Research Article Digital Image Forgery Detection...

Research ArticleDigital Image Forgery Detection Using JPEG Featuresand Local Noise Discrepancies

Bo Liu Chi-Man Pun and Xiao-Chen Yuan

Department of Computer and Information Science University of Macau Macau China

Correspondence should be addressed to Chi-Man Pun cmpunumacmo

Received 9 January 2014 Accepted 16 February 2014 Published 16 March 2014

Academic Editors F Di Martino and I Lanese

Copyright copy 2014 Bo Liu et al This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Wide availability of image processing software makes counterfeiting become an easy and low-cost way to distort or conceal factsDriven by great needs for valid forensic technique many methods have been proposed to expose such forgeries In this paper weproposed an integrated algorithm which was able to detect two commonly used fraud practices copy-move and splicing forgeryin digital picture To achieve this target a special descriptor for each block was created combining the feature from JPEG blockartificial grid with that from noise estimation And forehand image quality assessment procedure reconciled these different featuresby setting proper weights Experimental results showed that compared to existing algorithms our proposed method is effective ondetecting both copy-move and splicing forgery regardless of JPEG compression ratio of the input image

1 Introduction

Nowadays altering digital images via intuitive software is anoperation of simpleness with very low cost thus every indi-vidual can synthesize a fake picture For the widely accessibleInternet the false information disseminates extremely fastAs a consequence the facts may be distorted and the publicopinion may be affected yielding negative social influence Itcan be even worse in the justice when pictures are presentedas evidenceTherefore there is a strong demand for valid androbust authentication method to discern whether a picture isoriginal or not

Two means are commonly utilized to make a forgerycopy-move and splicing In the former case a part of apicture is duplicated and then pasted onto other regions tocover any unwanted portion within the same picture [1] Inthe latter case tampered image consists of two sources andretains the majority of one image for detail [1] Researchersand scientists have proposed many methods [2] to exposesuch intendedmanipulations Passive forensicmethods fulfillthe task without additional information except for the imageitself thus showing advantages over active algorithms likewatermarking and other signature schemas Hence most

research work is absorbed in developing blind authenticationmethods

The forged picture leaves some clues which can be used tolocate the manipulated regions In the practice of copy-moveoperation because the pasted area though it may probablybe altered geometrically shares some similar features with theoriginal region which is duplicated searching for analogousfeatures abstracted from local area is a possible solution SIFTfeature can be used to locate clone areas [3 4] For splicingtampered image detection considering that there may besome discrepancies between the host image and the splicedregion attempts to find the difference to expose that forgeriesmake sense For instance Kakar et al [5] took advantage ofmotion blur discrepancy to detect fake pictures For so manypictures stored or disseminated in JPEG format some tracesleft by JPEG compression algorithm can be used Estimatingthe quantization matrix used in JPEG compression regionsthat possess inconsistent DCT coefficients are regarded assplicing area since an intact JPEG format picture should becoded by only one quantization table Hamdy et al developedthis idea in [6] However this approach fails to deal withdouble compression Indeed this method is effective onlyon detecting BMP format image which is composed of two

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 230425 12 pageshttpdxdoiorg1011552014230425

2 The Scientific World Journal

JPEG pictures with different compression table Althoughcomplex situation in double compression was discussed [7]multicompressions more than twice are still very hard toanalyze

Our goal is to automatically detect copy-move as wellas splicing forgery within a single process without any priorknowledge about the forgery type of the doubtful pictureThe reason is obvious Rather than putting the same pictureinto different algorithm which may be effective on only onecertain type of forgery one single method is time saving andavoids evaluating every detection result which will be veryhard to discern the true output from various results Lin andWu in [8] proposed an integrated method to detect bothcopy-move and splicing forgery But this method just con-nects two separate processes togetherThe errors in the step offorgery type judgment will greatly affect the detection resultActually it is unnecessary to classify the pictures into certainforgery type if there is a tool or feature sensitive to two attackpractices Li et al [9] proposed amethod based on JPEGblockartificial grids (BAG) detection to expose both splicing andcopy-move forgeries But there are two major problems inthat paper the first is that the algorithm must be adjustedbefore applying to different forgery the author revised thesplicing detection algorithm to deal with copy-move practiceHowever in practice we do not know the forgery type fora doubtful image The other defect is that the algorithmis sensitive to highly compressed image and ineffective onhigh quality picture with little compression To overcomethis shortcoming we inducted noise feature to compensatefor BAG algorithm When the picture is less compressedclear BAG is very difficult to extract We consider that localnoise level and category can be used as a feature to identifydifferent source of regions in a picture Inconsistency anddiscrepancy from regions to regions provide the other cluesexcept for BAG to locate forged areaTherefore we created anintegrated feature combining BAG and noise feature to verifythe authenticity of a doubtful picture

The paper is organized as follows Section 2 introducesthe block artificial grids and noise patterns for detectingforgeries and the following Section 3 details our proposedintegrated method In Section 4 experimental results will bepresented to show the effectiveness of ourmethod andwe alsocompared it with existing methods Finally we conclude thispaper in Section 5

2 Block Artificial Grid and Noise Estimation

21 Block Artificial Grid Extraction It is universally knownthat the lossy JPEG compression will introduce some visuallyvertical or horizontal breaks in the imageThese breaks calledblock artificial grid (BAG) appear at the border of each 8 times 8pixel block This property can be used to determine whethera picture is altered or not If the picture is intact blockartificial grids should only present on block borders whilethere is a great possibility that copied and pasted or splicedregions will bring their original BAGs which may appearwithin the 8 times 8 block rather than at borders Some papers[9 10] noticed this and Figure 1 illustrates the phenomenon

Theoretically speaking if we extract all the BAGs from agiven image areas with BAGs within the block border areregarded as forged regions Li et al [9] introduced steps toextract BAGs As it is mentioned before artificial grids arevisually vertical and horizontal lines and they are very weakwhen comparing to the border lines of objects in the pictureAnd themain purpose of extraction procedures is to enhancethese weak lines and to make them visible However linesare also strengthened which may be the edges of objects orjust objects themselves This will interfere with the detectionresult because we only need BAGs To allay the side effectwe preprocessed the doubtful image by excluding the edgesof objects But it should be noticed that BAG can also beregarded as vertical or horizontal edges For preserving BAGswe only excluded edges within certain range

