CN108376413A - A kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature - Google Patents
A kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature Download PDFInfo
- Publication number
- CN108376413A CN108376413A CN201810064386.0A CN201810064386A CN108376413A CN 108376413 A CN108376413 A CN 108376413A CN 201810064386 A CN201810064386 A CN 201810064386A CN 108376413 A CN108376413 A CN 108376413A
- Authority
- CN
- China
- Prior art keywords
- matrix
- image
- indicate
- feature
- differnce
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 17
- 239000011159 matrix material Substances 0.000 claims abstract description 139
- 238000012549 training Methods 0.000 claims description 39
- 239000013598 vector Substances 0.000 claims description 33
- 238000012360 testing method Methods 0.000 claims description 27
- 230000007704 transition Effects 0.000 claims description 14
- 239000000284 extract Substances 0.000 claims description 11
- 238000000034 method Methods 0.000 claims description 11
- 238000013139 quantization Methods 0.000 claims description 7
- 230000009467 reduction Effects 0.000 claims description 5
- 230000008707 rearrangement Effects 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000012706 support-vector machine Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims 1
- 238000007906 compression Methods 0.000 abstract description 13
- 230000006835 compression Effects 0.000 abstract description 12
- 238000005516 engineering process Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/007—Transform coding, e.g. discrete cosine transform
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/005—Statistical coding, e.g. Huffman, run length coding
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
- H04N19/625—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding using discrete cosine transform [DCT]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Discrete Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Signal Processing (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
- Compression Of Band Width Or Redundancy In Fax (AREA)
Abstract
The present invention relates to digital image evidence collecting technical fields,More specifically,It is related to a kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature,For JPEG compression principle,After dct transform is utilized,The correlation of DCT coefficient in DCT blocks is the phenomenon that frequency domain direction enhances,Permutatation has been carried out to the DCT coefficient matrix of image,So that the DCT coefficient of adjacent frequency domain is horizontally adjacent in reorder matrix,Then the Markov features of the horizontal difference of reorder matrix are extracted,And the horizontal and vertical Markov features of the diagonally opposing corner difference matrix of the DCT coefficient matrix of combination image are as characteristic of division,The difference information of frequency domain in block is utilized in this feature well,It can obtain effective grader,The present invention can be effectively detected whether image passes through JPEG squeeze operations again,In conjunction with JPEG compression principle,Using difference statistical nature in the block in picture frequency domain as characteristic of division,With preferable detection result,Effectively increase the accuracy rate of detection.
Description
Technical field
The present invention relates to digital image evidence collecting technical fields, and frequency domain differential demodulation statistical nature is based on more particularly, to one kind
Jpeg image weight contracting detection method.
Background technology
With the development of internet, digital picture has had changed into a kind of very important information carrier, is propagated in information
It is middle to play great effect.But with the fast development of digital image editing software, distorting for digital picture becomes increasingly
It being easy, the image after distorting becomes closer to true picture, and in many occasions, such as administration of justice, medicine, news media, science are ground
Study carefully in equal fields, the authenticity of digital picture plays very important influence to result, this is undoubtedly to the authenticity of digital picture
Propose exact requirement so that digital image evidence collecting technology becomes increasingly eager.
Jpeg image weight compressed detected is an important branch of digital image evidence collecting technology, it is therefore an objective to detect an image
Whether twice JPEG compression have passed through.It, generally can be again in order to achieve the purpose that not to be therefore easily perceived by humans after jpeg image is distorted
It is secondary to save as jpeg format, in this way, tampered image will undergo JPEG compression twice, i.e. JPEG weight compression process.To image into
Row JPEG weight compressed detecteds are to carry out other bases for distorting operation detection, can be as the pre- of various jpeg image tampering detections
Judge.
Existing JPEG weight contracting detection technique is broadly divided into two kinds：One is based on DCT coefficient statistical distribution, this side
Method is mainly the difference being distributed according to the DCT coefficient of the distribution of the DCT coefficient of single jpeg compressed image and JPEG weight contract drawing pictures,
The DCT coefficient distribution of test image is judged, to show whether test image passes through JPEG squeeze operations again；It is another
Method is to be based on DCT coefficient difference statistics feature, and this method is calculated adjacent on spatial domain using JPEG compression quantization loss principle
The difference statistics feature of DCT coefficient, and image is learnt and classified using machine learning method.
