5. according to the element value of FBDM, synthesizing the tampered region that jpeg image relevant position mark is detected.

Advantage of the present invention is: to little image block, this method has clear superiority than existing algorithm aspect detection accuracy and the practicality.In addition, the present invention is used for the characteristic dimension-reduction treatment with principal component analysis, and under the prerequisite that descends significantly of intrinsic dimensionality, it is more stable to guarantee to detect accuracy rate.

Description of drawings

Fig. 1 is an image block feature extraction FB(flow block) of the present invention;

Fig. 2 is a SVMs training and testing FB(flow block) of the present invention;

Fig. 3 is that synthetic jpeg image copy-paste of the present invention is distorted the testing process block diagram;

Fig. 4, Fig. 5 are that the synthetic jpeg image of two width of cloth is distorted the synoptic diagram that the back is detected accurately and located by this method.

Embodiment

Below in conjunction with accompanying drawing and instantiation technical scheme of the present invention is elaborated.

The unpressed coloured image of the widely used in the world UCID of the present invention is implemented under the storehouse, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following technical scheme.

The UCID image library has comprised the tiff format image of 1338 width of cloth without the JPEG compression, can pass through http://www-staff.lboro.ac.uk/～cogs/datasets/UCID/ucid.html and obtain.

(1) foundation in experimental image piece sample storehouse

The detailed process that the present invention sets up experimental image piece sample storehouse is:

1. adopt the image block (promptly 128 * 128 and 256 * 256) of two kinds of sizes that this method is assessed;

2. JPEG compressed article prime factor QF ₁And QF ₂Span is [50,90], and step-length gets 10;

3. particularly, first 1000 unpressed images of picked at random in the UCID image library, then according to each to { QF ₁, QF ₂, (quality factor successively is QF to produce 1000 pairs of two image blocks that compress of displacement JPEG respectively ₁And QF ₂) and the JPEG list compresses, and (quality factor is QF ₂) image block.

(2) detection of the two compressed picture blocks of displacement JPEG

With reference to Fig. 1, Fig. 2, the support vector machine classifier among the present invention is selected LIBSVM for use, can pass through http://www.csie.ntu.edu.tw/～cjlin/libsvm/ and obtain.Kernel function is selected radially basic kernel function (RBF).For given image block size and quality factor q F ₂, make QF ₁50,60,70,80, and evenly choose at random among the 90}, according to respective image piece sample set in the experimental image piece sample storehouse, the detailed process that detects the two compressed picture blocks of displacement JPEG is following:

1. sample is divided into two sub-set, promptly from two compressed picture blocks of 1000 width of cloth displacement JPEG and the single compressed picture blocks of 1000 width of cloth JPEG, distinguishes picked at random 500 width of cloth as training sample database, remaining as the test sample book storehouse.

2. the feature extraction detailed process of each width of cloth image block in training sample database and the test sample book storehouse is following:

The first step: after the decompression of jpeg image piece, obtain brightness (Y) and colourity (C _bAnd C _r) the two-dimentional JPEG matrix of coefficients X (referring to the two-dimensional matrix that discrete cosine transform coefficient amplitude that JPEG quantizes is formed) of component.

Second step: the two-dimentional JPEG matrix of coefficients X of Y component is done four direction difference (being the secondary diagonal angle of level, vertical, main diagonal sum) processing, obtain four two-dimentional JPEG coefficient difference sub matrixs:

E _h(i,j)=X(i,j)-X(i+1,j)

E _v(i,j)=X(i,j)-X(i,j+1)

E _d(i,j)=X(i,j)-X(i+1,j+1)

E _m(i,j)=X(i+1,j)-X(i,j+1)

E wherein _h, E _v, E _dAnd E _mThe two-dimentional JPEG coefficient difference sub matrix of representing level, the secondary diagonal of vertical, main diagonal sum respectively;

The 3rd step: four two-dimentional JPEG coefficient difference sub matrix E that will obtain _h, E _v, E _dAnd E _mDoing thresholding handles:

