CN102509257A - Human visual characteristic compressive sensing-based grayscale image tampering and detection method - Google Patents

Abstract

The invention relates to a human visual characteristic compressive sensing-based grayscale image tampering and detection method, which comprises the following steps of: (1) generation and transmission of a watermark of an image: prior to the transmission of the image, a sending terminal firstly generates a Hash value of the image by using a human visual characteristic compressive sensing-based sparse base change and measurement matrix, then, transmits the Hash value through a secure channel, and transmits the image to a receiving terminal through a public channel; (2) image detection of the receiving terminal: the receiving terminal verifies whether the image obtained from the public channel is tampered or not by using the watermark of the image received from a trusted channel; and (3) tamper localization of the receiving terminal: after the condition that the received image is maliciously tampered is judged by the receiving terminal, a compressive sensing orthogonal matching pursuit reconstruction method is further used for restoring a difference value D, and therefore, an image E of a tampered position is obtained. Compared with the prior art, the human visual characteristic compressive sensing-based grayscale image tampering and detection method has the advantages of high rate in detection, low quantity in transmission, fastness in computation and the like.

Description

Gray level image altering detecting method based on the human visual system compressed sensing

Technical field

The present invention relates to technical field of image processing, especially relate to a kind of gray level image altering detecting method based on the human visual system compressed sensing.

Background technology

It is a kind of copyright protection technology of digital product that distorted image detects.Its core concept is: some identification informations (being digital watermarking) are directly embedded in the secondary indication (as revising the structure of specific region) of digital carrier (like multimedia, document, software etc.) or Digital Media; And do not influence the use value of original vector; Also be not easy to be found out and revise once more, but can be by producer identification and identification.Be hidden in the information in the carrier through these, can reach and confirm creator of content, buyer, transmission secret information or judge whether carrier purpose such as is distorted.Digital watermarking is a research direction of Information Hiding Techniques, also is the effective way that realizes copyright protection.

Compressed sensing is the signals collecting of a novelty and the theory of processing, and it is made up of sparse basic conversion, measurement matrix and signal reconstruction three parts.The basis of this theory is if signal is compressible or is sparse at certain transform domain; So just can use one with higher-dimension signal projection to the lower dimensional space of the incoherent observing matrix of transform-based with the conversion gained on, from the projection of these lower dimensional spaces, reconstruct original signal through finding the solution an optimization problem then with high probability.

Find through literature search; People such as M.Tagliasacchi, G.Valenzise and S.Tubaro are at paper " Localization of sparse image tampering via random projections " (" based on the discrete picture tampering location of stochastic sampling ") (IEEE International Conference on Image Processing; Pp.2092-2095 Oct.2008) has proposed the method that a kind of distorted image detects in (international conference of IEEE Flame Image Process).Wherein, transmitting terminal carries out accidental projection to original image information, and the compression measured value that obtains is carried out behind the distributed coding afterwards this Hash eigenwert being sent to Verification System through safe lane as the Hash eigenwert of original image again.Verification System according to the image Hash eigenwert that is received by safe lane, judges whether this image was distorted after receiving image information.If receive and distort attack, and tampered position is sparse on spatial domain, the reconstructing method positioning tampering information in the compressed sensing theory then capable of using.Emulation experiment shows that this method not only can reduce computing and transmission cost, also has and distorts detectability preferably.The outstanding feature of above-mentioned distorted image detection scheme is to have adopted compressed sensing theoretical.Because compressed sensing does not need scanning picture in high-frequency ground just can write down and recover this width of cloth picture, therefore need have remarkable advantages to the application that the view picture picture scans and handles like this for digital watermarking afterwards.

But the adaptive disposal route of compressed sensing right and wrong needs when specifically using to select different sparse bases, measure matrix and reconfiguration scheme according to different scene, this to a great extent limit the application performance of compressed sensing; Simultaneously, because how the human visual perception characteristic uses the responsive zone of human eye perception with Limited resources to greatest extent, also be the great approach of lifting scheme performance.This patent emphasis is proposing a kind of improved image watermark method based on compressed sensing.

Find through the document retrieval in addition,, then can effectively promote the recovery effects of image if human visual system is introduced in the compressed sensing.YingYu; People such as BinWang are at paper " Saliency-Based Compressive Sampling for image signals [C] " (" based on the compression of images perception studies of area-of-interest ") (IEEE SIGNAL PROCESSING LETTERS; Vol.17; No.11, pp.973-976 Sep.2010) points out in (IEEE signal Processing proceedings) that the responsive marking area of human vision can be broadly described as one group of statistical information.The compression of images perception framework based on area-of-interest that the document proposes mainly comprises 3 parts: at first image is analyzed and generated an area-of-interest mapping table; According to this mapping table,, calculate the compressed sensing measurement parameter of each piece successively then, generate and measure matrix, carry out the compressed sensing sampling the picture piecemeal; Use the compressed sensing recovery algorithms to carry out image reconstruction at last.Area-of-interest mapping table in the scheme is the core of entire method, and can the quality of this mapping table have determined distribute rational resource to give the emphasis part of human vision, therefore determined the quality of image recovery effects.