Invention content
It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of JPEG based on frequency domain differential demodulation statistical nature
Image weight contracting detection method, can be effectively detected whether image passes through JPEG squeeze operations again, in conjunction with JPEG compression principle,
Using difference statistical nature in the block in picture frequency domain as characteristic of division, there is preferable detection result.
To solve the above problems, technical solution provided by the invention is：A kind of JPEG based on frequency domain differential demodulation statistical nature
Image weight contracting detection method, wherein include the following steps：
S1. training set of images is chosen：Single jpeg compressed image that training set is compressed by various quality factor QF1 and
The JPEG weight contracting image constructions compressed by quality factor QF1, QF2, wherein
QF1 ∈ { 50,55,60,65,70,75,80,85,90,95 },
QF2 ∈ { 50,55,60,65,70,75,80,85,90,95 }, and QF1 ≠ QF2.
S2. permutatation is carried out to the DCT coefficient matrix in the channels training image Y：After extracting the quantization of the channels training image Y
DCT coefficient matrix, to obtained DCT coefficient matrix according to from left to right, sequence from top to bottom successively presses each DCT blocks
Zigzag sequences carry out permutatation；Each DCT blocks are rearranged into the row vector that a size is 64 dimensions, all DCT blocks are carried out
After rearrangement, the reorder matrix that a size is m × 64 is finally obtained, wherein m is the total quantity of DCT blocks, in this reorder matrix
In, the DCT coefficient of adjacent frequency domain can come adjacent position；Similarly, image is carried out according to anti-zigzag sequences identical
Reordering operations obtain the reorder matrix that a size is m × 64, finally obtain the reorder matrix of two m × 64.
S3. Markov features are extracted：
S31. according to following formula to the second differnce matrix in two obtained reorder matrix calculated level directions of step S2：
D1h(u, v)=F (u, v)-F (u, v+1)
D2h(u, v)=D1h(u,v)-D1h(u,v+1)
Wherein, F indicates that reorder matrix, (u, v) indicate that the position of u rows v row in image, F (u, v) indicate that (u, v) is right
The DCT coefficient answered, D1hIndicate reorder matrix first-order difference matrix in the horizontal direction, D2hIndicate reorder matrix in the horizontal direction
Second differnce matrix；
S32. to the coefficient of each difference matrix, break-in operation is carried out using threshold value T, the coefficient more than T all replaces with
T, the coefficient less than-T all replace with-T；
S33. the three rank Markov for calculating the difference matrix horizontal direction obtained in S32 steps according to following formula respectively turn
Move probability matrix：
Wherein, M3hIndicate that three rank Markov transition probability matrixs of horizontal direction, r indicate the line number of second differnce matrix, c
Indicating second differnce matrix column number, u and v indicate the position in second differnce matrix, i, j, x, y ∈-T,-T+1 ..., -1,0,
1 ..., T-1, T }, indicate the various possible values of second differnce matrix, and：
S34. the DCT coefficient matrix after quantifying successively to original image according to following formula calculates two scales in diagonally opposing corner direction
Sub-matrix：
D1d(u, v)=D (u, v)-D (u+1, v+1)
D2d(u, v)=D1d(u,v)-D1d(u+1,v+1)
Wherein, D indicates that the DCT coefficient matrix after artwork quantization, (u, v) indicate the position of u rows v row in image, D
(u, v) indicates (u, v) corresponding DCT coefficient, D1dIndicate DCT coefficient matrix in the first-order difference matrix in diagonally opposing corner direction, D2dTable
Show DCT coefficient matrix diagonally opposing corner direction second differnce matrix；
S35. the second differnce matrix obtained to S34 carries out break-in operation using threshold value T, and the coefficient more than T is all replaced
For T, the coefficient less than-T all replaces with-T, that is, carries out S32 operations；
S36. the level of the difference matrix obtained in S35 steps, three ranks of vertical direction are calculated according to following formula respectively
Markov transition probability matrixs：
Wherein, M3h、M3vIndicate that three rank Markov transition probability matrixs horizontally and vertically, r indicate two respectively
The line number of scale sub-matrix, c indicate that second differnce matrix column number, u and v indicate the position in second differnce matrix, i, j, x, y
∈ {-T,-T+1 ..., -1,0,1 ..., T-1, T } indicates the various possible values of second differnce matrix, and：
S37. the three rank Markov transition probability matrixs that S33 steps obtain and the three rank Markov that S36 steps obtain are turned
It moves probability matrix to connect together, obtains feature vector in the block in the picture frequency domain；
S38. S31 is carried out to the operation of S37 steps to all images of training set, obtains an eigenmatrix, wherein is every
The feature vector of a line represents three rank Markov feature vectors of each image, and line number is the total number of images in training set.
