Region Duplication Forgery Detection Technique Based on SURF … · 2017-11-15 · order derivative...

Hindawi Publishing CorporationThe Scientific World JournalVolume 2013 Article ID 267691 8 pageshttpdxdoiorg1011552013267691

Research ArticleRegion Duplication Forgery Detection TechniqueBased on SURF and HAC

Parul Mishra1 Nishchol Mishra1 Sanjeev Sharma1 and Ravindra Patel2

1 School of Information Technology RGPV Bhopal Madhya Pradesh 462036 India2University Institute of Technology RGPV Bhopal Madhya Pradesh 462036 India

Correspondence should be addressed to Parul Mishra parul123 mishrayahoocom

Received 16 August 2013 Accepted 17 September 2013

Academic Editors H Cheng H-E Tseng and Y Zhang

Copyright copy 2013 Parul Mishra et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Region duplication forgery detection is a special type of forgery detection approach and widely used research topic under digitalimage forensics In copy move forgery a specific area is copied and then pasted into any other region of the image Due to theavailability of sophisticated image processing tools it becomes very hard to detect forgery with naked eyes From the forged regionof an image no visual clues are often detected Formaking the tamperingmore robust various transformations like scaling rotationillumination changes JPEG compression noise addition gamma correction and blurring are applied So there is a need for amethod which performs efficiently in the presence of all such attacks This paper presents a detection method based on speededup robust features (SURF) and hierarchical agglomerative clustering (HAC) SURF detects the keypoints and their correspondingfeatures From these sets of keypoints grouping is performed on the matched keypoints by HAC that shows copied and pastedregions

1 Introduction

Today the use of digital images is increasing rapidly in almostevery area of human life like in education software com-panies television businesses journalism medical imagingand social media It is easy to learn and understand anythingvisually rather than only reading or listening Another aspectis that generally visual information is believed to be true Butas the technology advances and lots of sophisticated imageprocessing tools are available it becomes very easy to editvisual information Some of the tools are Adobe PhotoshopGIMP Macromedia Freehand and Corel Paint Shop [1 2]A big question arises how to distinguish the photographicimages from the photorealistic ones [3]

Digital image forensic is a branch that deals with crimeswhere images are used as a prime evidence in the court oflaw Forensic sciences have methods to identify the sourcedevice for example camera scanner and so forth with theirparticular model If any tampering is done on the image thenit can also be detected Tampering an image means eitheradding or removing some information from an image so thatthe original meaning will be changed [4]

If the alteration is intentional and related with somekind of benefits it can be termed as a digital image forgeryForgery detection methods are mainly divided into twodifferent categories active and passive In active methodsome information for example digital watermark or digitalsignature is preembedded into the image This procedureis performed for the sake of providing authenticity to animage But the images distributed on the web do not con-tain always preembedding information To overcome thedrawback of active methods passive methods have beendeveloped [5] The tampering again is of two types eitherit is performed on the same image or on multiple imagesIf some area is copied and pasted into another area of thesame image it is known as copy move forgery or regionduplication forgery When two or more images are involvedand their combination produces a fake image then it is calledsplicing or photomontage [6] Farid gives lots of examplesof real incidents with image forgery [7] On July 2008 aforged image of four Iranian missiles was posted on theweb and published in newspapers [8] Egyptrsquos newspaper Al-Ahram on Sept 2010 published a forged image In this imagepresident Mubarak was leading the group instead of Barak

2 The Scientific World Journal

Moment invariantBLUR Zernike Hu

LPM

DCT

PCT

DWT

PCA

FMT

SIFT SURF

Keypoint basedmethods

Block basedmethods

Region duplication forgerydetection techniques

Figure 1 Classification of copy move forgery detection techniques

Obama at the White House during Middle East peace talks[9]

The rest of the paper is organized as follows Section 2presents the related work Section 3 describes the proposedmethod for duplicate detection The results of forgery detec-tion by experimental evaluation are presented in Section 4and Section 5 describes the conclusion of the paper

2 Related Work

In region duplication detection the forged region can beidentified by applying proper detection techniques Thesetechniques are classified into block based and keypoint basedmethods [10] The classification of duplication detectionmethods is illustrated in Figure 1

21 Block Based Methods In this approach an image isdivided into different overlapping blocks of fixed size it isassumed that the block size is smaller than the duplicatedregion After dividing the image into different block sizesfeatures are extracted from it by applying different methodsThese features are then matched with other features of eachblock A match indicates the probability of forgery Fridrichet al [11] discuss the exhaustive search and autocorrelation offorgery detection Furthermore they applied discrete cosinetransform (DCT) on each separate block DCT is appliedfrom the upper left corner to the bottom right corner Forreducing the computation features are extracted from the lowfrequency component In thismethod feature dimension sizeis 64