$E(i,j)=\left\{\begin{array}{cc}E(i,j)& \left|E(i,j)\right|<T\\ T& E(i,j)\&GreaterEqual;T\\ -T& E(i,j)\&le;-T\end{array}\right.$

T=2 wherein;

The 4th step: adopt condition symbiosis probability matrix model that four two-dimentional JPEG coefficient difference sub matrixs of thresholding are carried out modeling, the condition symbiosis probability matrix of level, the secondary diagonal of vertical, main diagonal sum is following:

$p\left\{(E_h(i+1,j)=m,E_h(i+2,j)=n)\right|E_h(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l,E_h(i+1,j)=m,E_h(i+2,j)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l)}$

$p\left\{(E_v(i,j+1)=m,E_v(i,j+2)=n)\right|E_v(i,j)=l\}=\frac{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l,E_v(i,j+1)=m,E_v(i,j+2)=n))}{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l)}$

$p\left\{(E_d(i+1,j+1)=m,E_d(i+2,j+2)=n)\right|E_d(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l,E_d(i+1,j+1)=m,E_d(i+2,j+2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l)}$

$p\left\{(E_m(i+1,j-1)=m,E_m(i+2,j-2)=n)\right|E_m(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l,E_m(i+1,j-1)=m,E_m(i+2,j-2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l)}$

Wherein, M and N are respectively the level and the vertical direction sizes of the two-dimentional JPEG coefficient difference sub matrix of thresholding.L, m, n ∈ Z;-T≤l; M, n≤T, and

is therefore; The element of four direction condition symbiosis probability matrix is as characteristic, and its dimension is 500 dimensions;

3. training sample database characteristic and test sample book Al Kut are levied data normalization to [1,1] interval;

4. utilize PCA that training sample and test sample book are carried out dimension-reduction treatment, the intrinsic dimensionality d=250 behind the dimensionality reduction;

5. train the svm classifier device with the feature set of the training sample database after normalization and the dimension-reduction treatment.Wherein, radially the optimized parameter of basic kernel function (c is to obtain through cross validation and grid search algorithm g);

6. utilize the SVM that trained, the test set after normalization and the dimension-reduction treatment is classified, record detects accuracy.In order to eliminate the influence of randomness, above-mentioned experiment is repeated to carry out 20 times, detects the mean value that accuracy is got 20 experiments.

(3) synthetic jpeg image copy-paste is distorted detection

1. with reference to Fig. 3, at first confirm the image block of a B * B size, B is 8 multiple.Image block size B can not too greatly can not be too little, and accurately positioning tampering is regional for the too big algorithm of B if this is, the too little algorithm of B can not provide a good statistic property.Choose 128 * 128 image block in the experiment as the matrix size;

2. according to experimental image piece storehouse, 128 * 128 matrix size and JPEG compressed article prime factor QF ₂, make QF ₁{ 50,60,70; 80, picked at random among the 90} is utilized corresponding training set characteristic training SVM; The SVM that trains is obtained 1% wrong positive routine rate through the adjustment decision-making value, and (FP, i.e. the image block of the single compression of JPEG is by the ratio that is categorized into the two image blocks that compress of displacement JPEG of mistake.）；

3. use the sub-piece of 128 * 128 sizes from left to right and from top to bottom to travel through given synthetic jpeg image with the interval of 32 pixels; Obtain the sub-piece of 128 * 128 sizes to be detected successively; Extract the characteristic of the sub-piece of each 128 * 128 size to be detected, and through normalization and PCA dimensionality reduction.Then; (image block of the two compressions of displacement JPEG is marked as " 1 " with the SVM that trained every sub-block to be carried out classification and marking successively; The image block of the single compression of JPEG is marked as " 0 "); Mark can obtain a two-value decision matrix (BDM) after finishing all image blocks to be detected;

5. using window size is that 3 * 3 median filter carries out filtering to above-mentioned two-value decision matrix, its objective is in order further to remove the single compressed picture blocks of isolated JPEG that some possibly be mistaken for the two compressions of displacement JPEG.So, can obtain filtered two-value decision matrix (FBDM);

6. according to the value of FBDM, synthesizing the tampered region that jpeg image relevant position mark is detected.