Suppose that 119866 was the grayscale version of image 119868 andthen the edge119864was obtained by119864 = 119866lowast119878 where 119878 representsSobel operator and ldquolowastrdquo denotes convolutionThenwe definedwhether a pixel is excluded using

119877 (119898 119899) =

0 119863 (119864 (119898 119899)) isin [0 120579]

cup [120587

2minus 120579120587

2+ 120579] cup [120587 minus 120579 120587)

1 others

(1)

where 119863(sdot) denotes gradient of the pixel and 119877(119898 119899) = 1means excluded pixels Then we begin to extract BAGs

Firstly weak horizontal edges were extracted by calcu-lating second-order difference of an image For the testimage 119868(119898 119899) absolute second-order difference 119889(119898 119899) wasobtained by

119889 (119898 119899) = |2119868 (119898 119899) minus 119868 (119898 + 1 119899) minus 119868 (119898 minus 1 119899)| (2)

Then all differentials larger than 01 or 119877 = 1 arediscarded In subsequence enlarged horizontal lines areaccumulated from every 33 columns as shown in (3) Thena median filter Med[sdot] is used to refine the result in

119886 (119898 119899) =

16

sum

119894=119899minus16

119889 (119898 119894) (3)

119886119903(119898 119899) = 119886 (119898 119899) minusMed [119886 (119894 119899) | 119898 minus 16 le 119894 le 119898 + 16]

(4)

Weak horizontal edge 119887ℎis further periodical median

filtered as

119887ℎ(119898 119899) = Med [119886

119903(119894 119899) | 119894 = 119898 minus 16

119898 minus 8119898119898 + 8119898 + 16]

(5)

Similarly the vertical BAGs 119887V can also be attracted Asa result final BAG is obtained by adding two componentstogether in

119887 (119898 119899) = 119887ℎ(119898 119899) + 119887V (119898 119899) (6)

22 Noise Estimation Highly compressed by JPEG the pic-ture shows visual block artificial grids across the whole frame

The Scientific World Journal 3

(a) (b)

Figure 1 Illustration of BAG mismatch the region within the red circle in upper left picture is copied and spliced into upper right pictureBAGs appearing within 8 times 8 blockrsquos border are suspected to belong to regions from other pictures This mismatch may appear also incopy-move forgery practice

which can be extracted by algorithm described in Section 21However under some circumstances when the picture is nothighly compressed and stored in high quality the way byusing BAGonly becomes harder to detect forgery To increasethe versatility of the algorithm we use noise feature Thenoise comes from imaging sensor and internal circuits withina camera And the number of noise changes in accordancewith camera settings especially ISO sensitivity and exposuretime As an example Figure 2 shows that the visual noiseof images is captured from a Nikon D7000 camera We cansee that more noise appears in the image as the ISO speedrises In Figure 3 we can see that different camera modelfromdifferentmanufacture also shows unequal noise amountand forms although the pictures were taken in the samescenery with equal ISO speed So the noise can be used tohelp distinguish difference sources of a picture When twopictures are spliced together the noise level or patterns areinconsistent between regions By estimating the pattern orlevel of noise in different regions the forgery can be exposedvia noise discrepancies

In most cases the alien region has a specific shape suchas a tree a bird or a person The forged object may possessdifferent noise level compared to that of its surroundingsTo estimate every regionrsquos noise level the image should befirstly divided into small segments Most previous methodsdivide the picture into small overlapping blocks with equalsize But in our application this will lead to bad performancein next steps which need accurate noise estimation of eachregion to compare noise discrepancy This is because theforged area is not rectangle in most cases and the small blockwill contain original and alien pixels Therefore we segmentpicture into sets of pixels not regular shaped also known assuperpixels Employing this approach makes segments moremeaningful and easier to be processed in the following steps

because the segmentation algorithms locate the objects andboundaries other than the same-size blocks The result ofimage segmentation is a set of segments that collectively coverthe entire image or a set of contours extracted from the imageEach of the pixels in a region is similar with respect to somecharacteristic or computed property such as color intensityor texture Adjacent regions are significantly different withrespect to the same characteristic(s) [11]

In our application SLIC (simple linear iterative clus-tering) superpixels algorithm [12] was used to segmentpicture This algorithm is easy but better than other seg-mentation methods Given an 119872 times 119873 image 119868119888

119904where

119888 isin red green blue denotes different color channel Themeaning of subscript 119904 will be explained later and

119868 = (

119868119888

119904(1 1) sdot sdot sdot 119868

119888

119904(1119873)

d

119868119888

119904(119872 1) sdot sdot sdot 119868

119888

119904(119872119873)

) (7)

In essence SLIC is a clustering algorithm Similar toother clustering methods two steps are evolved with SLICsegmentation In the initialization step cluster centers 119862

119904=

(119897119904 119886119904 119887119904 119909119904 119910119904)119879 are assigned by sampling pixels at regular

grid Note that the picture is segmented in LAB color spaceThen cluster centers aremoved to the lowest gradient positionin a 3 times 3 neighborhood In the assignment step each pixelis associated with the nearest cluster center and an updatestep adjusts the cluster centers to be the mean (119897 119886 119887 119909 119910)119879vector of all the pixels belonging to the cluster A residualerror119864 between the new and previous cluster center locationsis computed Once 119864 le 119905ℎ119903119890119904ℎ119900119897119889 the algorithm stops Weassigned subscript 119904which denoted segment number to everypixel


(a) (b) (c) (d)

Figure 2 Visual noise comparison for pictures captured by the same cameraNikonD7000 under the same scenery with different ISO settings(a) ISO = 100 (b) ISO = 800 (c) ISO = 1600 and (d) ISO = 3200 Crops are 100 with ambient temperature approximately 22C