Popescu and Farid [12] presented a method based onprincipal component analysis (PCA) that reduces the dimen-sional size to 32 Extracted features are lexicographicallysorted therefore matched features come closer to each otherThis method is robust towards compression and additivenoise Bayram et al [13] proposed Fourier Mellin transforms(FMT) for copy move forgery detection Features extracted

are of length 45 and they are rotation invariant only to somedegree Bloom filter was used here instead of lexicographicalsorting which reduces the detection time

Moment invariant features are insensitive towards alltransformation hence it can be used to detect the regionduplication Blur moment invariant detects the forgery in thepresence of blur degradation and its performance remainsunaffected by additive zero mean noise Each separate blockis represented by a feature vector whose dimension length is24 in case of gray levelTheKD treemethod is applied here fornearest neighbour searching and then similar blocks denotethe duplicated regions [14] Ryu et al [15] proposed a methodwhich utilizes rotation invariant Zernike moment It alsogives significant performance in case of JPEG compressionblurring and noise contamination In [16] the authorsapplied Gaussian pyramid on image to decompose it intodifferent scales After that each scale space is divided intoseparate circular blocks From each circular block featuresare extracted by Hu moment of length 4 Features are thensorted and matching is performed on it The performanceof this method is not changed in the presence of a rotationaltransformation

Li et al [17] applied discrete wavelet transform (DWT)and singular value decomposition (SVD) to the image FirstlyDWT is applied to each block that reduces the size of theblock Then SVD is applied on low frequency componentswhich reduce the feature dimensional size to 4 After thatfeatures are lexicographically sorted and matched In [18]the image is divided into overlapping patches Features areextracted from each patch by applying PCT PCT utilizesthe orthogonal properties so that features of PCT are morecompact It is a rotational invariantmethod and also performsbetter in the presence of noise

Bravo-Solorio and Nandi [19] mapped all the pixels ofoverlapping blocks into log polar coordinates Then angleorientation creates a 1D descriptor by doing this synchro-nization problem can be removed Feature vectors are cal-culated from particular blocks depending on the colour andluminance factor This method is robust towards reflectionrotation and scale changes

22 Keypoints Based Methods In this approach variouskeypoints are selected from an image Feature descriptorsare calculated from each of these keypoints For detectingthe duplication forgery matching is performed on keypointfeature descriptors Lowe [20] invented SIFT which detectskeypoints and features from an image SIFT application existsin various fields It performs better in comparison to previousdescriptors [21]

In [22] SIFT is applied into the region duplicationdetection method In [23] the authors suggested that thekeypoint matching method suffers from some problems Forremoving these problems SIFT cluster matching is proposedwhere objects are matched rather than the point Points aregrouped here using agglomerative hierarchical clustering

Pan and Lyu [24] also applied SIFT in their work Foravoiding search from close adjacency search is applied farfrom 11 times 11 pixel window whose center is at the keypoint

The Scientific World Journal 3

They also applied Random Sample Consensus (RANSAC) foraffine transformation detection

Amerini et al [25] detect the duplication forgery and alsoestimate the transformation by RANSAC They developedg2NN nearest neighbour searching for multiple copy pastedetection Their method is robust to all transformationattacks Their procedure also works effectively for splicingattack detection The similar work is done in [26] where theauthors usedMPEG7 image signature tools for extracting thefeatures Least Median of Squares (LMedS) algorithm is usedinstead of RANSAC for estimating the geometrical transfor-mations Recently Amerini et al [27] proposed J linkage foreffective clustering In [28] the authors applied resamplingtraces with SIFT to distinguish the original region from thepasted region in a tamper image SIFT ring descriptors areapplied to an image for detecting the tampering The size offeature dimension is reduced to 24 from 128 that increasesthe speedThese feature descriptors are rotationally invariant[29] SURF is another keypoint based method which detects64 feature descriptors Obtained keypoints are divided intotwo subsetsMatching procedure is applied and repeated untilone keypoint remains in a set [30] In [31] SURF with KDtree method are used for the detection of a particular forgedregion In this proposed work HAC method is applied withSURF for a more accurate result in terms of all attacks

3 Proposed Method

This proposed method is based on a SURF algorithm for thedetection of keypoints and for extracting their correspond-ing feature descriptors Matching is performed in betweenselected keypoints by applying best bin first search procedure

For detecting the duplicated regions HAC technique isapplied The whole procedure of proposed work is depictedin Figure 2 and the related algorithm for detection methodis described in Section 31 An input image is inserted to thedetection system and the output is imaged with duplicatedregions if it is forged The first block of the detectionframework is keypoint detection and feature extraction thatwill be explained in Section 32 After that matching isperformed among selected keypoints the procedure forkeypoint matching will be described in Section 33 At lastclustering algorithm is applied on the matched keypointswhich will be explained in Section 34

31 Region Duplication Detection Algorithm If the imagesuffers from duplication forgery then it contains at least twosame regions one is copied and the other is pasted regionThe overall technique for detecting the duplicated region isas followsInput imageOutput detected duplicate regions with image