(4) testing result

(1) the two compressed picture blocks testing results of displacement JPEG

The image block size is 128 * 128 and 256 * 256 o'clock; (TP representes correct positive routine rate based on the detection accuracy (TP+TN)/2 of the two compressed picture blocks of the displacement JPEG of condition symbiosis probability matrix; TN representes correct counter-example rate; Positive example refer to the to be shifted image blocks of the two compressions of JPEG, counter-example refers to the single compressed picture blocks of JPEG) respectively shown in table 1 and 2.In order to compare with existing two kinds of two compressed picture blocks detection methods of displacement JPEG, with they emulation on the image block storehouse that this method is set up, the result shows, when the image block size is 128 * 128, and QF ₁And QF ₂Under the different value condition, this method has improved about 3.2%-40.7% than the method for Luo etc., has improved 2.7%-30.5% than the method for Qu etc.; When the image block size was 256 * 256, this method had improved 3.6%-55.7% than the method for Luo etc., had improved 1.8%-37.5% than the method for Qu etc.

Table 1 image block size is 128 * 128, QF ₁And QF ₂Detection accuracy (%) under the different value condition

?QF 1\QF 2	50	60	70	80	90
						50	60.5	65.2	72.1	82.3	92.0
60	59.9	63.4	68.8	79.1	91.2
						70	57.7	60.6	64.6	74.4	87.5
80	54.2	56.1	58.5	63.2	79.7
						90	51.8	51.7	52.3	52.2	58.1

Table 2 image block size is 256 * 256, QF ₁And QF ₂Detection accuracy (%) under the different value condition

QF 1＼QF 2	50	60	70	80	90
						50	72.0	76.3	83.6	91.0	97.5
60	69.8	74.4	80.6	89.0	97.5
						70	65.1	69.5	75.3	84.1	96.4
80	57.9	61.0	63.5	72.8	91.5

90

52.6

53.6

53.0

53.4

61.1

(2) synthetic jpeg image copy-paste tampering location

With reference to Fig. 4, Fig. 5, detect and the regional validity of positioning tampering in order to test this method, the example that two synthetic jpeg images are distorted detection is shown in Figure 4 and 5.Original image all is unpressed tiff format images, from the NRCS image library, can obtain through http://photogallery.nrcs.usda.gov/.The tampered region of synthetic jpeg image is experience quality factor q F ₁And QF ₂The two compressions of displacement JPEG, remainder only experiences quality factor q F ₂The single compression of JPEG.

Fig. 4 (a) and the synthetic jpeg image of Fig. 5 (a) expression two width of cloth, the actual tampered region of region representation that the black outline line in Fig. 4 (b) and Fig. 5 (b) expression is surrounded, Fig. 4 (c) and Fig. 5 (c) expression compressed article prime factor { QF ₁, QF ₂}={ 50, the tampered region of the synthetic jpeg image of 90} is detected and positioning result, Fig. 4 (d) and Fig. 5 (d) expression compressed article prime factor { QF ₁, QF ₂}={ 50, the tampered region of the synthetic jpeg image of 80} is detected and positioning result.Can find out that from Fig. 4 and Fig. 5 this method can detect and the positioning tampering zone exactly.

More than detect and show that to little image block, this method has clear superiority than existing method, utilize little image block to synthesize jpeg image and carry out the tampered region detection and positioning result also more accurate.

Claims (6)

1. blind checking method is distorted in two compressions based on the displacement JPEG of condition symbiosis probability matrix, and it is characterized in that: step comprises: the magnitude matrix of the discrete cosine transform coefficient that at first JPEG is quantized is carried out level, the secondary diagonal angle four direction difference of vertical, main diagonal sum and thresholding and is handled; Then, adopt condition symbiosis probability matrix that the difference matrix of these four thresholdings is carried out modeling, the element of choosing condition symbiosis probability matrix is as characteristic, through normalization and principal component analysis (PCA) dimension-reduction treatment; Whether adjudicate image block through the two compressions of displacement JPEG through the training and the detection of SVMs (SVM) at last.