The focusing on of this patent is applied to gray level image with above-mentioned compression of images aware scheme based on area-of-interest and distorts in the testing process, detects effect with further reinforcement distorted image.

Summary of the invention

The object of the invention is exactly the gray level image altering detecting method based on the human visual system compressed sensing that a kind of high detection rate, the throughput rate of passing at the low, fast computing are provided for the defective that overcomes above-mentioned prior art existence.

The object of the invention can be realized through following technical scheme:

A kind of gray level image altering detecting method based on the human visual system compressed sensing is characterized in that, may further comprise the steps:

1) image watermark generates and transmission, and transmitting terminal at first generated the cryptographic hash of this image before images, through safe lane transmission cryptographic hash, gives receiving end through the overt channel images then;

2) receiving end image detection, the image watermark that the receiving end utilization is received by trusted channel, whether the image that checking is obtained by overt channel has passed through is distorted;

3) receiving end tampering location; After receiving end is judged received image and is maliciously tampered, further utilize orthogonal matching pursuit (Orthogonal Matching Pursuit, the OMP) reconstructing method of compressed sensing; D recovers with difference, thereby obtains tampered position image E.

Described step 1) is specific as follows:

11) generating the area-of-interest mapping graph of image, specifically is that the dct transform gene polyadenylation signal to image carries out standardization;

For a given gray-scale map X ∈ R ^{N * N}, wherein n=N * N is total pixel of image, it is significantly schemed Map and is produced by following process:

P＝sign(C(X))

F＝abs(C ^-1(P))

Map＝G*F ²

Wherein C (X) and C ^-1(P) two-dimensional dct transform and the inverse transformation thereof of presentation video respectively, sign (C (X)) is-symbol function, abs (C ^-1(P)) be the function that takes absolute value, G is two-dimentional gauss low frequency filter, and the element of significantly scheming among the Map is designated as Map _i(1≤i≤n);

12) image is carried out piecemeal and handle, at first, the fixed size that segments the image into non-overlapping copies is the piecemeal of m=B * B, and establishing i piecemeal is B _i, wherein, 1≤i≤n/m supposes that n can be divided exactly by m, and each piecemeal i is calculated:

${\mathrm{\ξ}}_i=\frac{1}{m}{\mathrm{\Σ}}_{j\∈B_i}{\mathrm{Map}}_j$

13) by ξ _iCalculate arbitrary piecemeal B _i(the measurement number of times M in compressed sensing of 1≤i≤n/m) _iFor:

$M_i=\mathrm{rnd}({\mathrm{\θ}}_{\max }-\frac{{\mathrm{\ξ}}_i}{255}\×({\mathrm{\θ}}_{\max }-{\mathrm{\θ}}_{\min }))$

Wherein, θ _MaxAnd θ _MinThe maximal value and the minimum value of element in the matrix measured in expression respectively, and rnd () gets and the immediate integer of input variable;

Then, with the measured value M of all piecemeals _i(1≤i≤n/m) be recorded in the measurement number of times vector M, i.e. M=(M ₁, M ₂..., M _N/m);

14) image is carried out gray average, the steps include:

At first, the fixed size that former figure is divided into non-overlapping copies is that (dB≤B) supposes that m can be divided exactly each piecemeal B by t for the piecemeal of t=dB * dB _i(1≤i≤n/m) be divided into m/t fritter again is designated as dB _Ij(1≤i≤n/m, 1≤j≤m/t);

Then, calculate each piecemeal dB _Ij(the gray average g of all pixels among the 1≤i≤n/m, 1≤j≤m/t) _Ij(1≤i≤n/m, 1≤j≤m/t), and all gray values of pixel points in this piecemeal are set at this gray average;

At last, with piecemeal B _i(each sub-piece dB among 1≤i≤n/m) _Ij(gray average of 1≤j≤m/t) is stored in vectorial g _i=(g _{I, 1}g _{I, 2}... g _{I, j}... g _{I, m/t}) in, 1≤i≤n/m;

15) choose random seed S, image X is carried out the piecemeal compressed sensing calculate, this piecemeal compressed sensing is calculated and is meant each sub-piece B _i(1≤i≤n/m) carry out respectively calculating based on the compressed sensing of human visual system; Afterwards, again the compressed sensing result of calculation of each sub-piece is pressed its sequential combination in former figure X together, thereby obtain the compressed sensing measured value of image X;

Described piecemeal B _i(compressed sensing based on human visual system of 1≤i≤n/m) is calculated, and the steps include:

At first, to B _iCarry out wavelet transformation, obtain the sparse signal on the wavelet field

$x_{B_i}={\mathrm{\Ψ}}^T{B}_i$

Wherein, is the wavelet transformation base;

Secondly, be M according to S generation scale _iThe gaussian random of * m is measured matrix Φ _i, wherein m is piecemeal B _iPixel quantity, M _iBe the piecemeal B that obtains in the step 13) _iThe measurement number, the element in the said gaussian random matrix all produces at random, and its probability meets gaussian distribution nature;

At last, with sparse signal

Project in this measurement matrix, obtain Observed reading Y _i, that is:

$Y_i={\mathrm{\Φ}}_i{x}_{B_i}={\mathrm{\Φ}}_i{\mathrm{\Ψ}}^T{B}_i$

After all piecemeals have all carried out the compressed sensing measurement, with the compressed sensing result of calculation Y of each sub-piece _iBy its sequential combination in former figure X together, the compressed sensing that has just obtained image X is measured Y;

In above-mentioned compressed sensing computation process, utilize less, the compressible characteristics of image nonzero element coefficient in wavelet transformed domain, each image block B based on human visual system _iBe reduced to M from original m dimensional vector _iDimensional vector has reduced the storage data volume effectively; Simultaneously, the measurement number of times M of different piecemeals _iUtilize the remarkable characteristic of human vision to confirm; The vision signal portion can be given bigger measurement number of times, and the measurement number of times of the insensitive part of vision is then less, like this can the reasonable distribution computational resource; Both reduced the computing cost, for distorted image detects basis preferably was provided again;

16) use the LDPC lossy coding that the compressed sensing of image X is measured Y and encode, obtain the cryptographic hash Hash (X) of image X; Its process is specially, respectively to each piecemeal B _iObserved reading Y _iCarry out the LDPC lossy coding, obtain compressed code S _iAfterwards, again with each S _iConnect according to the order of sequence, the symbol string that obtains is as cryptographic hash Hash (X)=(S of image X ₁S ₂... S _N/m), 1≤i≤n/m;

17) watermark of image X generates and credible transmission, and the watermark watermark (X) of said image X is exactly by Hash (X), random seed S and measures the tlv triple that the number of times vector M is formed, promptly

watermark(X)＝(Hash(X)，S，M)

Transmitting terminal is calculating watermark (X) afterwards, and the trusted channel through itself and receiving end sends watermark (X) to receiving end, through overt channel issue image X.

Described LDPC lossy coding is according to the distributed coding theory of lossy compression method signal to be carried out compressed encoding, is specially, for compressed signal W=(w ₁w ₂... w _k), get its each element w _i(the plurality of continuous position bit of 1≤i≤k), the initial bits position of establishing the bit of getting is p, stopping bit is q, and with the interposition Bit String according to each element w _i(1≤i≤k) order in signal W connects successively, and the new Bit String that obtains is exactly the LDPC coding of signal W.

To twice minute block message: B, the dB of original image, and the initial bits position p of LDPC coding with stop bit q and can arrange in advance or send receiving end to through trusted channel.

Described step 2) receiving end image detection, the image watermark that the receiving end utilization is received by trusted channel, whether the image that checking is obtained by overt channel has passed through to distort and has been specially:

21) adopt the mode identical to receiving image X ' ∈ R with transmitting terminal ^{N * N}Carry out gray average, the steps include:

At first, the fixed size that X ' is divided into non-overlapping copies is the piecemeal B of m=B * B _i', 1≤i≤n/m; And then with each B _i' the size that is divided into non-overlapping copies is the piecemeal dB of t=dB * dB _Ij', 1≤i≤n/m, 1≤j≤m/t;

Secondly, calculate each piecemeal dB _Ij' in the gray average g of all pixels _Ij', and all gray values of pixel points in this piecemeal are set at this gray average;

At last, with piecemeal B _iIn each sub-piece dB _IjGray average be stored in vectorial g _i'=(g _{I, 1}' g _{I, 2}' ... g _{I, j}' ... g _{I, m/t}') in, 1≤i≤n/m, 1≤j≤m/t;

22) adopt the mode identical to receiving image X ' ∈ R with transmitting terminal ^{N * N}Carrying out the piecemeal compressed sensing calculates;

For piecemeal B _i', 1≤i≤n/m at first carries out wavelet transformation, obtains the sparse signal on the wavelet field

${x_{B_i}}^{\′}={\mathrm{\Ψ}}^T{B}_i$

Wherein, is the wavelet transformation base;

Secondly, according to random seed S that is received by trusted channel and mapping graph M, the generation scale is M _iThe gaussian random of * m is measured matrix Φ _i

At last, with sparse signal

Project in this measurement matrix, obtain Observed reading Y _i', that is:

${Y_i}^{\′}={\mathrm{\Φ}}_i{x_{B_i}}^{\′}={\mathrm{\Φ}}_i{\mathrm{\Ψ}}^T{B_i}^{\′}$

After all piecemeals have all carried out the compressed sensing measurement, with the compressed sensing result of calculation Y of each sub-piece _i' by its sequential combination in former figure X ' together, the compressed sensing that has just obtained image X ' is measured Y '=(Y ₁' Y ₂' ... Y _N/m');

23) image cryptographic hash Hash (X)=(S to receiving ₁S ₂... S _N/m), 1≤i≤n/m carries out the distributed decoding of the LDPC "=(Y that obtains Y ₁" Y ₂" ... Y _N/m"), that is, initial bits position p when encoding according to LDPC and termination bit q are with S _iAs Y _i" in the bit value of p to the q position of each element; Y _i" in all the other everybody by Y _i' correspondence everybody fill up successively;

24) distorted image detects, and calculates Y " and the difference of Y ':

D＝Y″-Y′

If D=0 explains that then the image of receiving end reception is consistent with the image that transmitting terminal is sent, do not distorted; If D ≠ 0, then presentation video was distorted, and carried out further tampering location, to point out in which position of image distorting.