S4. training characteristics prepare：After carrying out S3 operations to the image in all training sets, it can obtain owning in training set
The feature vector of image, then the feature vector of single jpeg compressed image is identified as 1, by the feature vector mark of weight contract drawing picture
It is -1 to know, and using the feature set identified as the feature training set of SVM, inputs in svm classifier model and learnt.
S5：SVM-RFE dimensionality reductions：Feature is trained using the recurrence feature removing method SVM-RFE based on support vector machines
Collection is ranked up, and effective feature is made to come before feature training set, after the feature list after being reset, selects feature list
Preceding n characteristic value constitute new feature vector, n feature of all images forms a new set of eigenvectors.
S6. optimal c, g parameters are found：Radial base kernel is used to the obtained set of eigenvectors of S5 and corresponding identification sets
SVM be trained, search for optimal penalty parameter c and nuclear parameter g using the method for grid search, obtain sorter model.
S7. test image extracts feature：The permutatation in S2 first is carried out to test image, then extracts the water of reorder matrix
The Markov feature vectors of the horizontal direction of adjustment sub-matrix, and extract the diagonally opposing corner difference matrix of DCT coefficient matrix level,
Vertical Markov features carry out the operation of S3, are then operated according to S5, take preceding n of the Ordered list of features after resetting
Characteristic value forms the feature vector of test image.
S8. classification prediction：The feature vector of the obtained test images of S7 is input in the svm classifier model that S6 is obtained,
Obtain the prediction result of test image；Wherein, 1 represent test image as single jpeg compressed image, -1 represent test image as
JPEG weight contract drawing pictures.
Compared with prior art, advantageous effect is：The present invention is directed to JPEG compression principle, after dct transform is utilized, DCT
The correlation of DCT coefficient in block has carried out permutatation the phenomenon that frequency domain direction enhances, to the DCT coefficient matrix of image, makes
The DCT coefficient for obtaining adjacent frequency domain is adjacent in reorder matrix, then extracts the Markov features of the horizontal difference of reorder matrix, separately
Outside, the frequency domain in DCT blocks increases along the upper left corner to bottom right angular direction, and the difference statistics feature in diagonally opposing corner direction can fine land productivity
With the frequency domain differential demodulation statistical nature in DCT blocks, in conjunction with the horizontal difference of reorder matrix Markov features and DCT coefficient matrix it is oblique
The Markov features of diagonal difference matrix, obtain frequency domain differential demodulation statistical nature in block, and frequency in block is utilized in this feature well
The difference information in domain can obtain effective grader, effectively increase the accuracy rate of detection.
Description of the drawings
Fig. 1 is flow chart of the method for the present invention.
Fig. 2 is average standard of the jpeg image weight compressed detected at identical quality factor QF2 and QF1 in the embodiment of the present invention
True rate definition graph.