2. blind checking method is distorted in the two compressions of the displacement JPEG based on condition symbiosis probability matrix according to claim 1, and it is characterized in that: the characteristic of said colored jpeg image piece is extracted, and may further comprise the steps:

1. after colored jpeg image piece being decompressed, obtain brightness (Y) and colourity (C _bAnd C _r) the two-dimentional JPEG matrix of coefficients X of component, refer to the two-dimensional matrix that discrete cosine transform coefficient amplitude that JPEG quantizes is formed;

2. the two-dimentional JPEG matrix of coefficients X with the Y component does level, the secondary diagonal angle of vertical, main diagonal sum four direction difference processing, obtains four two-dimentional JPEG coefficient difference sub matrixs:

E _h(i,j)=X(i,j)-X(i+1,j)

E _v(i,j)=X(i,j)-X(i，j+1)

E _d(i,j)=X(i,j)-X(i+1,j+1)

E _m(i,j)=X(i+1,j)-X(i,j+1)

E wherein _h, E _v, E _dAnd E _mThe two-dimentional JPEG coefficient difference sub matrix of representing level, the secondary diagonal of vertical, main diagonal sum respectively;

Four two-dimentional JPEG coefficient difference sub matrix E that 3. will obtain _h, E _v, E _dAnd E _mCarrying out thresholding handles:

$E(i,j)=\left\{\begin{array}{cc}E(i,j)& \left|E(i,j)\right|<T\\ T& E(i,j)\&GreaterEqual;T\\ -T& E(i,j)\&le;-T\end{array}\right.$

T=2 wherein;

4. utilize condition symbiosis probability matrix that four two-dimentional JPEG coefficient difference sub matrixs of thresholding are carried out modeling, the condition symbiosis probability matrix of level, the secondary diagonal of vertical, main diagonal sum is following:

$p\left\{(E_h(i+1,j)=m,E_h(i+2,j)=n)\right|E_h(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l,E_h(i+1,j)=m,E_h(i+2,j)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_h(i,j)=l)}$

$p\left\{(E_v(i,j+1)=m,E_v(i,j+2)=n)\right|E_v(i,j)=l\}=\frac{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l,E_v(i,j+1)=m,E_v(i,j+2)=n))}{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_v(i,j)=l)}$

$p\left\{(E_d(i+1,j+1)=m,E_d(i+2,j+2)=n)\right|E_d(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l,E_d(i+1,j+1)=m,E_d(i+2,j+2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-3}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_d(i,j)=l)}$

$p\left\{(E_m(i+1,j-1)=m,E_m(i+2,j-2)=n)\right|E_m(i,j)=l\}=\frac{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l,E_m(i+1,j-1)=m,E_m(i+2,j-2)=n))}{\underset{i=0}{\overset{M-3}{\mathrm{\&Sigma;}}}\underset{j=2}{\overset{N-1}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(E_m(i,j)=l)}$

Wherein, M and N are respectively the level and the vertical direction sizes of the two-dimentional JPEG coefficient difference sub matrix of thresholding, l; M; N ∈ Z ,-T≤l, m; N≤T, and

3. blind checking method is distorted in the two compressions of the displacement JPEG based on condition symbiosis probability matrix according to claim 1, it is characterized in that:

Described principal component analysis characteristic dimensionality reduction makes that eigenmatrix is C _X, its size is m ₁* n ₁, proper vector of each line display of matrix, therefore, intrinsic dimensionality is m ₁, the PCA dimensionality reduction is intended to find transformation matrix W, obtains new feature C _Y=W ^T(C _X-μ _X), make new feature C _YComposition with better classification capacity, its step comprises:

1. calculate the covariance matrix ∑ _X:

${\mathrm{\&Sigma;}}_X=\frac{1}{n_1}\left((C_X-{\mathrm{\&mu;}}_X){(C_X-{\mathrm{\&mu;}}_X)}^T\right)$

Wherein, μ _XBe eigenvectors matrix C _XAverage, ∑ _XIt is symmetric matrix;