(1) If RGB image then converted into gray scale(2) Applying SURF method

(a) Keypoints are detected from an image(1 2 3 119872)

(b) From the above detected keypoints features areextracted (119873

1 1198732 1198733 119873

119872)

(c) This matrix is stored in a variable119863 = 119872 times119873

(3) For each 119894 = 1 to119872for 119895 = 1 to119872

(a) If 119894 = 119895 then go to step (b) else return(b) Dot products are calculated between each fea-

ture descriptor

End of For 119895

(c) Inverse cosine angle of dot products will becomputed

(d) Sorting is applied on the result and values arestored

[Value index] = sort (cosminus1(dot prods))

(e) If (Value(1)Value(2)) lt 06 then match existsand index will be stored

Else index = 0End of For 119894

(4) For each keypoint

(a) If match exists then go to step (b)

else return

(b) If the matched points are far from 10 times 10 squareregion then go to step (c)

else return

(c) Store the coordinates of matched points in119898 by119899 data matrix119883 and set flag = 1

End of For(5) If flag gt 0 then

(a) euclidean distance computed between each pairof objects of119883

Dist (119860 119861) = radic(1199092minus 1199091)2

+ (1199102minus 1199101)2

(1)

where 119860 = (1199091 1199101) 119861 = (119909

2 1199102)

(b) linkage function is applied for linking theobjects into hierarchal tree

(c) the smallest height is taken for cutting thehierarchal tree into clusters

(d) a line is drawn between the matched objectsfrom different clusters

(e) objects of different clusters are shown from thedifferent colours


Input image

Keypoint detection

Feature extraction

SURF

No

Sort the inverse cosine angle ofdot products

Dot products between featuredescriptors

Yes

Two different keypoints aretaken

No

Line drawn between the matchedobjects of different clusters

Duplicated regionsshown from clusters

If allkeypoints are

processed

Match exists store thecoordinates

Yes

Euclidean distance computedbetween all matched keypoints

Hierarchal tree is created on thebasis of distance

Clusters are constructed

HAC

If ratio of twonearest

neighbourslt06

Figure 2 Flow chart of proposed work

A B

C D

sum = A + Dminus (C + B)

Figure 3 Integral image calculation by rectangular region

32 Keypoint Detection and Feature Extraction Bay et al [32]proposed SURF method whose computation is faster thanSIFT How keypoints are detected and feature descriptor isgenerated from SURF are discussed below

321 Integral Image Integral image increases the computa-tion speed as well as the performance its value is calculatedfrom an upright rectangular area

In Figure 3 the sum of all pixel intensities is calculated bythe formula which is written in the rectangular area whose

vertices are119860119861119862 and119863 Suppose that an input image 119868 anda point (119909 119910) are givenThe integral image 119868

Σis calculated by

the sum of the values between the point and the origin Thefollowing formula is used to calculate the integral image

119868Σ(119909 119910) =

119894le119909

sum

119894=0

119895le119910

sum

119895=0

119868 (119909 119910) (2)

322 Keypoint Detection This step requires scale spacegeneration for the extraction of keypoints In SURF Laplacianof Gaussian is approximated with a box filter Convolution isapplied to an imagewith varying size box filter for creating thescale space After constructing the scale space determinant ofthe Hessian matrix is calculated for detecting the extremumpoint If determinant of the Hessian matrix is positive thatmeans both the Eigen values are of the same sign eitherboth are negative or both are positive In case of the positiveresponse points will be taken as extrema otherwise it will bediscarded

Hessian matrix is represented by

119867(119909 120590) = [119871119909119909

(119909 120590) 119871119909119910(119909 120590)

119871119909119910(119909 120590) 119871

119910119910(119909 120590)

] (3)


where 119871119909119909(119909 120590) is the convolution of the Gaussian second

order derivative with the image 119868 in point 119909 and simi-larly 119871

119909119910(119909 120590) and 119871

119910119910(119909 120590) These derivatives are called

Laplacian of Gaussian The approximate determinant of theHessian matrix is calculated by

det (119867approx) = 119863119909119909119863119910119910

minus (09119863119909119910)2

(4)

323 Orientation Assignment At first a circular area is con-structed around the keypoints Then Haar wavelets are usedfor the orientation assignment It also increases the robust-ness and decreases the computational cost Haar wavelets arefilters that detect the gradients in 119909 and 119910 directions In orderto make rotation invariant a reproducible orientation for theinterest point is identified A circle segment of 60∘ is rotatedaround the interest point The maximum value is chosen as adominant orientation for that particular point

324 Feature Descriptor Generation For generating thedescriptors first construct a square region around an interestpoint where interest point is taken as the center point Thissquare area is again divided into 4 times 4 smaller subareas Foreach of these cells Haar wavelet responses are calculatedHere 119889

119909termed as horizontal response and 119889

119910as vertical

response For each of these subregions 4 responses arecollected as

Vsubregion = [sum119889119909sum119889119910sum |119889119909| sum1003816100381610038161003816119889119910