Specific implementation mode
As shown in Figure 1, a kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature, wherein including
Following steps：
Step 1. chooses training set of images：Single jpeg compressed image that training set is compressed by various quality factor QF1 and
The JPEG weight contracting image constructions compressed by quality factor QF1, QF2, wherein
QF1 ∈ { 50,55,60,65,70,75,80,85,90,95 },
QF2 ∈ { 50,55,60,65,70,75,80,85,90,95 }, and QF1 ≠ QF2；
S11. the UCID numbers that the data set used in this example is provided by The Nottingham Trent University
Image in the NRCS databases provided according to library and USDANatural Resources Conservation Service is made,
Contain the 3000 big small uncompressed image for being 384 × 512；
S12. JPEG compression is carried out to this 3000 uncompressed images according to quality factor QF1, it is right under each QF1 to obtain
The 3000 single jpeg compressed images answered, then to the single jpeg compressed image of each QF1, with QF2 (QF2 ≠ QF1) into
Second of JPEG compression of row obtains final JPEG weight contract drawing pictures；
S13. in this example, the corresponding single jpeg compressed images of each QF1 have 3000, per a pair of QF1, QF2
(QF1 ≠ QF2) has 3000 JPEG weight contract drawing pictures, according to training image and test image 5:1 ratio, at random from each
2500 are chosen in kind situation and is used as training image, and remaining 500 are used as testing image；
It is right when the single jpeg compressed image compressed with quality factor QF1=Q is positive sample in this example
The negative sample answered is the single jpeg image not compressed for the quality factor of Q with some first, then is compressed with QF2=Q
The JPEG weight contract drawing pictures arrived correspond to for example, when positive sample is the single jpeg compressed image of quality factor QF1=60
JPEG weight contract drawing have QF1 ∈ 50,55,65,70,75,80,85,90,95 as negative sample, nine kinds of situations of QF2=60, respectively
This nine kinds of situations are tested, then to different (Q ∈ { 50,55,60,65,70,75,80,85,90,95 }) Q the case where
Also it is tested respectively.
Step 2. carries out permutatation to the DCT coefficient matrix in the channels training image Y：After extracting the quantization of the channels training image Y
DCT coefficient matrix, to obtained DCT coefficient matrix according to from left to right, sequence from top to bottom, successively to each DCT blocks
Permutatation is carried out by zigzag sequences；By each DCT blocks be rearranged into a size be 64 dimension row vectors, to all DCT blocks into
After rearrangement, the reorder matrix that a size is m × 64 is finally obtained, wherein m is the total quantity of DCT blocks, in this rearrangement square
In battle array, the DCT coefficient of adjacent frequency domain can come adjacent position；Similarly, image is carried out according to anti-zigzag sequences identical
Reordering operations, obtain a size be m × 64 reorder matrix, finally obtain the reorder matrix of two m × 64.
Step 3. extracts Markov features：
S31. according to following formula to the second differnce matrix in two obtained reorder matrix calculated level directions of step S2：
D1h(u, v)=F (u, v)-F (u, v+1)
D2h(u, v)=D1h(u,v)-D1h(u,v+1)
Wherein, F indicates that reorder matrix, (u, v) indicate that the position of u rows v row in image, F (u, v) indicate that (u, v) is right
The DCT coefficient answered, D1hIndicate reorder matrix first-order difference matrix in the horizontal direction, D2hIndicate reorder matrix in the horizontal direction
Second differnce matrix；
S32. to the coefficient of each difference matrix, break-in operation is carried out using threshold value T, the coefficient more than T all replaces with
T, the coefficient less than-T all replace with-T, in this example, take T=3；
S33. the three rank Markov for calculating the difference matrix horizontal direction obtained in S32 steps according to following formula respectively turn
Move probability matrix：
Wherein, M3hIndicate that three rank Markov transition probability matrixs of horizontal direction, r indicate the line number of second differnce matrix, c
Indicating second differnce matrix column number, u and v indicate the position in second differnce matrix, i, j, x, y ∈-T,-T+1 ..., -1,0,
1 ..., T-1, T }, indicate the various possible values of second differnce matrix, and：
S34. the DCT coefficient matrix after quantifying successively to original image according to following formula calculates two scales in diagonally opposing corner direction
Sub-matrix：
D1d(u, v)=D (u, v)-D (u+1, v+1)
D2d(u, v)=D1d(u,v)-D1d(u+1,v+1)
Wherein, D indicates that the DCT coefficient matrix after artwork quantization, (u, v) indicate the position of u rows v row in image, D
(u, v) indicates (u, v) corresponding DCT coefficient, D1dIndicate DCT coefficient matrix in the first-order difference matrix in diagonally opposing corner direction, D2dTable
Show DCT coefficient matrix diagonally opposing corner direction second differnce matrix；
S35. the second differnce matrix obtained to S34 carries out break-in operation using threshold value T, and the coefficient more than T is all replaced
For T, the coefficient less than-T all replaces with-T, that is, carries out S32 operations；
S36. the level of the difference matrix obtained in S35 steps, three ranks of vertical direction are calculated according to following formula respectively
Markov transition probability matrixs：
Wherein, M3h、M3vIndicate that three rank Markov transition probability matrixs horizontally and vertically, r indicate two respectively
The line number of scale sub-matrix, c indicate that second differnce matrix column number, u and v indicate the position in second differnce matrix, i, j, x, y
∈ {-T,-T+1 ..., -1,0,1 ..., T-1, T } indicates the various possible values of second differnce matrix, and：
S37. the three rank Markov transition probability matrixs that S33 steps obtain and the three rank Markov that S36 steps obtain are turned
It moves probability matrix to connect together, obtains feature vector in the block in the picture frequency domain；
S38. S31 is carried out to the operation of S37 steps to all images of training set, obtains an eigenmatrix, wherein is every
The feature vector of a line represents three rank Markov feature vectors of each image, and line number is the total number of images in training set.