(a) (b) (c) (d)

Figure 3 Visual noise comparison for pictures taken by different cameras under the same scenery with ISO = 1600 (a) Canon 550D (b)Nikon D7000 (c) Sony A77 and (d) Pentax K5

Before construction of noise feature for every segment weexcluded sharp transitional area since noise estimation wasadversely affected by heterogeneous image content [13] Weestimated sharp area using its gray-scale image 119866 which wascalculated by

119866 = 02989119868119903+ 05870119868

119892+ 01140119868

119887 (8)

The sharpness edge of image was then obtained by 119864 =119866 lowast 119878 where 119878 represents Sobel operator and ldquolowastrdquo denotesconvolution We then define whether a pixel is in the sharparea using

119867(119898 119899) = 1 (119898 119899) isin 119864

0 (119898 119899) notin 119864(9)

where 119867(119898 119899) = 1 means that the pixel (119898 119899) is located insharp transitional area To guarantee that these areas will notaffect noise estimation in the next step we expand boundariesvia dilation by

119885 = 119867 oplus 119881 (10)

where119881 is a structure element of 3times3 ones119885 is the expandedsharp area

To extract noise feature of each segment produced byprevious SLIC algorithm we firstly employed denoisingalgorithm across the whole picture The estimated noise 119891 atlocation (119898 119899) of image 119868119888 was calculated by

119891119888119889(119898 119899) = 119868

119888(119898 119899) minus 119867

119888119889(119898 119899) (11)

where 119867119888119889 = 119868119888 lowast 119875119889 and filter 119875119889 119889 = 1 2 5 representsfive different filters used to trace different aspects of the noise[14] They are median filter Gaussian filter averaging filterand adaptive Wiener denoising with two neighborhood sizes3 times 3 and 5 times 5 respectively For instance high frequencynoise can be detected by using Gaussian filter and medianfilter addresses ldquosalt and pepperrdquo noise

For each combination of color channel 119888 and denoisingfilter 119875119889 we calculated the mean and standard deviation 120590119888119889

119904

values of each segment 119904 as the noise feature 119865119888119889119904= (120583119888119889

119904 120590119888119889

119904)

where

120583119888119889

119904=1

119877sum

119885(119898119899)=0

119891119888119889

119904(119898 119899)

120590119888119889

119904=1

119877( sum

119885(119898119899)=0

(119891119888119889

119904(119898 119899) minus 120583

119888119889

119904)2

)

12

(12)


Input image

Block artificial grids (BAG) extraction

BAG feature generation

BAG map

BAG feature

for each 8 times 8 block

Bxy

Figure 4 Flowchart of proposed method BAG feature generation

As a result we computed 3 times 5 times 2 = 30 dimensional featurevector 119865

119904of a segment

3 Integrated Method for Forgery Detection

We proposed an integrated method effective to both copy-move and splicing forgery Based on combination of blockartificial grid extraction with analysis of local noise dis-crepancies the algorithm showed valid performance to highcompression JPEG pictures as well as high quality imageslack of BAGs To implement the authentication process webuilt an indicator for every 8 times 8 nonoverlapped blocksof the doubtfulful picture The indicator mathematicallydescribed the possibility of the block being a forged areawith higher value denoting higher probability As describedin Figures 4 5 and 6 for each 8 times 8 block BAG feature andassigned label based on noise discrepancies were integratedby estimated compression indicator (see Algorithm 1 for adetailed description)

31 Block BAG Feature Construction In Section 21 we intro-duced how to extract BAGs For an intact picture the BAGsappear at the border of each 8 times 8 block while for a picturewith intentional copy-move or splicing operation some BAGswill be presented at some abnormal positions such as

Input image

SLIC segmentation

Segmented image

Noise estimation for each segment

Noise feature generation for each

Noise feature

Label map

Graph cut

Noise discrepancy map

Axy

8 times 8 block

60

50

40

30

20

10

Figure 5 Flowchart of proposed method noise feature generation

the center of the block For a fixed 8 times 8 block 119868119909119910 these

abnormal BAGs can be calculated [9] by

119861119909119910= Max

7

sum

119894=2

119887 (119894 119899) | 2 le 119899 le 7

minusMin7

sum

119894=2

119887 (119894 119899) | 119899 = 1 8

+Max7

sum

119894=2

119887 (119898 119894) | 2 le 119898 le 7

minusMin7

sum

119894=2

119887 (119898 119894) | 119898 = 1 8

(13)


BeginLoad image 119868(119898 119899)Generate BAG feature 119861

119909119910 119909 = 1 2 lfloor1198988rfloor 119910 = 1 2 lfloor1198998rfloor for 119868(119898 119899)

Divide 119868 into 119870 segments 119878119896 119896 isin 1 2 119870 by SLIC superpixels

For each 119878119896

Extract noise feature 119865119896

Assign label 119871119896(119898 119899) isin 0 1 for 119878

119896by graph cut

Generate noise feature 119860119909119910= (164)sum

8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)

Calculate image quality score 119876Calculate coefficient 120572 = 119891(119876)Set 119862119909119910= 0

Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910

Morphological operation-closing and opening (119862 ∙ 119872119888) ∘ 119872119900

End

Algorithm 1 Algorithm description

Table 1 Edge weights for graph cuts

Edge Weight For119905120572

119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573

32 Noise Discrepancy Label Assignment Thenoise feature ofeach segment had been calculated in Section 22 and then ourgoal was to segment the image into two regions with noisediscrepancies To achieve the target the energy-based graphcuts can be used

Energy minimization via graph cuts is proposed byBoykov et al [15] to solve labeling problems with low compu-tation cost In a common label assignment problem the labelsshould vary smoothly almost everywhere while preservingsharp discontinuities existing at object boundariesThese twoconstraints can be formulated as119864(119891) = 119864smooth(119891)+119864data(119891)where 119891 is a labeling that assigns each pixel 119901 isin 119875 a label119891119901isin L and 119864smooth measures the extent to which 119891 is

not piecewise smooth while 119864data measures the disagreementbetween 119891 and observed data The goal is to minimize thefunction Specifically the energy function can be rewritten as

119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)

where 119873 is neighboring pixels 119881 is the penalty of pairs inthe first term and 119863

119901is nonnegative and measures how well

label fits pixel Local minimum value can be obtained withthe help of graph cuts The simplified problem is illustratedin Figure 7 Since many algorithms have been proposed tosolve min-cut problem if proper weight value is assignedto each edge the problem of minimizing energy functionchanges to min-cut problem The weight is seen in Table 1The calculation result is a cut 119862 which separates two labelsFigure 7 shows two possible cuts and the label is assigned tothe pixel when cut 119862 contains the edge connecting that label

to the pixel For example in left case of Figure 7 label 120572 isassigned to pixel 119901 while 120573 is assigned to 119902 because cut 119862contains edge 119905120572

119901and 119905120573119902

Our forgery detection task can also be regarded as alabeling problem In our application there are two labelsthat need to be assigned to each segment produced bypreviously introduced SLIC algorithm forged area as theyshow inconsistency to rest segments in terms of noise levelor pattern and the original area And each segment isprocessed as a pixel The reason why we avoid employingwidely used outlier detection algorithms [16] and Otsursquosautomatic thresholding method [17] is the property of noiseFrom Figure 2 we observe that even the picture is takenby one camera and the amount of noise differs in differentillumination The color of object may also affect the noiselevel Accordingly the ideal algorithm should tolerate theselocal deviations and inconsistencies In other words it shouldkeep ldquosmoothrdquo across the image while preserving ldquosharprdquodiscontinuity in inconsistent boundariesThis requirement isidentical to label assignment problems described previouslywhile normal outlier detection algorithms are not capable ofthis

ldquoSmoothrdquo constraint is realized by proper assignment of119881(120572 120573) and ldquosharprdquo discontinuity requirement is supportedby 119863119901(lowast) We firstly discuss the weight of edge 119905120572

119901and 119905120573119901 We

computed average value of feature vector of all segments in 30dimensions and named it the mean vector 119865 Then we foundthe vector whose Euclidean distance was the largest from 119865by searching for all segments and called it 119865max For a featurevector 119865

119904the weight 119908 was obtained by

119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)

where 120572 was ldquooriginalrdquo label while 120573 was ldquoforgedrdquo and sdot denoted Euclidean distance between two vectors

From (15) we can find that if the noise level of a segmentis close to the average value across the whole picture theweight 119908

120572assigned is small while 119908

120573is large and vice versa

Thismeets the requirement of discontinuity preservingThenit is the turn to discuss smooth constraint Proper value of


Noise feature BAG feature

Input image

Quality estimation

Quality score Q

Calculating coefficient

Coefficient 120572

Closing and opening morphological operations

Detection result

Rough map

Axy 8 times 8

Generating integrated feature Cxy

Figure 6 Flowchart of proposed method combination of twofeatures

interaction penalty 119881(120572 120573) tolerates local deviations of noisewhich is affected by illumination or color There are manyforms proposed For an instance 119881(120572 120573) = min(119870 |120572 minus 120573|)or an important function given by the Potts model 119881(120572 120573) =119870sdot119879(120572 = 120573)where119879(sdot) is 1 if its argument is true and otherwise0 This penalty function possesses good feature of piecewisesmooth so we used it in the experiment

Graph cut based on noise discrepancy assigned everysegment 119878

119896a label L isin 0 1 indicating whether the area

was classified as forged (L119896= 1) or not (L

119896= 0) And

we assigned every pixel belonging to the segment the samelabelL(119898 119899) = L

119896 (119898 119899) isin 119878

119896 At last block indicator 119860

119909119910

described the possibility of forgery and was calculated by

119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)

33 Feature Generation for Forgery Detection In this stepwe combined together two features described already withproper coefficient Since the method based on BAG extrac-tion is only sensitive and feasible to highly compressedimages the form of combined feature is described as 119862

119909119910=

120572119861119909119910+ (1 minus 120572)119860

119909119910and 120572 denotes the coefficient assigned and

is a function of evaluated image compression rate or the JPEGimage quality namely 120572 = 119891(Q)

We firstly evaluated the quality of the picture and thenfound the function 119891 Proposed by Wang et al [18] thequality assessment algorithm is nonreferenced and sensitiveto JPEG compression rather than noise which was testedand verified by our experiment We took 20 pictures in rawfile (no compression) and then saved them as JPEG formatpictures with different compression ratio In our experiment100 means saving with the highest quality and the lowestcompression We assessed the picture quality of compressionrate as 100 80 60 40 20 and 5 respectively andaveraged the scores See Figure 8 for result less compressedpictures show higher quality scores

However the algorithm is less sensitive to noise effect Inthe experiment for each image set with certain compressionrate we added 10 20 and 40 monologue Gaussiannoise to the image respectively and then obtained theaverage quality scores See Figure 9 for result noise does notlargely affect quality scores Therefore we consider that thedominated factor affecting quality score in algorithm [18] isJPEG compression rate

Then we discuss how to generate the function 120572 = 119891(Q)In the experiment we made 60 fake pictures and every10 pictures were compressed in a certain rate And thenwe used BAG feature only to detect the forgeries Table 2shows detection accuracy in different compression rateThe experimental result confirmed that the BAG method isgood at dealing with the pictures with low quality scoreTherefore the value of 120572 should approach 1 when 119876 declinesnear to 2 for its detection accuracy is 100 Meanwhile 120572should be set to 0 when 119876 rises to 9 or so because of its lowaccuracy

The function 119891 we recommended based on experimentalresult is

119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)

In order to filter out some isolated false marked areas andimprove the integrity of suspect forged regionmorphologicaloperations including closing and opening are used The finalresult comes from (119862∙119872

119888)∘119872119900 where119872

119888and119872

119900are circular

structure with radius of 5 and 3 pixels respectively


p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)

Figure 7 Two possible graph cuts result 120572120573 are two labels and 119901119902 are pixels

1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e

Actual image compression rate ()

Figure 8 Image quality score in different compression rate thenumbers in the figure denote average value of quality scores of 20pictures in the same compression rate

Table 2 Detection accuracy in different image quality

Compression rate Quality score 119876 Accuracy5 224 10020 684 9040 790 7060 888 6080 968 0100 1057 0

4 Experimental Results and Discussion

This part firstly exhibits the experimental results and com-pares our results with existing algorithm Then we considerthe situation when the input image is slightly compressed Inthis circumstance there are few conspicuous block artificial

000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise

Figure 9 Image quality score in different noise level

grids noise feature becomes predominated since 120572 = 119891(119876) =0 In order to verify the effectiveness of the proposedmethodwe tested under two situations noise discrepancies fromartificial added noise and from digital cameras

41 Detection Results and Comparison As it is mentionedat the beginning our proposed method can deal with bothcopy-move and splicing forgery with one authenticationprocess Two detection results are shown in Figure 10 themarked white area is detected forged region Our algorithmshows good performance in these two types of forgery

Then we compared our proposed method with existingalgorithm in [9] In the experiment we prepared six setsof test images In each set there were 25 pictures includingintact and fake pictures with copy-move or splicing forgeryThe difference between sets was the image quality-JPEGcompression rate Table 3 shows the comparison of detectionaccuracy between two methods when the image is greatlydegraded by high JPEG compression two methods present


(a) (b)

(c) (d)

(e) (f)

Figure 10 The detection result of copy-move and splicing forgeries (a) and (b) intact pictures (c) and (d) forged pictures and (e) and (f)detection result

valid performance However if the forged image is saved withslightly compression the detection accuracy of Lirsquos methoddrops significantly while our method still maintains highaccuracy Figure 11 shows an instance of detection resultcomparison between two methods

42 Simulation Results In this part we present a simulatedforgery case that the noise is added to implanted regionThissimulation also reflects a real splicing attack that in orderto make the alien area visually resemble the rest part ofpicture noise may be applied Since Photoshop is a popularimage editing tool we add noise to picture with providedfilters by software There are two noise distribution optionsGaussian and uniform and two noise patterns monochromeand colored Therefore four combinations are available andthe user can alter the noise amount in percentage Theexperiment is designed to demonstrate the sensibility of

Table 3 Comparison between two methods

Set Compressionratio Accuracy of Li [9] Accuracy of our

method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100

algorithm which is the lower limit amount of added noisethat can be detected by our method Figure 12 shows thedetection accuracy of four groups each of which contains fiveforged pictures We conclude that the effective lower limit for


(a) (b)

(c) (d)

Figure 11 Comparison between the proposed method and Lirsquos algorithm (a)intact image (b) picture with splicing forgery (c) detectionresult of Lirsquos algorithm and (d) result of our method

Table 4 Combination of ISO speed and respective TP rate Source pictures are taken by Nikon D7000

ISO 100 200 400 800 1600 3200100 mdashlowast 20 90 100 100 100200 20 mdash 40 80 100 100400 90 40 mdash 50 100 100800 100 80 50 mdash 80 1001600 100 100 100 80 mdash 703200 100 100 100 100 70 mdash

020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()

Gaussian distributed monochrome noiseUniform distributed monochrome noiseGaussian distributed color noiseUniform distributed color noise

Figure 12 Finding lower limit amount of added noise that thealgorithm can detect

detection is 14 for Gaussian noise and 22 for uniformnoise regardless of monochrome or colored noise pattern

43 ISO and Detection Results Two image datasets areprepared to verify the effectiveness of our proposed methodIn the first set all source pictures were taken by a NikonD7000 DSLR camera and used to make splicing forgeries incombination of different ISO speed seen in Table 4There are10 forged pictures in the test set The data in this table is thedetection accuracy or true positive rate

The ISO speed setting in camera is discrete without thesame interval and we find that the higher TP rate appears atcombination of two ISO speeds with big gap In order to seethis phenomenon clearly we can see Figure 13The horizontal


Table 5 Combination of ISO speed and respective TP rate Source pictures are taken by Canon 550D

ISO 100 200 400 800 1600 3200100 mdashlowast 30 80 100 100 100200 30 mdash 20 80 100 100400 80 20 mdash 40 90 100800 100 80 40 mdash 60 901600 100 100 90 60 mdash 803200 100 100 100 90 80 mdashlowastnot verified in experiment

100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)

Figure 13 TP rate in different interval stop(s)

100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)

Figure 14 TP rate in different ISO speed

axis is marked by interval stop(s) which denotes the intervalISO speed For instance the interval stop of ISO 100 and 200is 1 this is the same with ISO 1600 and 3200 while that ofISO 200 and 1600 is 3 The average TR rate is calculated fromTables 4 and 5 We conclude that our method shows goodperformance in two or more interval stops

The second experiment is to verify the effectiveness ofdetecting forgery in pictures combined from two differentcameras And in the paper we just show an extremely hardsituation when the source pictures are taken in the sameISO speed Two cameras are Nikon D7000 and Canon 550D

respectively And 10 forged images in the set are used to thetest The TP rate is shown in Figure 14 And the accuracyincreases as the ISO speed rises The reason is that the imageprocessing ability of two camera models is not the same Inlower ISO speed less noise appears in the picture and thisprocessing difference is small therefore the TP rate is verylow at 10 while in high ISO settings the method showseffectiveness again Note that in real situation the ISO of twosource pictures may not be the same only one interval stopwill highly enhance the accuracy as it is shown in the firstexperiment

5 Conclusions

In this paper we concentrated on exposing the two maintypes of imagemanipulation copy-move and splicing forgeryWe proposed an integrated algorithm to locate forged regionsby a single authentication process In ourmethod JPEGblockartificial grids and local noise discrepancies were used togenerate features which were combined with image qualityscore as coefficient Experimental result shows that ourapproach is valid to both highly compressed and high qualitypictures Comparing to existing algorithms our method hascompetitive advantages and a larger range of application