1003816100381610038161003816] (5)

So each subregion contributes 4 values Therefore thedescriptor is calculated as 4 times 4 times 4 = 64

33 Keypoint Matching A set of keypoints and their corre-sponding feature descriptors are obtained from SURF Thecomparison is performed between each keypoint with theremaining other keypoints feature descriptor As matchingthese keypoints with their high dimensional feature vector 64takes time therefore best bin first (BBF) method is chosenfor selecting two nearest neighbours [33] Dot products arecalculated between each keypoint feature descriptor withthe others After that sort the inverse cosine angles of dotproducts Store their values as well as their correspondingindex number The ratio between two nearest neighboursvalue is compared to a predefined threshold In this workthe threshold is set to 06 because above this value theprobability of false matches arises If the ratio is less thanthe given threshold they satisfy the similarity criterion andmatch exists In case ofmatching their relative index numberwill be stored This procedure continues for all keypoints

34 Keypoint Clustering HAC is also known as hierarchyof clusters in which each keypoint behaves as a singlecluster at the starting stage Euclidean distance betweeneach keypoint with the remaining other keypoints will becalculatedMerging is performed if two clusters are dissimilarto each other This step is repeated until there is one clusterleft or dissimilarity criterion unsatisfied [34] Single average

and ward methods are types of linkage used for merging andcreating a hierarchal tree

Single Linkage It uses the smallest distance between objectsin two clusters

119889 (119860 119861) = min (dist (119909119860119894 119909119861119895)) (6)

Average Linkage It uses the average distance between all pairsof objects in the two clusters

119889 (119860 119861) =1

119899119860119899119861

119899119860

sum

119894=1

119899119861

sum

119895=1

dist (119909119860119894 119909119861119895) (7)

Ward Linkage It is based on the increment or decrementin the value of error sum of squares (ESS) In other wordsdistance between the clusters is the difference between theESS for unified cluster and ESS of the individual clusters

119889 (119860 119861) = ESS (119860119861) minus [ESS (119860) + ESS (119861)] (8)

where

ESS (119860) =119899119860

sum

119894=1

10038161003816100381610038161003816119909119860119894

minus 119909119860

10038161003816100381610038161003816

2

(9)

Here119860119861 indicates the combined cluster 119899119860indicates number

of objects in cluster 119860 119899119861indicates number of objects in

cluster B 119909119860119894

indicates 119894th object in the cluster 119860 and 119909119860

indicates centroids of cluster whose value is calculated by

119909119860=

1

119899119860

119899119860

sum

119894=1

119909119860119894 (10)

4 Experimental Results

In this section experiment of duplication detection is per-formed on theMICC-F220 dataset [35]This dataset contains220 images from them 110 are real and 110 are fake 10different combinations of scaling and rotation attacks arealready applied to each forged image of the dataset [25]Images shown in Figure 4 represent the detection results inthe presence of various scaling and rotation attacks

For checking the robustness of this method we applieddifferent attacks on images Figure 5 depicts the detectionresult in the presence of compression for JPEG quality factor20 40 60 and 80 respectively Figure 6 denotes the detectionresults in addition of white Gaussian noise whose SNR valuesare 20 30 40 and 50 respectively In Figure 7 detectionresults are shown in the presence of Gaussian blurring Thevalue of the window size is 5 times 5 7 times 7 and the value of120590 is taken as 05 1 Figure 8 represents the detection resultsin the presence of Gamma correction values 12 14 16and 18 In all these images copied and pasted regions arerepresented separately by clusters A line drawn between twokey points indicates that this point matches with each other


12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)

Figure 4 Copy paste detection results in the presence of 10 different rotation and scaling attacks applied on the image

12

(a)

12

(b)

12

(c)

12

(d)

Figure 5 Copy paste detection results for compression (a) JPEG image quality factor 20 (b) JPEG image quality factor 40 (c) JPEG imagequality factor 60 and (d) JPEG image quality factor 80

12

(a)

12

(b)

12

(c)

12

(d)

Figure 6 Copy paste detection results for noise addition (a) SNR value 20 (b) SNR value 30 (c) SNR value 40 and (d) SNR value 50

12

(a)

12

(b)

12

(c)

12

(d)

Figure 7 Copy paste detection results for blurring (a) window size 5times5 120590 = 05 (b) window size 5times5 120590 = 1 (c) window size 7times7 120590 = 05and (d) window size 7 times 7 120590 = 1


12

(a)

12

(b)

12

(c)

12

(d)

Figure 8 Copy paste detection results for (a) gamma value = 12 (b) gamma value = 14 (c) gamma value = 16 and (d) gamma value = 18

Table 1 TPR FPR values () and processing time (average perimage) for each method

Methods FPR TPR Time (s)Fridrich et al [11] 84 89 29469Popescu and Farid [12] 86 87 7097Amerini et al [25] 8 100 494SURF and HAC 364 7364 285