Step 4. training characteristics prepare：After carrying out step 3 operation to the image in all training sets, training set can be obtained
In all images feature vector, then the feature vector of single jpeg compressed image is identified as 1, by the feature of weight contract drawing picture
Vectorial is -1, and using the feature set identified as the feature training set of SVM, inputs in svm classifier model and learnt.
In this example, it is 2500 × 9604 that the training set for the single jpeg image that each corresponding QF1=Q compresses, which is a size,
Matrix, each QF1 ≠ Q, QF2=Q correspond to lower JPEG weight contracting training set of images be a size for 2500 × 9604 square
Battle array.
Step 5：SVM-RFE dimensionality reductions：Feature is instructed using the recurrence feature removing method SVM-RFE based on support vector machines
Practice collection to be ranked up, effective feature is made to come before feature training set, after the feature list after being reset, selects characteristic series
The preceding n characteristic value of table constitutes new feature vector, and n feature of all images forms a new set of eigenvectors；This reality
In example, n=300, so, the eigenvectors matrix size of the single jpeg compressed image obtained after dimensionality reduction is 2500 × 300, and
The JPEG weight contracting image characteristic matrix sizes obtained after dimensionality reduction are 2500 × 300.
Step 6. finds optimal c, g parameters：The set of eigenvectors obtained to step 5 and corresponding identification sets use radial
The SVM of base kernel is trained, and is searched for optimal penalty parameter c and nuclear parameter g using the method for grid search, is obtained grader
Model；In this example, optimized parameter c=5.656854, g=0.02 are finally obtained.
Step 7. test image extracts feature：The permutatation in step 2 first is carried out to test image, obtains two m × 64
Reorder matrix, then calculate the horizontal difference matrix of reorder matrix, extract the Markov features of horizontal direction, and combine DCT
The Markov features of the diagonally opposing corner difference matrix of coefficient matrix, obtain feature vector, that is, carry out the operation of step 3, then according to
Step 5 operates, and takes preceding 300 characteristic values of the Ordered list of features after resetting, forms the feature vector of test image.This example
In, to the single jpeg image of each QF1 compressions, the eigenvectors matrix that a size is 500 × 300 can be obtained, it is right
Per a pair of QF1, QF2, the eigenvectors matrix that a size is 500 × 300 can be obtained.
Step 8. classification prediction：The feature vector for the test image that step 7 obtains is input to the SVM that step 6 obtains to divide
In class model, the prediction result of test image is obtained；Wherein, 1 test image is represented as single jpeg compressed image, -1 representative is surveyed
Attempt as being JPEG weight contract drawing pictures.In this example, 10 random selections are carried out to training image and test image, and to 10 times
Test result average, obtained prediction result is as shown in Fig. 2, Fig. 2 (a) indicates that JPEG weight contract drawings picture is kept (to survey
Try negative sample image) in the case of QF2 is identical, the Average Accuracy detected in the case of nine kinds of different Q F1 compressions, Fig. 2 (b) is indicated
In the case of keeping JPEG weight contract drawings picture (test negative sample image) QF1 being identical, detected in the case of nine kinds of different Q F2 compressions
Average Accuracy, as can be seen that the accuracy rate of image prediction all exists in the accuracy rate of 93% or more, Fig. 2 (b) from Fig. 2 (a)
QF1<When 95, accuracy rate is up to 97%, and in QF1=95, accuracy rate is 87% or so.