2. through finding the solution secular equation

det(λ _iI-∑ _X)=0

Obtain eigenvalue _i(1≤i≤m ₁), the eigenwert of being asked constitutes the diagonal covariance matrix ∑ _Y

3. to each eigenvalue _i(1≤i≤m ₁), according to (λ _iThe I-∑ _X) w _i=0 finds the solution corresponding w _iThereby, obtaining proper vector, proper vector should be normalization and linear independence;

4. eigenwert is sorted, choose bigger preceding d (＜m ₁) the pairing proper vector composition of individual eigenwert transformation matrix W;

5. new feature Matrix C _Y=W ^T(C _X-μ _X).

4. blind checking method is distorted in the two compressions of the displacement JPEG based on condition symbiosis probability matrix according to claim 1, it is characterized in that:

Described SVMs training and testing process may further comprise the steps:

1. detection, quality factor q F are distorted in two compressions for the displacement JPEG of image block ₁Generally be unknown, suppose QF ₁50,60,70,80, evenly choose quality factor q F among the 90} at random ₂Can from the jpeg image header file, calculate; According to quality factor q F ₁And QF ₂, use the size that comprises of known class to be QF as B * B, quality factor ₂Single compressed picture blocks of JPEG and quality factor be QF ₁And QF ₂The two compressed picture blocks composing training collection of displacement JPEG; Each image block to training set extracts above-mentioned condition symbiosis probability matrix characteristic; The characteristic that obtains is through normalization and PCA dimension-reduction treatment; And indicate classification under its image block (it is the single compressed picture blocks of JPEG to zone bit for-1 expression, the two compressed picture blocks of 1 expression displacement JPEG) with zone bit;

2. svm classifier device kernel function is selected radially basic kernel function for use, and the training set characteristic is sent into SVM, utilizes grid search and cross validation, obtains cross validation accuracy rate the highest optimized parameter c and g, and trains SVMs with them;

3. extract the condition symbiosis probability matrix characteristic of test pattern piece, the characteristic that obtains is through normalization and PCA dimension-reduction treatment, and judges classification under it with the SVM that trained.

5. blind checking method is distorted in the two compressions of the displacement JPEG based on condition symbiosis probability matrix according to claim 4, it is characterized in that: detection is distorted in the two compressions of the displacement JPEG of said image block, may further comprise the steps:

1. from the header file of synthetic jpeg image, calculate quality factor q F ₂

2. according to the matrix of B * B size and quality factor q F ₂, make QF ₁{ 50,60,70; 80, evenly choose at random among the 90}, utilize said method training SVM; The SVM that trains is obtained 1% wrong positive routine rate (FP refers to that the image block of the single compression of JPEG is by the ratio that is categorized into the two image blocks that compress of displacement JPEG of mistake) through adjusting decision-making value;

3. use the big or small sub-piece of B * B from left to right and from top to bottom to travel through given synthetic jpeg image with the interval of 32 pixels; Obtain the sub-piece of B to be detected * B size successively; Extract the characteristic of the sub-piece of each B to be detected * B size, and through normalization and PCA dimension-reduction treatment;

Then; (image block of the two compressions of displacement JPEG is marked as " 1 " with the SVM that trains every sub-block to be carried out classification and marking successively; The image block of the single compression of JPEG is marked as " 0 "); Behind the intact all image blocks to be detected of mark, can obtain a two-value decision matrix (binary decision matrix, BDM);

4. using window size is that 3 * 3 median filter carries out filtering to above-mentioned two-value decision matrix; Further remove the single compressed picture blocks of isolated JPEG that some possibly be mistaken for the two compressions of displacement JPEG; Can obtain filtered two-value decision matrix (filtered binary decision matrix, FBDM);

5. according to the element value of FBDM, synthesizing the tampered region that jpeg image relevant position mark is detected.

6. distort blind checking method according to claim 1 or the two compressions of 2 or 4 described displacement JPEG, it is characterized in that: image block size B=128 * 128 and B=256 * 256 based on condition symbiosis probability matrix.

Images