The performance of detection method is measured in termsof true positive rate (TPR) false positive rate (FPR) and timecomplexity where

TPR =images detected as forged being forged

total number of forged images (11)

FPR =images detected as forged being original

total number of original images (12)

TPR is the percentage of forged images which arecorrectly identified FPR is the percentage of the originalimage which is wrongly identified as a tampered

The values of FPR TPR and time (in seconds) for SURFand HAC methods will be computed Then a comparisonis performed with other methods The starting three rowsshown in Table 1 are taken from [25] as a benchmark andthe fourth row represents the value obtained from SURFand HAC based methods Figure 9 represents the result ingraphical formThe graph indicates that this method reducesthe FPR rate as well as the time complexity FPR value isapproximately 4 which is lower than DCT [11] and PCA [12]methods Also the time required to detect the forgery is verylow compared to [11] and [12] TPR value is low which showsone drawback of this method

5 Conclusion and Future Work

In this paper a method was presented for detecting theduplicate region based on SURF andHACThe integral imageused in SURF reduces the time complexity SURF has lessfeature descriptor dimensional size So that matching appliedon SURF descriptor is faster and increases the computa-tion speed as well The Haar wavelets are used for featuredescriptors computation from each keypoint so descriptorsare robust to illumination changes The experimental results

0

50

100

150

200

250

300

Fridrichet al [11]

Popescu and Farid Ameriniet al [25]

SURF andHAC

FPRTPRTime (s)

[12]

Figure 9 Performance of the methods represented in graph

show that SURF feature descriptors are invariant towardsdifferent combination of scaling and rotation In the presenceof JPEG compression Gaussian noise addition and gammacorrection attack this method gives good result HAC is usedhere for creating the regions from matched keypoints HACis easy to implement and create regions in less time butsatisfactory result is not obtained in terms of the true positiverate So in the future we would like to replace clusteringwith suitable image segmentation technique also we wantto utilize this method for multiple duplication detection ina single image

Conflict of Interests

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

References

[1] S Bayram H T Sencar and N Memon ldquoA survey of copy-move forgery detection techniquesrdquo in Proceedings of the IEEEWestern New York Image Processing Workshop pp 538ndash542September 2008

[2] W Luo Z Qu F Pan and J Huang ldquoA survey of passivetechnology for digital image forensicsrdquo Frontiers of ComputerScience in China vol 1 no 2 pp 166ndash179 2007


[3] S Lyu and H Farid ldquoHow realistic is photorealisticrdquo IEEETransactions on Signal Processing vol 53 no 2 pp 845ndash8502005

[4] J A Redi W Taktak and J Dugelay ldquoDigital image forensics abooklet for beginnersrdquo Multimedia Tools and Applications vol51 no 1 pp 133ndash162 2011

[5] H Farid ldquoImage forgery detectionrdquo IEEE Signal ProcessingMagazine vol 26 no 2 pp 16ndash25 2009

[6] M Sridevi C Mala and S Sanyam ldquoComparative study ofimage forgery and copy-move techniquesrdquo in Advances inComputer Science Engineering amp Applications pp 715ndash723Springer Berlin Germany 2012

[7] H Farid ldquoSeeing is not believingrdquo IEEE Spectrum vol 46 no8 pp 44ndash51 2009

[8] M Nizza and P J Lyons In an Iranian Image A Missile TooMany The Lede News Blog The New York Times New YorkNY USA 2008

[9] S D Mahalakshmi K Vijayalakshmi and S PriyadharsinildquoDigital image forgery detection and estimation by exploringbasic image manipulationsrdquo Digital Investigation vol 8 no 3pp 215ndash225 2012

[10] V Christlein C Riess J Jordan and E Angelopoulou ldquoAn eval-uation of popular copy-move forgery detection approachesrdquoIEEE Transactions on Information Forensics and Security vol 7no 6 pp 1841ndash1854 2012

[11] A J Fridrich B D Soukal and A J Lukas ldquoDetection of copy-move forgery in digital imagesrdquo in Proceedings of the DigitalForensic Research Workshop Cleveland Ohio USA August2003

[12] A C Popescu and H Farid ldquoExposing digital forgeries bydetecting duplicated image regionsrdquo Tech Rep TR2004-515Department of Computer Science Dartmouth College 2004

[13] S BayramH T Sencar andNMemon ldquoAn efficient and robustmethod for detecting copy-move forgeryrdquo in Proceedings of theIEEE International Conference on Acoustics Speech and SignalProcessing (ICASSP rsquo09) pp 1053ndash1056 IEEE April 2009

[14] B Mahdian and S Saic ldquoDetection of copy-move forgery usinga method based on blur moment invariantsrdquo Forensic ScienceInternational vol 171 no 2 pp 180ndash189 2007

[15] S J Ryu M J Lee and H K Lee ldquoDetection of copy-rotate-move forgery using Zernike momentsrdquo in Information Hidingpp 51ndash65 Springer Berlin Germany 2010