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
All any modification, equivalent and improvement etc., should be included in the claims in the present invention made by within the spirit and principle of invention
Protection domain within.
Claims (3)
1. a kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature, which is characterized in that including following step
Suddenly：
S1. training set of images is chosen：Training set by the various quality factor QF1 single jpeg compressed images compressed and by quality because
The JPEG weight contracting image constructions that sub- QF1, QF2 compress, wherein
QF1 ∈ { 50,55,60,65,70,75,80,85,90,95 },
QF2 ∈ { 50,55,60,65,70,75,80,85,90,95 }, and QF1 ≠ QF2；
S2. permutatation is carried out to the DCT coefficient matrix in the channels training image Y；
S3. Markov features are extracted：The first second differnce matrix to the reorder matrix calculated level obtained in S2, obtains original image
Then second differnce matrix in the block of frequency domain direction carries out truncation using threshold value T to the difference matrix, is more than the numerical value of T
It is all replaced with T, the numerical value whole less than-T is replaced with-T, and finally to treated, difference matrix extracts three horizontal ranks
Markov transition probability matrixs, and the level of the second differnce matrix in the diagonally opposing corner direction of the DCT coefficient matrix of combination original image,
Three vertical rank Markov transition probability matrixs, obtain the Markov feature vectors in the image block；
S4. training characteristics prepare：After carrying out S3 operations to the image in all training sets, all images in training set can be obtained
Feature vector, then the feature vector of single jpeg compressed image is identified as 1, the feature vector of weight contract drawing picture is identified
It is -1, and using the feature set identified as the feature training set of SVM, inputs in svm classifier model and learnt；
S5：SVM-RFE dimensionality reductions：Using the recurrence feature removing method SVM-RFE based on support vector machines to feature training set into
Row sequence, makes effective feature come before feature training set, after the feature list after being reset, before selecting feature list
N characteristic value constitutes new feature vector, and n feature of all images forms a new set of eigenvectors；
S6. optimal c, g parameters are found：Radial base kernel is used to the obtained set of eigenvectors of S5 and corresponding identification sets
SVM is trained, and is searched for optimal penalty parameter c and nuclear parameter g using the method for grid search, is obtained sorter model；
S7. test image extracts feature：The permutatation in S2 first is carried out to test image, then extracts the level error of reorder matrix
The Markov feature vectors of the horizontal direction of sub-matrix, and extract DCT coefficient matrix diagonally opposing corner difference matrix it is horizontal, vertical
Markov features, that is, carry out S3 operation, then operated according to S5, take reset after Ordered list of features preceding n feature
Value, forms the feature vector of test image；
S8. classification prediction：The feature vector of the obtained test images of S7 is input in the svm classifier model that S6 is obtained, is obtained
The prediction result of test image；Wherein, it 1 represents test image and represents test image as single jpeg compressed image, -1 as JPEG
Weight contract drawing picture.
2. a kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature according to claim 1,
It is characterized in that, the S2 steps include：
S21. the DCT coefficient matrix after the quantization of extraction training image Y channels, to obtained DCT coefficient matrix according to from left to right,
Sequence from top to bottom carries out permutatation to each DCT blocks by zigzag sequences successively；
S22. each DCT blocks are rearranged into the row vectors that a size is 64 dimensions, to all DCT blocks into after rearrangement, final
To the reorder matrix that a size is m × 64, wherein m is the total quantity of DCT blocks, in this reorder matrix, adjacent frequency domain
DCT coefficient can come adjacent position；
S23. identical as S21, S22 method, but identical reordering operations are carried out to image according to anti-zigzag sequences, obtain one
Size is the reorder matrix of m × 64, finally obtains the reorder matrix of two m × 64.