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

The authors would like to thank the referees for their valuablecomments This research was supported in part by theResearch Committee of the University of Macau and theScience and Technology Development Fund of Macau SAR(Project nos 0342010A2 and 0082013A1)

References

[1] J Granty Regina Elwin T S Aditya and S Madhu ShankarldquoSurvey on passive methods of image tampering detectionrdquo inProceedings of the International Conference on Communicationand Computational Intelligence (INCOCCI rsquo10) pp 431ndash436December 2010


[2] J A RediW Taktak and J-L Dugelay ldquoDigital image forensicsa booklet for beginnersrdquoMultimedia Tools andApplications vol51 no 1 pp 133ndash162 2011

[3] I Amerini L Ballan R Caldelli A Del Bimbo and GSerra ldquoA SIFT-based forensic method for copy-move attackdetection and transformation recoveryrdquo IEEE Transactions onInformation Forensics and Security vol 6 no 3 pp 1099ndash11102011

[4] X Pan and S Lyu ldquoDetecting image region duplication usingsift featuresrdquo in Proceedings of the IEEE International Conferenceon Acoustics Speech and Signal Processing (ICASSP rsquo10) pp1706ndash1709 March 2010

[5] P KakarN Sudha andW Ser ldquoExposing digital image forgeriesby detecting discrepancies in motion blurrdquo IEEE Transactionson Multimedia vol 13 no 3 pp 443ndash452 2011

[6] S Hamdy H El-Messiry M Roushdy and E Kahlifa ldquoQuanti-zation table estimation in jpeg imagesrdquo International Journal ofAdvanced Computer Science and Applications vol 1 no 6 pp17ndash23 2010

[7] F Huang J Huang and Y Q Shi ldquoDetecting double JPEG com-pression with the same quantization matrixrdquo IEEE Transactionson Information Forensics and Security vol 5 no 4 pp 848ndash8562010

[8] S D Lin and T Wu ldquoAn integrated technique for splicing andcopy-move forgery image detectionrdquo in Proceedings of the 4thInternational Congress on Image and Signal Processing (CISP rsquo11)vol 2 pp 1086ndash1090 October 2011

[9] W Li Y Yuan and N Yu ldquoPassive detection of doctored JPEGimage via block artifact grid extractionrdquo Signal Processing vol89 no 9 pp 1821ndash1829 2009

[10] W Li N Yu and Y Yuan ldquoDoctored JPEG image detectionrdquo inProceedings of the IEEE International Conference onMultimediaand Expo (ICME rsquo08) pp 253ndash256 June 2008

[11] L G Shapiro and G C Stockman Computer Vision Prentice-Hall New Jersey NY USA 2011

[12] R Achanta A Shaji K Smith A Lucchi P Fua and SSusstrunk ldquoSLIC superpixels compared to state-of-the-artsuperpixel methodsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 34 no 11 pp 2274ndash2282 2012

[13] J Fan H Cao and A C Kot ldquoEstimating EXIF parametersbased onnoise features for imagemanipulation detectionrdquo IEEETransactions on Information Forensics and Security vol 8 no 4pp 608ndash618 2013

[14] H Gou A Swaminathan and M Wu ldquoIntrinsic sensor noisefeatures for forensic analysis on scanners and scanned imagesrdquoIEEE Transactions on Information Forensics and Security vol 4no 3 pp 476ndash491 2009

[15] Y Boykov O Veksler and R Zabih ldquoFast approximate energyminimization via graph cutsrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 23 no 11 pp 1222ndash12392001

[16] V J Hodge and J Austin ldquoA survey of outlier detectionmethodologiesrdquo Artificial Intelligence Review vol 22 no 2 pp85ndash126 2004

[17] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 285ndash296 1975

[18] Z Wang H R Sheikh and A C Bovik ldquoNo referenceperceptual quality assessment of JPEG compressed imagesrdquo inProceedings of the International Conference on Image Processing(ICIP rsquo02) vol 1 pp I477ndashI480 September 2002

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



JPEG pictures with different compression table Althoughcomplex situation in double compression was discussed [7]multicompressions more than twice are still very hard toanalyze

Our goal is to automatically detect copy-move as wellas splicing forgery within a single process without any priorknowledge about the forgery type of the doubtful pictureThe reason is obvious Rather than putting the same pictureinto different algorithm which may be effective on only onecertain type of forgery one single method is time saving andavoids evaluating every detection result which will be veryhard to discern the true output from various results Lin andWu in [8] proposed an integrated method to detect bothcopy-move and splicing forgery But this method just con-nects two separate processes togetherThe errors in the step offorgery type judgment will greatly affect the detection resultActually it is unnecessary to classify the pictures into certainforgery type if there is a tool or feature sensitive to two attackpractices Li et al [9] proposed amethod based on JPEGblockartificial grids (BAG) detection to expose both splicing andcopy-move forgeries But there are two major problems inthat paper the first is that the algorithm must be adjustedbefore applying to different forgery the author revised thesplicing detection algorithm to deal with copy-move practiceHowever in practice we do not know the forgery type fora doubtful image The other defect is that the algorithmis sensitive to highly compressed image and ineffective onhigh quality picture with little compression To overcomethis shortcoming we inducted noise feature to compensatefor BAG algorithm When the picture is less compressedclear BAG is very difficult to extract We consider that localnoise level and category can be used as a feature to identifydifferent source of regions in a picture Inconsistency anddiscrepancy from regions to regions provide the other cluesexcept for BAG to locate forged areaTherefore we created anintegrated feature combining BAG and noise feature to verifythe authenticity of a doubtful picture

The paper is organized as follows Section 2 introducesthe block artificial grids and noise patterns for detectingforgeries and the following Section 3 details our proposedintegrated method In Section 4 experimental results will bepresented to show the effectiveness of ourmethod andwe alsocompared it with existing methods Finally we conclude thispaper in Section 5

2 Block Artificial Grid and Noise Estimation

21 Block Artificial Grid Extraction It is universally knownthat the lossy JPEG compression will introduce some visuallyvertical or horizontal breaks in the imageThese breaks calledblock artificial grid (BAG) appear at the border of each 8 times 8pixel block This property can be used to determine whethera picture is altered or not If the picture is intact blockartificial grids should only present on block borders whilethere is a great possibility that copied and pasted or splicedregions will bring their original BAGs which may appearwithin the 8 times 8 block rather than at borders Some papers[9 10] noticed this and Figure 1 illustrates the phenomenon

Theoretically speaking if we extract all the BAGs from agiven image areas with BAGs within the block border areregarded as forged regions Li et al [9] introduced steps toextract BAGs As it is mentioned before artificial grids arevisually vertical and horizontal lines and they are very weakwhen comparing to the border lines of objects in the pictureAnd themain purpose of extraction procedures is to enhancethese weak lines and to make them visible However linesare also strengthened which may be the edges of objects orjust objects themselves This will interfere with the detectionresult because we only need BAGs To allay the side effectwe preprocessed the doubtful image by excluding the edgesof objects But it should be noticed that BAG can also beregarded as vertical or horizontal edges For preserving BAGswe only excluded edges within certain range

Suppose that 119866 was the grayscale version of image 119868 andthen the edge119864was obtained by119864 = 119866lowast119878 where 119878 representsSobel operator and ldquolowastrdquo denotes convolutionThenwe definedwhether a pixel is excluded using

119877 (119898 119899) =

0 119863 (119864 (119898 119899)) isin [0 120579]

cup [120587

2minus 120579120587

2+ 120579] cup [120587 minus 120579 120587)

1 others

(1)

where 119863(sdot) denotes gradient of the pixel and 119877(119898 119899) = 1means excluded pixels Then we begin to extract BAGs

Firstly weak horizontal edges were extracted by calcu-lating second-order difference of an image For the testimage 119868(119898 119899) absolute second-order difference 119889(119898 119899) wasobtained by

119889 (119898 119899) = |2119868 (119898 119899) minus 119868 (119898 + 1 119899) minus 119868 (119898 minus 1 119899)| (2)

Then all differentials larger than 01 or 119877 = 1 arediscarded In subsequence enlarged horizontal lines areaccumulated from every 33 columns as shown in (3) Thena median filter Med[sdot] is used to refine the result in

119886 (119898 119899) =

16

sum

119894=119899minus16

119889 (119898 119894) (3)

119886119903(119898 119899) = 119886 (119898 119899) minusMed [119886 (119894 119899) | 119898 minus 16 le 119894 le 119898 + 16]

(4)

Weak horizontal edge 119887ℎis further periodical median

filtered as

119887ℎ(119898 119899) = Med [119886

119903(119894 119899) | 119894 = 119898 minus 16

119898 minus 8119898119898 + 8119898 + 16]

(5)

Similarly the vertical BAGs 119887V can also be attracted Asa result final BAG is obtained by adding two componentstogether in

119887 (119898 119899) = 119887ℎ(119898 119899) + 119887V (119898 119899) (6)

22 Noise Estimation Highly compressed by JPEG the pic-ture shows visual block artificial grids across the whole frame


(a) (b)






119904where


119868 = (

119868119888

119904(1 1) sdot sdot sdot 119868

119888

119904(1119873)

d

119868119888

119904(119872 1) sdot sdot sdot 119868

119888

119904(119872119873)

) (7)


119904=




(a) (b) (c) (d)


(a) (b) (c) (d)



119866 = 02989119868119903+ 05870119868

119892+ 01140119868

119887 (8)


119867(119898 119899) = 1 (119898 119899) isin 119864

0 (119898 119899) notin 119864(9)


119885 = 119867 oplus 119881 (10)



119891119888119889(119898 119899) = 119868

119888(119898 119899) minus 119867

119888119889(119898 119899) (11)



119904


119904 120590119888119889

119904)

where

120583119888119889

119904=1

119877sum

119885(119898119899)=0

119891119888119889

119904(119898 119899)

120590119888119889

119904=1

119877( sum

119885(119898119899)=0

(119891119888119889

119904(119898 119899) minus 120583

119888119889

119904)2

)

12

(12)


Input image



BAG map

BAG feature


Bxy



119904of a segment




Input image

SLIC segmentation

Segmented image



Noise feature

Label map

Graph cut


Axy

8 times 8 block

60

50

40

30

20

10




119861119909119910= Max

7

sum

119894=2

119887 (119894 119899) | 2 le 119899 le 7

minusMin7

sum

119894=2

119887 (119894 119899) | 119899 = 1 8

+Max7

sum

119894=2

119887 (119898 119894) | 2 le 119898 le 7

minusMin7

sum

119894=2

119887 (119898 119894) | 119898 = 1 8

(13)





For each 119878119896



119896by graph cut


8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)


Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910


End




119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573




119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)





119901and 119905120573119902



119901and 119905120573119901 We



119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)








Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)






119904where


119868 = (

119868119888

119904(1 1) sdot sdot sdot 119868

119888

119904(1119873)

d

119868119888

119904(119872 1) sdot sdot sdot 119868

119888

119904(119872119873)

) (7)


119904=




(a) (b) (c) (d)


(a) (b) (c) (d)



119866 = 02989119868119903+ 05870119868

119892+ 01140119868

119887 (8)


119867(119898 119899) = 1 (119898 119899) isin 119864

0 (119898 119899) notin 119864(9)


119885 = 119867 oplus 119881 (10)



119891119888119889(119898 119899) = 119868

119888(119898 119899) minus 119867

119888119889(119898 119899) (11)



119904


119904 120590119888119889

119904)

where

120583119888119889

119904=1

119877sum

119885(119898119899)=0

119891119888119889

119904(119898 119899)

120590119888119889

119904=1

119877( sum

119885(119898119899)=0

(119891119888119889

119904(119898 119899) minus 120583

119888119889

119904)2

)

12

(12)


Input image



BAG map

BAG feature


Bxy



119904of a segment




Input image

SLIC segmentation

Segmented image



Noise feature

Label map

Graph cut


Axy

8 times 8 block

60

50

40

30

20

10




119861119909119910= Max

7

sum

119894=2

119887 (119894 119899) | 2 le 119899 le 7

minusMin7

sum

119894=2

119887 (119894 119899) | 119899 = 1 8

+Max7

sum

119894=2

119887 (119898 119894) | 2 le 119898 le 7

minusMin7

sum

119894=2

119887 (119898 119894) | 119898 = 1 8

(13)





For each 119878119896



119896by graph cut


8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)


Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910


End




119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573




119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)





119901and 119905120573119902



119901and 119905120573119901 We



119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)








Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b) (c) (d)


(a) (b) (c) (d)



119866 = 02989119868119903+ 05870119868

119892+ 01140119868

119887 (8)


119867(119898 119899) = 1 (119898 119899) isin 119864

0 (119898 119899) notin 119864(9)


119885 = 119867 oplus 119881 (10)



119891119888119889(119898 119899) = 119868

119888(119898 119899) minus 119867

119888119889(119898 119899) (11)



119904


119904 120590119888119889

119904)

where

120583119888119889

119904=1

119877sum

119885(119898119899)=0

119891119888119889

119904(119898 119899)

120590119888119889

119904=1

119877( sum

119885(119898119899)=0

(119891119888119889

119904(119898 119899) minus 120583

119888119889

119904)2

)

12

(12)


Input image



BAG map

BAG feature


Bxy



119904of a segment




Input image

SLIC segmentation

Segmented image



Noise feature

Label map

Graph cut


Axy

8 times 8 block

60

50

40

30

20

10




119861119909119910= Max

7

sum

119894=2

119887 (119894 119899) | 2 le 119899 le 7

minusMin7

sum

119894=2

119887 (119894 119899) | 119899 = 1 8

+Max7

sum

119894=2

119887 (119898 119894) | 2 le 119898 le 7

minusMin7

sum

119894=2

119887 (119898 119894) | 119898 = 1 8

(13)





For each 119878119896



119896by graph cut


8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)


Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910


End




119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573




119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)





119901and 119905120573119902



119901and 119905120573119901 We



119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)








Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Input image



BAG map

BAG feature


Bxy



119904of a segment




Input image

SLIC segmentation

Segmented image



Noise feature

Label map

Graph cut


Axy

8 times 8 block

60

50

40

30

20

10




119861119909119910= Max

7

sum

119894=2

119887 (119894 119899) | 2 le 119899 le 7

minusMin7

sum

119894=2

119887 (119894 119899) | 119899 = 1 8

+Max7

sum

119894=2

119887 (119898 119894) | 2 le 119898 le 7

minusMin7

sum

119894=2

119887 (119898 119894) | 119898 = 1 8

(13)





For each 119878119896



119896by graph cut


8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)


Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910


End




119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573




119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)





119901and 119905120573119902



119901and 119905120573119901 We



119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)








Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







For each 119878119896



119896by graph cut


8119910

119895=8119910minus7sum8119909

119894=8119909minus7119871(119894 119895)


Calculate 119862119909119910= 120572 sdot 119861

119909119910+ (1 minus 120572) sdot 119860

119909119910


End




119901119863119901(120572) + sum

119902isin119873119901

119902notin119875120572120573119881(120572 119891

119902) 119901 isin 119875

120572120573

119905120573

119901119863119901(120573) + sum

119902isin119873119901

119902notin119875120572120573119881(120573 119891

119902) 119901 isin 119875

120572120573

119890119901119902

119881(120572 120573)119901 119902 isin 119873

119901 119902 isin 119875120572120573




119864 (119891) = sum

119901119902isin119873

119881119901119902(119891119901 119891119902) + sum

119901isin119875

119863119901(119891119901) (14)





119901and 119905120573119902



119901and 119905120573119901 We



119908120572=10038171003817100381710038171003817119865119904minus 11986510038171003817100381710038171003817 119908

120573=1003817100381710038171003817119865119904 minus 119865max

1003817100381710038171003817 (15)








Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





Input image

Quality estimation

Quality score Q


Coefficient 120572


Detection result

Rough map

Axy 8 times 8







119896= 0) And


119896 (119898 119899) isin 119878


119909119910


119860119909119910=1

64

8119910

sum

119895=8119910minus7

8119909

sum

119894=8119909minus7

119871 (119894 119895) (16)


119909119910=

120572119861119909119910+ (1 minus 120572)119860







119891 (119876)

=

1 119876 lt 2

minus00213119876 + 10469 2 le 119876 lt 69

minus029801198762+ 42584119876 minus 142952 69 le 119876 lt 89

0 119876 ge 89

(17)


119888)∘119872119900 where119872

119888and119872

119900are circular



p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




p q

t120572p

t120573p

t120572q

t120573q

120572

120573

Cut

epq

(a)

p q

120572

120573

Cut

(b)


1095 988

911 829

733

294

000

200

400

600

800

1000

1200

1400

0 20 40 60 80 100

Imag

e qua

lity

scor

e







000

200

400

600

800

1000

1200

0 005 01 015 02 025 03 035 04

Imag

e qua

lity

scor

e

Added noise






(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)

(c) (d)

(e) (f)






method1 100 0 802 80 0 843 60 56 804 40 84 885 20 92 926 5 100 100



(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)

(c) (d)




020406080

100

0 05 1 15 2 25 3Det

ectio

n ac

cura

cy (

)

Added noise ()









100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






100

90

80

70

60

50

40

30

20

10

01

46

85

52

925100100

100100

100100

2 3 4 5

TP-NikonTP-Canon

Interval stop (s)

Aver

age T

P (

)


100

90

808080

7070

6060

5050

4040

30

20

1010

0100 200 400 800 1600 3200 6400

ISO speed

TP (

)





5 Conclusions




Acknowledgments


References



























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Research Article Digital Image Forgery Detection...

Documents

Transcript of Research Article Digital Image Forgery Detection...