[16] G Liu J Wang S Lian and Z Wang ldquoA passive imageauthentication scheme for detecting region-duplication forgerywith rotationrdquo Journal of Network and Computer Applicationsvol 34 no 5 pp 1557ndash1565 2011

[17] G Li Q Wu D Tu and S Sun ldquoA sorted neighborhoodapproach for detecting duplicated regions in image forgeriesbased on DWT and SVDrdquo in Proceedings of the IEEE Inter-national Conference on Multimedia and Expo (ICME rsquo07) pp1750ndash1753 IEEE July 2007

[18] Y Li ldquoImage copy-move forgery detection based on polarcosine transform and approximate nearest neighbor searchingrdquoForensic Science International vol 224 pp 159ndash367 2012

[19] S Bravo-Solorio and A K Nandi ldquoAutomated detection andlocalisation of duplicated regions affected by reflection rotationand scaling in image forensicsrdquo Signal Processing vol 91 no 8pp 1759ndash1770 2011

[20] D G Lowe ldquoDistinctive image features from scale-invariantkeypointsrdquo International Journal of Computer Vision vol 60 no2 pp 91ndash110 2004

[21] K Mikolajczyk and C Schmid ldquoA performance evaluation oflocal descriptorsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 27 no 10 pp 1615ndash1630 2005

[22] H Huang W Guo and Y Zhang ldquoDetection of copy-moveforgery in digital images using sift algorithmrdquo in Proceedingsof the Pacific-Asia Workshop on Computational Intelligence andIndustrial Application (PACIIA rsquo08) vol 2 pp 272ndash276 IEEEDecember 2008

[23] E Ardizzone A Bruno and G Mazzola ldquoDetecting multiplecopies in tampered imagesrdquo in Proceedings of the 17th IEEEInternational Conference on Image Processing (ICIP rsquo10) pp2117ndash2120 IEEE September 2010

[24] X Pan and S Lyu ldquoRegion duplication detection using imagefeature matchingrdquo IEEE Transactions on Information Forensicsand Security vol 5 no 4 pp 857ndash867 2010

[25] 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

[26] P Kakar and N Sudha ldquoExposing postprocessed copy-pasteforgeries through transform-invariant featuresrdquo IEEE Transac-tions on Information Forensics and Security vol 7 no 3 pp1018ndash1028 2012

[27] I Amerini L Ballan R Caldelli A del Bimbo L del Tongoand G Serra ldquoCopy-move forgery detection and localization bymeans of robust clusteringwith J-linkagerdquo Signal Processing vol28 no 6 pp 659ndash669 2013

[28] D Vazquez-Padın and F Perez-Gonzalez ldquoExposing originaland duplicated regions using SIFT features and resamplingtracesrdquo in Digital Forensics and Watermarking pp 306ndash320Springer Berlin Germany 2012

[29] L N Zhou Y B Guo and X G You ldquoBlind copy-pastedetection using improved SIFT ring descriptorrdquo in DigitalForensics and Watermarking pp 257ndash267 Springer BerlinGermany 2012

[30] X Bo W Junwen L Guangjie and D Yuewei ldquoImage copy-move forgery detection based on SURFrdquo in Proceedings ofthe 2nd International Conference on Multimedia InformationNetworking and Security (MINES rsquo10) pp 889ndash892 IEEENovember 2010

[31] B L Shivakumar and S S Baboo ldquoDetection of region duplica-tion forgery in digital images using SURFrdquo International Journalof Computer Science Issues vol 8 no 4 p 199 2011

[32] H Bay T Tuytelaars and L van Gool ldquoSurf speeded uprobust featuresrdquo in Computer Vision-ECCV 2006 pp 404ndash417Springer Berlin Germany 2006

[33] J S Beis and D G Lowe ldquoShape indexing using approximatenearest-neighbour search in high-dimensional spacesrdquo in Pro-ceedings of the 1997 IEEE Conference on Computer Vision andPattern Recognition pp 1000ndash1006 June 1997

[34] T Hastie R Tibshirani and J J H Friedman The Elements ofStatistical Learning Volume 1 Springer New York NY USA2001

[35] httpwwwmiccunifiitballanresearchimage-forensics


Moment invariantBLUR Zernike Hu

LPM

DCT

PCT

DWT

PCA

FMT

SIFT SURF

Keypoint basedmethods

Block basedmethods

Region duplication forgerydetection techniques

Figure 1 Classification of copy move forgery detection techniques

Obama at the White House during Middle East peace talks[9]

The rest of the paper is organized as follows Section 2presents the related work Section 3 describes the proposedmethod for duplicate detection The results of forgery detec-tion by experimental evaluation are presented in Section 4and Section 5 describes the conclusion of the paper

2 Related Work

In region duplication detection the forged region can beidentified by applying proper detection techniques Thesetechniques are classified into block based and keypoint basedmethods [10] The classification of duplication detectionmethods is illustrated in Figure 1

21 Block Based Methods In this approach an image isdivided into different overlapping blocks of fixed size it isassumed that the block size is smaller than the duplicatedregion After dividing the image into different block sizesfeatures are extracted from it by applying different methodsThese features are then matched with other features of eachblock A match indicates the probability of forgery Fridrichet al [11] discuss the exhaustive search and autocorrelation offorgery detection Furthermore they applied discrete cosinetransform (DCT) on each separate block DCT is appliedfrom the upper left corner to the bottom right corner Forreducing the computation features are extracted from the lowfrequency component In thismethod feature dimension sizeis 64

Popescu and Farid [12] presented a method based onprincipal component analysis (PCA) that reduces the dimen-sional size to 32 Extracted features are lexicographicallysorted therefore matched features come closer to each otherThis method is robust towards compression and additivenoise Bayram et al [13] proposed Fourier Mellin transforms(FMT) for copy move forgery detection Features extracted

are of length 45 and they are rotation invariant only to somedegree Bloom filter was used here instead of lexicographicalsorting which reduces the detection time

Moment invariant features are insensitive towards alltransformation hence it can be used to detect the regionduplication Blur moment invariant detects the forgery in thepresence of blur degradation and its performance remainsunaffected by additive zero mean noise Each separate blockis represented by a feature vector whose dimension length is24 in case of gray levelTheKD treemethod is applied here fornearest neighbour searching and then similar blocks denotethe duplicated regions [14] Ryu et al [15] proposed a methodwhich utilizes rotation invariant Zernike moment It alsogives significant performance in case of JPEG compressionblurring and noise contamination In [16] the authorsapplied Gaussian pyramid on image to decompose it intodifferent scales After that each scale space is divided intoseparate circular blocks From each circular block featuresare extracted by Hu moment of length 4 Features are thensorted and matching is performed on it The performanceof this method is not changed in the presence of a rotationaltransformation

Li et al [17] applied discrete wavelet transform (DWT)and singular value decomposition (SVD) to the image FirstlyDWT is applied to each block that reduces the size of theblock Then SVD is applied on low frequency componentswhich reduce the feature dimensional size to 4 After thatfeatures are lexicographically sorted and matched In [18]the image is divided into overlapping patches Features areextracted from each patch by applying PCT PCT utilizesthe orthogonal properties so that features of PCT are morecompact It is a rotational invariantmethod and also performsbetter in the presence of noise

Bravo-Solorio and Nandi [19] mapped all the pixels ofoverlapping blocks into log polar coordinates Then angleorientation creates a 1D descriptor by doing this synchro-nization problem can be removed Feature vectors are cal-culated from particular blocks depending on the colour andluminance factor This method is robust towards reflectionrotation and scale changes

22 Keypoints Based Methods In this approach variouskeypoints are selected from an image Feature descriptorsare calculated from each of these keypoints For detectingthe duplication forgery matching is performed on keypointfeature descriptors Lowe [20] invented SIFT which detectskeypoints and features from an image SIFT application existsin various fields It performs better in comparison to previousdescriptors [21]

In [22] SIFT is applied into the region duplicationdetection method In [23] the authors suggested that thekeypoint matching method suffers from some problems Forremoving these problems SIFT cluster matching is proposedwhere objects are matched rather than the point Points aregrouped here using agglomerative hierarchical clustering

Pan and Lyu [24] also applied SIFT in their work Foravoiding search from close adjacency search is applied farfrom 11 times 11 pixel window whose center is at the keypoint




3 Proposed Method







1 1198732 1198733 119873

119872)


(3) For each 119894 = 1 to119872for 119895 = 1 to119872


ture descriptor

End of For 119895








else return


else return




Dist (119860 119861) = radic(1199092minus 1199091)2

+ (1199102minus 1199101)2

(1)

where 119860 = (1199091 1199101) 119861 = (119909

2 1199102)






Input image

Keypoint detection

Feature extraction

SURF

No



Yes


No



If allkeypoints are

processed


Yes




HAC


neighbourslt06


A B

C D







Σis calculated by


119868Σ(119909 119910) =

119894le119909

sum

119894=0

119895le119910

sum

119895=0

119868 (119909 119910) (2)



119867(119909 120590) = [119871119909119909

(119909 120590) 119871119909119910(119909 120590)

119871119909119910(119909 120590) 119871

119910119910(119909 120590)

] (3)




119909119910(119909 120590) and 119871



det (119867approx) = 119863119909119909119863119910119910

minus (09119863119909119910)2

(4)




119910as vertical



1003816100381610038161003816] (5)






119889 (119860 119861) = min (dist (119909119860119894 119909119861119895)) (6)


119889 (119860 119861) =1

119899119860119899119861

119899119860

sum

119894=1

119899119861

sum

119895=1

dist (119909119860119894 119909119861119895) (7)


119889 (119860 119861) = ESS (119860119861) minus [ESS (119860) + ESS (119861)] (8)

where

ESS (119860) =119899119860

sum

119894=1

10038161003816100381610038161003816119909119860119894

minus 119909119860

10038161003816100381610038161003816

2

(9)



cluster B 119909119860119894



119909119860=

1

119899119860

119899119860

sum

119894=1

119909119860119894 (10)





12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)



12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References








































3 Proposed Method







1 1198732 1198733 119873

119872)


(3) For each 119894 = 1 to119872for 119895 = 1 to119872


ture descriptor

End of For 119895








else return


else return




Dist (119860 119861) = radic(1199092minus 1199091)2

+ (1199102minus 1199101)2

(1)

where 119860 = (1199091 1199101) 119861 = (119909

2 1199102)






Input image

Keypoint detection

Feature extraction

SURF

No



Yes


No



If allkeypoints are

processed


Yes




HAC


neighbourslt06


A B

C D







Σis calculated by


119868Σ(119909 119910) =

119894le119909

sum

119894=0

119895le119910

sum

119895=0

119868 (119909 119910) (2)



119867(119909 120590) = [119871119909119909

(119909 120590) 119871119909119910(119909 120590)

119871119909119910(119909 120590) 119871

119910119910(119909 120590)

] (3)




119909119910(119909 120590) and 119871



det (119867approx) = 119863119909119909119863119910119910

minus (09119863119909119910)2

(4)




119910as vertical



1003816100381610038161003816] (5)






119889 (119860 119861) = min (dist (119909119860119894 119909119861119895)) (6)


119889 (119860 119861) =1

119899119860119899119861

119899119860

sum

119894=1

119899119861

sum

119895=1

dist (119909119860119894 119909119861119895) (7)


119889 (119860 119861) = ESS (119860119861) minus [ESS (119860) + ESS (119861)] (8)

where

ESS (119860) =119899119860

sum

119894=1

10038161003816100381610038161003816119909119860119894

minus 119909119860

10038161003816100381610038161003816

2

(9)



cluster B 119909119860119894



119909119860=

1

119899119860

119899119860

sum

119894=1

119909119860119894 (10)





12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)



12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References






































Input image

Keypoint detection

Feature extraction

SURF

No



Yes


No



If allkeypoints are

processed


Yes




HAC


neighbourslt06


A B

C D







Σis calculated by


119868Σ(119909 119910) =

119894le119909

sum

119894=0

119895le119910

sum

119895=0

119868 (119909 119910) (2)



119867(119909 120590) = [119871119909119909

(119909 120590) 119871119909119910(119909 120590)

119871119909119910(119909 120590) 119871

119910119910(119909 120590)

] (3)




119909119910(119909 120590) and 119871



det (119867approx) = 119863119909119909119863119910119910

minus (09119863119909119910)2

(4)




119910as vertical



1003816100381610038161003816] (5)






119889 (119860 119861) = min (dist (119909119860119894 119909119861119895)) (6)


119889 (119860 119861) =1

119899119860119899119861

119899119860

sum

119894=1

119899119861

sum

119895=1

dist (119909119860119894 119909119861119895) (7)


119889 (119860 119861) = ESS (119860119861) minus [ESS (119860) + ESS (119861)] (8)

where

ESS (119860) =119899119860

sum

119894=1

10038161003816100381610038161003816119909119860119894

minus 119909119860

10038161003816100381610038161003816

2

(9)



cluster B 119909119860119894



119909119860=

1

119899119860

119899119860

sum

119894=1

119909119860119894 (10)





12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)



12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References








































119909119910(119909 120590) and 119871



det (119867approx) = 119863119909119909119863119910119910

minus (09119863119909119910)2

(4)




119910as vertical



1003816100381610038161003816] (5)






119889 (119860 119861) = min (dist (119909119860119894 119909119861119895)) (6)


119889 (119860 119861) =1

119899119860119899119861

119899119860

sum

119894=1

119899119861

sum

119895=1

dist (119909119860119894 119909119861119895) (7)


119889 (119860 119861) = ESS (119860119861) minus [ESS (119860) + ESS (119861)] (8)

where

ESS (119860) =119899119860

sum

119894=1

10038161003816100381610038161003816119909119860119894

minus 119909119860

10038161003816100381610038161003816

2

(9)



cluster B 119909119860119894



119909119860=

1

119899119860

119899119860

sum

119894=1

119909119860119894 (10)





12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)



12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References






































12

(a)

12

(b)

12

(c)

12

(d)

12

(e)12

(f)

12

(g)

12

(h)

12

(i)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)


12

(a)

12

(b)

12

(c)

12

(d)



12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References






































12

(a)

12

(b)

12

(c)

12

(d)













0

50

100

150

200

250

300

Fridrichet al [11]


SURF andHAC

FPRTPRTime (s)

[12]





References





































Region Duplication Forgery Detection Technique Based on SURF … · 2017-11-15 · order derivative...

Documents

Transcript of Region Duplication Forgery Detection Technique Based on SURF … · 2017-11-15 · order derivative...