3. a kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature according to claim 1,
It is characterized in that, the S3 steps specifically include：
S31. according to following formula to the second differnce matrix in two obtained reorder matrix calculated level directions of step S2：
D1h(u, v)=F (u, v)-F (u, v+1)
D2h(u, v)=D1h(u,v)-D1h(u,v+1)
Wherein, F indicates that reorder matrix, (u, v) indicate that the position of u rows v row in image, F (u, v) indicate that (u, v) is corresponding
DCT coefficient, D1hIndicate reorder matrix first-order difference matrix in the horizontal direction, D2hIndicate reorder matrix in the horizontal direction two
Scale sub-matrix；
S32. to the coefficient of each difference matrix, break-in operation is carried out using threshold value T, the coefficient more than T all replaces with T, small
- T is all replaced in the coefficient of-T；
S33. the three rank Markov transfers for calculating the difference matrix horizontal direction obtained in S32 steps according to following formula respectively are general
Rate matrix：
Wherein, M3hIndicate that three rank Markov transition probability matrixs of horizontal direction, r indicate that the line number of second differnce matrix, c indicate
Second differnce matrix column number, u and v indicate the position in second differnce matrix, i, j, x, y ∈-T,-T+1 ..., -1,0,
1 ..., T-1, T }, indicate the various possible values of second differnce matrix, and：
S34. the DCT coefficient matrix after quantifying successively to original image according to following formula calculates the second differnce square in diagonally opposing corner direction
Battle array：
D1d(u, v)=D (u, v)-D (u+1, v+1)
D2d(u, v)=D1d(u,v)-D1d(u+1,v+1)
Wherein, D indicates that the DCT coefficient matrix after artwork quantization, (u, v) indicate the position of u rows v row in image, D (u, v)
Indicate (u, v) corresponding DCT coefficient, D1dIndicate DCT coefficient matrix in the first-order difference matrix in diagonally opposing corner direction, D2dIndicate DCT
Second differnce matrix of the coefficient matrix in diagonally opposing corner direction；
S35. the second differnce matrix obtained to S34 carries out break-in operation using threshold value T, and the coefficient more than T all replaces with T,
Coefficient less than-T all replaces with-T, that is, carries out S32 operations；
S36. the level of the difference matrix obtained in S35 steps, three ranks of vertical direction are calculated according to following formula respectively
Markov transition probability matrixs：
Wherein, M3h、M3vIndicate that three rank Markov transition probability matrixs horizontally and vertically, r indicate two scales respectively
The line number of sub-matrix, c indicate second differnce matrix column number, and u and v indicate the position in second differnce matrix, i, j, x, y ∈-
T,-T+1 ..., -1,0,1 ..., T-1, T }, indicate the various possible values of second differnce matrix, and：
S37. the three rank Markov transition probability matrixs that S33 steps obtain and the three rank Markov transfers that S36 steps obtain are general
Rate matrix connects together, and obtains feature vector in the block in the picture frequency domain；
S38. S31 is carried out to the operation of S37 steps to all images of training set, obtains an eigenmatrix, wherein per a line
Feature vector represent three rank Markov feature vectors of each image, line number is the total number of images in training set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810064386.0A CN108376413B (en) | 2018-01-23 | 2018-01-23 | JPEG image recompression detection method based on frequency domain difference statistical characteristics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810064386.0A CN108376413B (en) | 2018-01-23 | 2018-01-23 | JPEG image recompression detection method based on frequency domain difference statistical characteristics |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108376413A true CN108376413A (en) | 2018-08-07 |
| CN108376413B CN108376413B (en) | 2021-08-06 |
Family
ID=63016568
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810064386.0A Active CN108376413B (en) | 2018-01-23 | 2018-01-23 | JPEG image recompression detection method based on frequency domain difference statistical characteristics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108376413B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112001429A (en) * | 2020-08-06 | 2020-11-27 | 中山大学 | Depth forgery video detection method based on texture features |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0855681A2 (en) * | 1997-01-28 | 1998-07-29 | Nippon Telegraph And Telephone Corporation | Method of embedding watermark-information into digital data |
| CN102568016A (en) * | 2012-01-03 | 2012-07-11 | 西安电子科技大学 | Compressive sensing image target reconstruction method based on visual attention |