Described step 3) receiving end tampering location is specially:

31) initializaing variable is set:

A) maximum iteration time Iter and acceptable error threshold e are set;

B) initialization Increment Matrix Aug_t is a null matrix, scale and the step 24 of matrix A ug_t) in the matrix of differences D that calculates identical;

C) residual values error initial value being set is error threshold e; Iterations time=0 is set;

D) temporary variable product being set is the matrix that has identical size with D;

32) with matrix product zero clearing, be changed to null matrix; Time=time+1;

33) recomputate the procuct matrix:

product ⁱ＝abs(D ⁱgerror)

Product wherein ⁱThe i row of expression product matrix, D ⁱBe the i row of D, abs is an ABS function; Be each column vector of calculated difference matrix D and the inner product of residual values error, with its column vector as product matrix respective column;

34) confirm the maximal value val of each matrix element of product and the row pos at place thereof

35) expand Increment Matrix Aug_t, with product ^PosRow are added in the former Aug_t matrix as last row, obtain the Aug_t matrix after epicycle expands, that is: [Aug_t, product ^Pos] → Aug_t;

36) with the pos row zero setting of matrix of differences D;

37) calculate recovery matrix A ug_y:

Aug_y＝(Aug_t′×Aug_t) ^-1×Aug_t′×e

Wherein, the transposed matrix of Aug_t ' expression Increment Matrix Aug_t; (Mat) ^-1The inverse matrix of representing matrix Mat;

38) each matrix element among the Aug_t connected line by line obtain the vectorial L.Aug_t of row, each matrix element among the Aug_y is connected line by line obtain the vectorial L.Aug_y of row equally, calculate the residual values error of epicycle iteration then:

error＝e-L.Aug_t×L.Aug_y′

Wherein, L.Aug_y ' expression is by the column vector that obtains of the vectorial L.Aug_y transposition of row;

39) judge that whether error＜e satisfies, if satisfy, then goes to step 310);

Otherwise, at first judge whether time=Iter, be then to go to step 310), otherwise return step 32) the continuation iterative computation;

310) algorithm finishes, and the recovery matrix A ug_y that obtains i.e. tampering location figure E for being recovered by D.

Compared with prior art, the present invention owing to selected resume speed compressed sensing restoration methods faster, and has been merged the human vision characteristics of cognition except on transinformation, reducing greatly, and its performance also is able to guarantee; Because the reliability of compressed sensing itself, a whole set of verification and measurement ratio that detects based on the distorted image of visual cognition compressed sensing also is able to ensure.

Description of drawings

Fig. 1 is the former figure of the employed Bird of embodiment;

The Bird synoptic diagram that Fig. 2 distorts for the employed process of embodiment;

Fig. 3 is the location map of the resulting tampered position of embodiment;

Fig. 4 is the location map of the tampered position implementing common compressed sensing and obtain.

Embodiment

Below in conjunction with accompanying drawing and specific embodiment the present invention is elaborated.

Embodiment

What this instance adopted is that size is 512 * 512 Bird gray scale picture, is designated as X ∈ R ^{N * N}, N=512 wherein.

The first step, the area-of-interest mapping graph of generation image.Specifically be that dct transform gene polyadenylation signal to image carries out standardization, with the marking area of further outstanding image.

At first, P=sign (C (X)) obtains significantly figure Map by following three steps.

F＝abs(C ^-1(P))

$\mathrm{Map}=G*F^2\sqrt{a^2+b^2}$

Wherein C () and C ^-1() be the two-dimensional dct transform and the inverse transformation thereof of presentation video respectively, sign () is-symbol function, and abs () is the function that takes absolute value, G is two-dimentional gauss low frequency filter.Significantly the element among the figure Map is designated as Map _i(1≤i≤n).

Then, we segment the image into the piecemeal that does not have overlapping fixed measure 2 * 2, use B _iNote is made i piece piecemeal, wherein, and 0≤i＜65536.We can handle each piecemeal as follows so:

${\mathrm{\ξ}}_i=\frac{1}{m}{\mathrm{\Σ}}_{j\∈B_i}{\mathrm{Map}}_j$

Wherein, m=4,0≤j＜65536.

Note and since interest maps figure in the convolution process of the first step by standardization, so the brightness value of all pixels promptly satisfies 0≤ξ all in 0 to 255 scope _i≤255.

Afterwards, calculate the measured value M of each piecemeal i _i:

$M_i=\mathrm{rnd}({\mathrm{\θ}}_{\max }-\frac{{\mathrm{\ξ}}_i}{255}\×({\mathrm{\θ}}_{\max }-{\mathrm{\θ}}_{\min }))$

Wherein, 0≤i＜65536, θ _MaxAnd θ _MinRespectively expression the maximal value and the minimum value of receptible random measurement matrix measured value, be respectively 100 and 80 in this example.Function rnd is an immediate integer of getting input variable.

At last, with the measured value M of all piecemeals _i(1≤i≤65536) are recorded in and measure in the number of times vector M, i.e. M=(M ₁, M ₂..., M ₆₅₅₃₆).

In second step, image is carried out gray average.

At first, the fixed size that former figure is divided into non-overlapping copies is dB _IjPiecemeal, be 2 * 2 piecemeal in this instance.

Then, calculate each piecemeal dB _IjIn the gray average g of all pixels _Ij(1≤i≤65536 j=1), and are set at this gray average with all gray values of pixel points in this piecemeal.

At last, with the piecemeal B that obtains in the first step _iEach sub-piece dB in (1≤i≤65536) _Ij(1≤i≤65536, gray average j=1) is stored in vectorial g _i=(g _{I, 1}g _{I, 2}... g _{I, j}... g _{I, m/t}) in, 1≤i≤65536.In the present embodiment, because piecemeal B _iWith dB _IjSize is identical, therefore vectorial g _iOnly contain element, i.e. a g _i=(g _{I, 1}).

The 3rd step, choose random seed S, be used to generate the seed of random gaussian matrix.Image X is carried out the piecemeal compressed sensing to be calculated.In this patent, sparse base adopts the wavelet transformation base, and measuring matrix is that gaussian random is measured matrix.

To piecemeal B _iThe compressed sensing based on human visual system of (1≤i≤65536) is calculated, and the steps include:

At first, to B _iCarry out wavelet transformation, obtain the sparse signal on the wavelet field

$x_{B_i}={\mathrm{\Ψ}}^T{B}_i$

Wherein,

is the wavelet transformation base

Secondly, be M according to S generation scale _iThe gaussian random of * m is measured matrix Φ _i, M _iBe in the piecemeal B that obtains _iThe measurement number.Element in the gaussian random matrix all produces at random, and its probability meets gaussian distribution nature.

At last, with sparse signal

Project in this measurement matrix, obtain

Observed reading Y _i, that is:

$Y_i={\mathrm{\Φ}}_i{x}_{B_i}={\mathrm{\Φ}}_i{\mathrm{\Ψ}}^T{B}_i$

After all piecemeals have all carried out the compressed sensing measurement, with the compressed sensing result of calculation Y of each sub-piece _iBy its sequential combination in former figure X together, the compressed sensing that has just obtained image X is measured Y.

The 4th step, use the LDPC lossy coding that the compressed sensing of image X is measured Y and encode, obtain the cryptographic hash Hash (X) of image X.

Detailed process does, respectively to each piecemeal B _iObserved reading Y _iCarry out the LDPC lossy coding, promptly for compressed signal W=(w ₁w ₂... w _k), get its each element w _i(the 8th to the 18th bit in centre of 1≤i≤k), and with the interposition Bit String according to each element w _i(1≤i≤k) order in signal W connects successively, and the new Bit String that obtains is exactly the LDPC coding of signal W.

To Y _iCarry out obtaining compressed code S behind the LDPC lossy coding _iAfterwards, again with each S _iConnect according to the order of sequence, the symbol string that obtains is as cryptographic hash Hash (X)=(S of image X ₁S ₂... S _N/m), 1≤i≤n/m.

Finally, with watermark (X)=(Hash (X), S M) is transferred to Verification System as watermark through trusted channel.

In addition, to twice minute block message: B=2, the dB=2 of original image, and the initial bits position 8 of LDPC coding with stop bit 18 agreement in advance, send receiving end to through trusted channel.

In the 5th step, receiving end adopts the mode identical with transmitting terminal to receiving image X ' ∈ R ^{N * N}Carry out gray average.This example uses Fig. 2 under attack through the Bird synoptic diagram simulation picture of distorting.

At first, the fixed size that X ' is divided into non-overlapping copies is the piecemeal B of m=2 * 2 _i' (1≤i≤65536); And then with each B _i' the size that is divided into non-overlapping copies is the piecemeal dB of t=2 * 2 _Ij' (1≤i≤65536, j=1).

Secondly, calculate each piecemeal dB with the mode of transmitting terminal _Ij' in the gray average g of all pixels _Ij', and all gray values of pixel points in this piecemeal are set at this gray average.

At last, with piecemeal B _iEach sub-piece dB in (1≤i≤65536) _Ij(1≤i≤65536, gray average j=1) is stored in vectorial g _i'=(g _{I, 1}') in, 1≤i≤65536.

The 6th step is to image cryptographic hash Hash (X)=(S that receives ₁S ₂... S _N/m), the distributed decoding of the LDPC "=(Y that obtains Y is carried out in 1≤i≤65536 ₁" Y ₂" ... Y _N/m"), that is, initial bits position 8 when encoding according to LDPC and termination bit 18 are with S _iAs Y _i" in the 8th to the 18th bit value; Y _i" in all the other everybody by Y _i' correspondence everybody fill up successively.

In the 7th step, distorted image detects.Calculate Y " and the difference of Y ':

D＝Y″-Y′

If D=0 (representing null matrix here) explains that then the image of receiving end reception is consistent with the image that transmitting terminal is sent before, do not distorted; If D ≠ 0, then presentation video was distorted, and need further carry out tampering location, with which position of pointing out image was distorted.In this instance, the D non-zero, then presentation video is distorted, and needs further to use tampering location to come checking image by the part of malicious modification.

Described receiving end tampering location; Be meant after receiving end is judged received image and is maliciously tampered; Further utilize orthogonal matching pursuit (the Orthogonal Matching Pursuit of compressed sensing; OMP) signal reconfiguring method, D recovers with difference, thereby obtains tampered position image E.Concrete steps are following:

31) initializaing variable is set:

A) maximum iteration time Iter=128 and acceptable error threshold e=9 are set;

B) initialization Increment Matrix Aug_t is a null matrix, the scale of matrix A ug_t and 24) in the matrix of differences D that calculates identical;

C) residual values error initial value being set is error threshold e; Iterations time=0 is set;

D) temporary variable product being set is the matrix that has identical size with D;

32) with matrix product zero clearing, be changed to null matrix; Time=time+1;

33) recomputate procuct matrix: product ⁱ=abs (D ⁱGerror)

Product wherein ⁱThe i row of expression product matrix, D ⁱBe the i row of D, abs is an ABS function;

Be each column vector of calculated difference matrix D and the inner product of residual values error, with its column vector as product matrix respective column.

34) confirm the maximal value val of each matrix element of product and the columns pos at place thereof.

35) expand Increment Matrix Aug_t, with product ^PosRow are added in the former Aug_t matrix as last row, obtain the Aug_t matrix after epicycle expands, that is: [Aug_t, product ^Pos] → Aug_t;

36) with the pos row zero setting of matrix of differences D;

37) calculate recovery matrix A ug_y:

Aug_y＝(Aug_t′×Aug_t) ^-1×Aug_t′×e

Wherein, the transposed matrix of Aug_t ' expression Increment Matrix Aug_t; (Mat) ^-1The inverse matrix of representing matrix Mat.

38) each matrix element among the Aug_t connected line by line obtain the vectorial L.Aug_t of row, each matrix element among the Aug_y is connected line by line obtain the vectorial L.Aug_y of row equally, calculate the residual values error of epicycle iteration then:

error＝e-L.Aug_t×L.Aug_y′

Wherein, L.Aug_y ' expression is by the column vector that obtains of the vectorial L.Aug_y transposition of row.

39) judge that whether error＜e satisfies, if satisfy, then goes to 310);

Otherwise, at first judge whether time=Iter, be then to go to 310), otherwise return 32) the continuation iterative computation.

310) algorithm finishes, and the recovery matrix A ug_y that obtains i.e. tampering location figure E for being recovered by D.

Finally, the effect oriented of instance is as shown in Figure 3.Distort that to detect effect as shown in Figure 4 and document " Localization of sparse image tampering via random projections " adopts that common compression sensing method obtains.Can find out; The image recovery effects that adopts patent art to obtain is better than the result of conventional compression cognition technology; This mainly is that the measurement number of times of the insensitive part of vision is then less because present technique has been given more measurement number of times at the vision signal portion, like this can the reasonable distribution computational resource; Both reduced the computing cost, for distorted image detects basis preferably was provided again.

Claims (6)

1. the gray level image altering detecting method based on the human visual system compressed sensing is characterized in that, may further comprise the steps:

1) image watermark generates and transmission; Transmitting terminal is before images; At first utilize sparse basic conversion and the cryptographic hash of measuring this image of matrix generation, through safe lane transmission cryptographic hash, give receiving end then through the overt channel images based on the compressed sensing of human visual system;

2) receiving end image detection, the image watermark that the receiving end utilization is received by trusted channel, whether the image that checking is obtained by overt channel has passed through is distorted;

3) receiving end tampering location after receiving end is judged received image and is maliciously tampered, is further utilized the orthogonal matching pursuit reconstructing method of compressed sensing, and D recovers with difference, thereby obtains tampered position image E.

2. a kind of gray level image altering detecting method based on the human visual system compressed sensing according to claim 1 is characterized in that described step 1) is specific as follows:

11) generating the area-of-interest mapping graph of image, specifically is that the dct transform gene polyadenylation signal to image carries out standardization;

For a given gray-scale map X ∈ R ^{N * N}, wherein n=N * N is total pixel of image, it is significantly schemed Map and is produced by following process:

P＝sign(C(X))

F＝abs(C ^-1(P))

Map＝G*F ²

Wherein C (X) and C ^-1(P) two-dimensional dct transform and the inverse transformation thereof of presentation video respectively, sign (C (X)) is-symbol function, abs (C ^-1(P)) be the function that takes absolute value, G is two-dimentional gauss low frequency filter, and the element of significantly scheming among the Map is designated as Map _i(1≤i≤n);

12) image is carried out piecemeal and handle, at first, the fixed size that segments the image into non-overlapping copies is the piecemeal of m=B * B, and establishing i piecemeal is B _i, wherein, 1≤i≤n/m supposes that n can be divided exactly by m, and each piecemeal i is calculated:

${\mathrm{\ξ}}_i=\frac{1}{m}{\mathrm{\Σ}}_{j\∈B_i}{\mathrm{Map}}_j$

13) by ξ _iCalculate arbitrary piecemeal B _i(the measurement number of times M in compressed sensing of 1≤i≤n/m) _iFor:

$M_i=\mathrm{rnd}({\mathrm{\θ}}_{\max }-\frac{{\mathrm{\ξ}}_i}{255}\×({\mathrm{\θ}}_{\max }-{\mathrm{\θ}}_{\min }))$

Wherein, θ _MaxAnd θ _MinThe maximal value and the minimum value of element in the matrix measured in expression respectively, and rnd () gets and the immediate integer of input variable;

Then, with the measured value M of all piecemeals _i(1≤i≤n/m) be recorded in the measurement number of times vector M, i.e. M=(M ₁, M ₂..., M _N/m);

14) image is carried out gray average, the steps include:

At first, the fixed size that former figure is divided into non-overlapping copies is that (dB≤B) supposes that m can be divided exactly each piecemeal B by t for the piecemeal of t=dB * dB _i(1≤i≤n/m) be divided into m/t fritter again is designated as dB _Ij(1≤i≤n/m, 1≤j≤m/t);

Then, calculate each piecemeal dB _Ij(the gray average g of all pixels among the 1≤i≤n/m, 1≤j≤m/t) _Ij(1≤i≤n/m, 1≤j≤m/t), and all gray values of pixel points in this piecemeal are set at this gray average;

At last, with piecemeal B _i(each sub-piece dB among 1≤i≤n/m) _Ij(gray average of 1≤j≤m/t) is stored in vectorial g _i=(g _{I, 1}g _{I, 2}... g _{I, j}... g _{I, m/t}) in, 1≤i≤n/m;

15) choose random seed S, image X is carried out the piecemeal compressed sensing calculate, this piecemeal compressed sensing is calculated and is meant each sub-piece B _i(1≤i≤n/m) carry out respectively calculating based on the compressed sensing of human visual system; Afterwards, again the compressed sensing result of calculation of each sub-piece is pressed its sequential combination in former figure X together, thereby obtain the compressed sensing measured value of image X;

Described piecemeal B _i(compressed sensing based on human visual system of 1≤i≤n/m) is calculated, and the steps include:

At first, to B _iCarry out wavelet transformation, obtain the sparse signal on the wavelet field

$x_{B_i}={\mathrm{\Ψ}}^T{B}_i$

Wherein,

is the wavelet transformation base;

Secondly, be M according to S generation scale _iThe gaussian random of * m is measured matrix Φ _i, wherein m is piecemeal B _iPixel quantity, M _iBe the piecemeal B that obtains in the step 13) _iThe measurement number, the element in the said gaussian random matrix all produces at random, and its probability meets gaussian distribution nature;

At last, with sparse signal Project in this measurement matrix, obtain

Observed reading Y _i, that is:

$Y_i={\mathrm{\Φ}}_i{x}_{B_i}={\mathrm{\Φ}}_i{\mathrm{\Ψ}}^T{B}_i$

After all piecemeals have all carried out the compressed sensing measurement, with the compressed sensing result of calculation Y of each sub-piece _iBy its sequential combination in former figure X together, the compressed sensing that has just obtained image X is measured Y;

16) use the LDPC lossy coding that the compressed sensing of image X is measured Y and encode, obtain the cryptographic hash Hash (X) of image X; Its process is specially, respectively to each piecemeal B _iObserved reading Y _iCarry out the LDPC lossy coding, obtain compressed code S _iAfterwards, again with each S _iConnect according to the order of sequence, the symbol string that obtains is as cryptographic hash Hash (X)=(S of image X ₁S ₂... S _N/m), 1≤i≤n/m;

17) watermark of image X generates and credible transmission, and the watermark watermark (X) of said image X is exactly by Hash (X), random seed S and measures the tlv triple that the number of times vector M is formed, promptly

watermark(X)＝(Hash(X)，S，M)

Transmitting terminal is calculating watermark (X) afterwards, and the trusted channel through itself and receiving end sends watermark (X) to receiving end, through overt channel issue image X.

3. a kind of gray level image altering detecting method according to claim 2 based on the human visual system compressed sensing; It is characterized in that; Described LDPC lossy coding is according to the distributed coding theory of lossy compression method signal to be carried out compressed encoding, is specially, for compressed signal W=(w ₁w ₂... w _k), get its each element w _i(the plurality of continuous position bit of 1≤i≤k), the initial bits position of establishing the bit of getting is p, stopping bit is q, and with the interposition Bit String according to each element w _i(1≤i≤k) order in signal W connects successively, and the new Bit String that obtains is exactly the LDPC coding of signal W.

4. a kind of gray level image altering detecting method according to claim 2 based on the human visual system compressed sensing; It is characterized in that; To twice minute block message: B, the dB of original image, and the initial bits position p of LDPC coding with stop bit q and can arrange in advance or send receiving end to through trusted channel.

5. a kind of gray level image altering detecting method according to claim 4 based on the human visual system compressed sensing; It is characterized in that; Described step 2) receiving end image detection; The image watermark that the receiving end utilization is received by trusted channel, whether the image that checking is obtained by overt channel has passed through is distorted; Be specially:

21) adopt the mode identical to receiving image X ' ∈ R with transmitting terminal ^{N * N}Carry out gray average, the steps include:

At first, the fixed size that X ' is divided into non-overlapping copies is the piecemeal B of m=B * B _i', 1≤i≤n/m; And then with each B _i' the size that is divided into non-overlapping copies is the piecemeal dB of t=dB * dB _Ij', 1≤i≤n/m, 1≤j≤m/t;

Secondly, calculate each piecemeal dB _Ij' in the gray average g of all pixels _Ij', and all gray values of pixel points in this piecemeal are set at this gray average;

At last, with piecemeal B _iIn each sub-piece dB _IjGray average be stored in vectorial g _i'=(g _{I, 1}' g _{I, 2}' ... g _{I, j}' ... g _{I, m/t}') in, 1≤i≤n/m, 1≤j≤m/t;

22) adopt the mode identical to receiving image X ' ∈ R with transmitting terminal ^{N * N}Carrying out the piecemeal compressed sensing calculates;

For piecemeal B _i', 1≤i≤n/m at first carries out wavelet transformation, obtains the sparse signal on the wavelet field

${x_{B_i}}^{\′}={\mathrm{\Ψ}}^T{B}_i$

Wherein,

is the wavelet transformation base;

Secondly, according to random seed S that is received by trusted channel and mapping graph M, the generation scale is M _iThe gaussian random of * m is measured matrix Φ _i

At last, with sparse signal

Project in this measurement matrix, obtain

Observed reading Y _i', that is:

${Y_i}^{\′}={\mathrm{\Φ}}_i{x_{B_i}}^{\′}={\mathrm{\Φ}}_i{\mathrm{\Ψ}}^T{B_i}^{\′}$

After all piecemeals have all carried out the compressed sensing measurement, with the compressed sensing result of calculation Y of each sub-piece _i' by its sequential combination in former figure X ' together, the compressed sensing that has just obtained image X ' is measured Y '=(Y ₁' Y ₂' ... Y _N/m');

23) image cryptographic hash Hash (X)=(S to receiving ₁S ₂... S _N/m), 1≤i≤n/m carries out the distributed decoding of the LDPC "=(Y that obtains Y ₁" Y ₂" ... Y _N/m"), that is, initial bits position p when encoding according to LDPC and termination bit q are with S _iAs Y _i" in the bit value of p to the q position of each element; Y _i" in all the other everybody by Y _i' correspondence everybody fill up successively;

24) distorted image detects, and calculates Y " and the difference of Y ':

D＝Y″-Y′

If D=0 explains that then the image of receiving end reception is consistent with the image that transmitting terminal is sent, do not distorted; If D ≠ 0, then presentation video was distorted, and carried out further tampering location, to point out in which position of image distorting.

6. a kind of gray level image altering detecting method based on the human visual system compressed sensing according to claim 5 is characterized in that, described step 3) receiving end tampering location is specially:

31) initializaing variable is set:

A) maximum iteration time Iter and acceptable error threshold e are set;

B) initialization Increment Matrix Aug_t is a null matrix, scale and the step 24 of matrix A ug_t) in the matrix of differences D that calculates identical;

C) residual values error initial value being set is error threshold e; Iterations time=0 is set;

D) temporary variable product being set is the matrix that has identical size with D;

32) with matrix product zero clearing, be changed to null matrix; Time=time+1;

33) recomputate the procuct matrix:

product ⁱ＝abs(D ⁱ?gerror)

Product wherein ⁱThe i row of expression product matrix, D ⁱBe the i row of D, abs is an ABS function; Be each column vector of calculated difference matrix D and the inner product of residual values error, with its column vector as product matrix respective column;

34) confirm the maximal value val of each matrix element of product and the row pos at place thereof;

35) expand Increment Matrix Aug_t, with product ^PosRow are added in the former Aug_t matrix as last row, obtain the Aug_t matrix after epicycle expands, that is: [Aug_t, product ^Pos] → Aug_t;

36) with the pos row zero setting of matrix of differences D;

37) calculate recovery matrix A ug_y:

Aug_y＝(Aug_t′×Aug_t) ^-1×Aug_t′×e

Wherein, the transposed matrix of Aug_t ' expression Increment Matrix Aug_t; (Mat) ^-1The inverse matrix of representing matrix Mat;

38) each matrix element among the Aug_t connected line by line obtain the vectorial L.Aug_t of row, each matrix element among the Aug_y is connected line by line obtain the vectorial L.Aug_y of row equally, calculate the residual values error of epicycle iteration then:

error＝e-L.Aug_t×L.Aug_y′

Wherein, L.Aug_y ' expression is by the column vector that obtains of the vectorial L.Aug_y transposition of row;

39) judge that whether error＜e satisfies, if satisfy, then goes to step 310);

Otherwise, at first judge whether time=Iter, be then to go to step 310), otherwise return step 32) the continuation iterative computation;

310) algorithm finishes, and the recovery matrix A ug_y that obtains i.e. tampering location figure E for being recovered by D.

Images