| CN102819842A (en) * | 2012-07-30 | 2012-12-12 | 桂林电子科技大学 | Displacement JPEG (joint photographic experts group) double-compression tampering blind detection method based on condition symbiotic probability matrix |
| CN104331913A (en) * | 2014-11-19 | 2015-02-04 | 西安电子科技大学 | Polarized SAR polarization method based on sparse K-SVD (Singular Value Decomposition) |
| CN104766273A (en) * | 2015-04-20 | 2015-07-08 | 重庆大学 | Infrared image super-resolution reestablishing method based on compressed sensing theory |
| CN106056600A (en) * | 2016-05-26 | 2016-10-26 | 中山大学 | Contourlet transform-based image splicing detection method |
| CN106960435A (en) * | 2017-03-15 | 2017-07-18 | 华中师范大学 | A kind of double compression automatic testing methods of jpeg image |
-
2018
- 2018-01-23 CN CN201810064386.0A patent/CN108376413B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0855681A2 (en) * | 1997-01-28 | 1998-07-29 | Nippon Telegraph And Telephone Corporation | Method of embedding watermark-information into digital data |
| CN102568016A (en) * | 2012-01-03 | 2012-07-11 | 西安电子科技大学 | Compressive sensing image target reconstruction method based on visual attention |
| CN102819842A (en) * | 2012-07-30 | 2012-12-12 | 桂林电子科技大学 | Displacement JPEG (joint photographic experts group) double-compression tampering blind detection method based on condition symbiotic probability matrix |
| CN104331913A (en) * | 2014-11-19 | 2015-02-04 | 西安电子科技大学 | Polarized SAR polarization method based on sparse K-SVD (Singular Value Decomposition) |
| CN104766273A (en) * | 2015-04-20 | 2015-07-08 | 重庆大学 | Infrared image super-resolution reestablishing method based on compressed sensing theory |
| CN106056600A (en) * | 2016-05-26 | 2016-10-26 | 中山大学 | Contourlet transform-based image splicing detection method |
| CN106960435A (en) * | 2017-03-15 | 2017-07-18 | 华中师范大学 | A kind of double compression automatic testing methods of jpeg image |
Non-Patent Citations (3)
| Title |
|---|
| FEI XUE 等: ""MSE period based estimation of first quantization step in double compressed JPEG images"", 《SIGNAL PROCESSING: IMAGE COMMUNICATION》 * |
| 巫兰英: "" JPEG图像重压缩特性研究及篡改区域定位"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 张玉金 等: ""基于条件共生概率矩阵的移位"", 《光电子·激光》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112001429A (en) * | 2020-08-06 | 2020-11-27 | 中山大学 | Depth forgery video detection method based on texture features |
| CN112001429B (en) * | 2020-08-06 | 2023-07-11 | 中山大学 | Depth fake video detection method based on texture features |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108376413B (en) | 2021-08-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104508682B (en) | Key frame is identified using the openness analysis of group | |
| CN106851437A (en) | A kind of method for extracting video frequency abstract | |
| Lv et al. | Difference of Gaussian statistical features based blind image quality assessment: A deep learning approach | |
| CN105718555A (en) | Hierarchical semantic description based image retrieving method | |
| CN104680173A (en) | Scene classification method for remote sensing images | |
| CN103853724A (en) | Multimedia data sorting method and device | |
| Wang et al. | COME for no-reference video quality assessment | |
| Reta et al. | Color uniformity descriptor: An efficient contextual color representation for image indexing and retrieval | |
| CN107509079B (en) | Text block coding method and device | |
| CN108376413A (en) | A kind of jpeg image weight contracting detection method based on frequency domain differential demodulation statistical nature | |
| CN106683074A (en) | Image tampering detection method based on haze characteristic | |
| CN111652260A (en) | Method and system for selecting number of face clustering samples | |
| CN109271997A (en) | A kind of image texture classification method based on jump subdivision local mode | |
| CN107944340A (en) | A kind of combination is directly measured and the pedestrian of indirect measurement recognition methods again | |
| CN115063692B (en) | Remote sensing image scene classification method based on active learning | |
| Gorisse et al. | IRIM at TRECVID 2010: semantic indexing and instance search | |
| CN108804988B (en) | Remote sensing image scene classification method and device | |
| CN107870923B (en) | Image retrieval method and device | |
| Wang et al. | Image quality assessment based on local orientation distributions | |
| CN108875572A (en) | The pedestrian's recognition methods again inhibited based on background | |
| Sebai et al. | Improving high resolution satellite images retrieval using color component features | |
| CN113269218B (en) | Video classification method based on improved VLAD algorithm | |
| Rapantzikos et al. | On the use of spatiotemporal visual attention for video classification | |
| Xuemei et al. | Novel shot boundary detection method based on support vector machine | |
| CN106375768B (en) | Video steganalysis method based on intra prediction mode calibration |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |