International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

Analysis of Digital Image Forgery Detection using Adaptive Over-

Segmentation Based on Feature Point Extraction and Matching

Ch. SUDARSHAN1, U. SATHISH KUMAR2

1(M.Tech.) Dept of CSE,Acharya Nagarjuna University, Guntur, AP, (522510), INDIA

2Assistant Professor (,M.Tech,(Ph.d)), Dept of CSE,Acharya Nagarjuna University, Guntur, AP, (522510), INDIA
--------------------------------------------------------------------------***------------------------------------------------------------------
Abstract--- The innovation of the web has presented the the privacy and the confidential information with greater
unimaginable development and improvements in the robustness. The forgery has become the major concerned
prestigious research fields, for example, pharmaceutical, area in the 21st and a lot of research is carried out in the
satellite symbolism, picture handling, security, biometrics, literature but still achieving the desired results remained
and genetics. The methods introduced in the 21st century as unsolved issue. The digital images are considered as the
has made the human life more comfortable and secure, primary source of the medium used for too meet the very
however the security to the original reports has a place with purpose which includes the data transmission, the data
the verified individual is stayed as worried in the digital compression, the data hiding and the various other
image processing area. In this paper, two methods are applicative research areas. The forgery of the images has
proposed to detect the forged region of an image such as reach to the new level to pose serious issues in the 21 st
Adaptive over-segmentation and feature point matching. century and it creates the situation where the difference
Thus Adaptive over-segmentation is used to extract the between the forged and non forged documents
features by both block-based and key point-based forgery identification become the biggest drawback, which is
detection methods. Firstly, the host image is segmented into addressed in efficient way using the proposed work.
non-overlapping and irregular blocks adaptively. Then these
block as compared to extract the feature points of a host Copy-move forgery, which is to paste one or several copied
image; this procedure can approximately indicate the region of an image into other part of the same image.
suspected forgery regions. To trace out the forged region During the copy and move operations, some image
efficiently, the forgery region extraction algorithm is used to processing methods such as rotation, scaling, blurring,
replaces the features point with the super pixels as feature compression, and noise addition are occasionally applied
blocks and then merges the neighboring blocks that have to make convincing forgeries. Earlier blocked based
similar local color features into the feature block to forgery detection was used to detect forged image but this
generate the merged region this merged region is processed algorithm faced some drawbacks such as the host image is
into morphological operation to detect the extracted forged divided into over-lapping rectangular blocks, which would
regions in an image. In cut-paste image forgery detection, be computationally expensive as the size of the image
image forensic techniques are used to detect the global and increases and it was less efficient as it take more time to be
local contrast enhancement, identifying the use of histogram process. To avoid such drawbacks along with the blocked
equalization. based forgery, we proposed an image-blocking method
called Adaptive Over-Segmentation that divided the host
Keywords: Copy-move, Forgery detection, the adaptive image into non overlapping blocks adaptively with the
over-segmentation, feature point matching, help of two algorithm those are Simple Linear Iterative
neighboring blocks, super pixels, feature points. Clustering (SLIC) to segment the host image into irregular
blocks and Discrete Wavelength Transform (DWT) which
is employed to analyze the frequencies of the super pixel.
1. INTRODUCTION Further the image block formed are pass to the Block
Feature Extraction method where the block feature are
extracted by using Scale Invariant Feature Transform
The digital image processing is the prominent research
(SIFT) as it possessed constant and better performance
domain in the 21st century where its presence is clearly
compared with the other extraction method. Further the
observed in various fields. The digital image processing is
process of Block Feature Matching is carried out which
a important constituent of the electromagnetic spectrum
used Simple Linear Iterative Clustering (SLIC) for
and the security field remain as one of the major research
calculating super pixel and Discrete Wavelength
areas on which lot of research needs to be done to secure
Transform for finding super pixel from one block and
2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1568
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

checking other for other blocks. When the features are properties such as noise, color and texture dont change
extracted and matched then we get to know which regions and make the recognition process troublesome.
the host image has been forged.
B. Image Forgery using Splicing
These methods are common in use for forgery detection,
but they are having following drawbacks: 1) there is very Image splicing uses cut-and-paste systems from one or a
high computational complexity as there is division of lot of pictures to make another pretend image. once
image into overlapped regions. 2) To deal with the conjunction is performed exactly, the borders between the
geometrical transformation of the forgery area is difficult. spliced regions will visually be unbearable. Splicing,
3) There is a low recall rate due to host image division in however, disturbs the high order Fourier statistics. These
regular blocks. insights will so be utilized as a locality of identifying
phony. Figure 2, demonstrates an honest sample of image
To address the shortcomings of the prevailing methods, we conjunction during which images the photographs the
tend to propose a unique copy-move forgery detection images of the shark and also the eggbeater ar unified into
scheme exploitation adaptive over-segmentation and have one picture.
purpose matching during this paper. The adaptive Over-
Segmentation algorithm is projected to adaptively divide
the host image into non-overlapping and irregular blocks.
Then the feature points are extracted from every block and
matched with every other to seek out the tagged feature
points which might approximately indicate the suspected
forgery regions. And finally the tagged feature points are
processed and also the morphological operation is applied
to get the detected forgery regions.
Fig 1. (a) original Image (i) ; (b) copy Image (ii);
2. TYPES OF DIGITAL IMAGE FORGERY (c) forged image

Now days fake images have become more in society. C. Image Resampling
Tampering images are common for making controversies.
For example, it can be used for sensational news, spread To make associate astounding fake image, some elite
political news and rumors. As the quality of pictures regions got to undergo geometric transformations like
suffers, it's necessary to plot techniques so as to verify rotation, scaling, stretching, skewing, flipping and then
their genuineness and trait of pictures. forth. The interpolation step plays a vital role within the
resampling method and introduces non-negligible applied
Picture sterilization is characterized as adding, changing, mathematics changes. Resampling introduces specific
or deleting some vital options from a picture while not periodic correlations into the image. These correlations
exploit any obvious trace. There are totally different are often utilized to acknowledge phony caused by
techniques used for formation a picture. Taking under resampling. In Figure three, the image on the left is that the
consideration the ways wont to create cast pictures, digital original image whereas the one on the proper is that the
image forgery are often isolated into 3 primary cast image obtained by rotation and scaling it.
classifications: Copy-Move forgery, Image splice, and
Image resampling.

A. Copy-Move Forgery
In copy move forgery, in original image some part of the
image with any size is copies and pasted in that only in
some area of image it can be show in figure1. As the copied Fig 2. (a) The real image (b) final result of image
part originated from the same image, its essential retouching

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1569
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

3. LITERATURE SURVEY 3.5 A Sift-based Forensic Method

3.1 Detecting Duplicated Image Irene Amerini et al. (2011) proposed a method to support
image forgery detection based on SIFT algorithm. Thus, the
A technique that works by 1st applying principal element algorithm is used to detect the regions which are
analysis to little mounted - size image blocks to yield a duplicated and determine the geometric transformation
reduced dimension illustration was planned by Alin C applied to perform such tampering. But, the main
Popescu et al. (2004). Whereas performing arts the on top drawbacks of this technique, it is unable to detect the
of technique we are able to realize some duplicate pictures image with uniform texture and salient keypoints.
(noises). Then the duplicate regions are detected by
lexicographically (the follow of aggregation 3.6 Exposing Transform-invariant Features
dictionaries).Sorting the whole image blocks. This can be
terribly wonderful and actual appropriate technique to Pravin Kakar et al. (2012) has proposed a method based
yield a reduced dimension illustration. Its sensitive to jpeg on transforming-invariant features. These got y utilizing
lossy compression and additionally it's additive to noise. the features from MPEG-7 image signature devices.This
method achieved good results, accuracy and extremely low
3.2 Fast Copy-Move Forgery Detection false positives. Thus, these features are invariant to
common image processing operations. This method cannot
A methodology to discover copy- move forgery by detect regions which have undergone affine
dividing the image into overlapping blocks of equal size, transformations and/or multiply copied.
extracting feature for every block and representing it as a
vector and typing all the extracted feature vectors 4. PROPOSED METHOD
victimization the base sort, was planned by Hwei-jen
sculpture et.al (2009). Base type dramatically reduces the For the forgery detection process lot of importance is
time complexness and also the adopted options enhance given from the past years. In this paper adaptive over
the aptitude of resisting of varied attacks like JPEG segmentation and feature point matching methods are
compression and mathematician noise. Each potency and used to detect the forgery region. The overall overview of
high detection rates are incontestable. this method is explained in the Fig.1.

3.3 Robust Copy-move Forgery Adaptive over-segmentation method is used to

segment the host image. The image is divided into
Sevinc Bayram et al.(2009) projected to use Fourier- non-overlapping and irregular blocks. The
Mellin Transform (FMT) options that square measure segmented blocks are called as image blocks (IB).
invariant to scaling and translation. A replacement The irregular block segmentation can be done by
detection scheme that creates use of investigation bloom the Scale Invariant Feature Transform (SIFT)
filters is additionally introduced by them. It detects copy technique, SIFT is applied to each block to get the
move forgery terribly accurately albeit the cast image is perfect block features (BF).
turned, scaled or extremely compressed. This detection
scheme improves the potency. However the hardiness of The main important parameter of the
the tactic is reduced. proposed method is detecting the suspected
forgery region, it can be detected by performing
3.4 Detection Digital Images Using SURF the matching between block features with one
another. The block features which are matched
B.L.Shivakumar et al. (2011)proposed a method to detect are named as Labeled Feature Points (LFP).
duplication regions. Because one of the common image The LFP will plays a major role in detecting
forgery methods is copy move forgery (CMF). the forgery because it is used as a reference. In
Identification of the CMF can be detected by the the final forgery region extraction
duplication regions using Speeded Up Robust Features method is proposed to detect the forgery
(SURF) keypoints. These SURF keypoints are extracted region from the host image.
from images. The duplication region can be detected with
different sizes. The result shows that CMF with minimum
false match for images with high resolution. A few small
copied regions were not successfully detected.
2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1570
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

the proposed method different superpixels sizes will

produce different forgery results.

To get the relationship between the frequency distribution

of host images and the initial size of the superpixels a large
number of experiments are performed. By using the haar
Input : Host image wavelet 4-level DWT technique is applied to the host
image then the image I divided into the low-frequency
energy ELF and high-frequency energy EHF can be
Adaptive Over- calculated using (1) and (2), respectively. The percentage
Segmentation of the low-frequency distribution in (3) can be
calculated by using the low-frequency ELF and high-
Output: Detected frequency energy EHF. The initial size of the superpixels
Forgery Regions can be defined in (4)
| |

Forgery (| | | | | |)
Region
Input Blocks (IB)

Block Feature
Extraction

{
Labeled Feature
Points (LFP)
Where S means the initial size of the superpixels; M N
indicates the size of the host image; and P LF means the
Adaptive Block percentage of the low-frequency distribution
Block Feature (BF) feature matching

Fig.3: The proposed copy-move forgery detection

scheme framework

A. Adaptive over segmentation algorithm

Input : Host image
When the size of the host image is increased the matching
computation of the overlapping blocks will become Discrete Ouput: Image
difficult. To overcome this problem we are implementing Wavelet Blocks (IB)
adaptive over segmentation. By using this method the Transform
image is divided into non-overlapping and irregular
regions called as image blocks. The use of non-overlapping Wavelet Coefficients
segmentation compared to the overlapping method is to
reducing the computational complexity and compared to Adaptive Block Superpixel
the regular blocks irregular blocks can give the better
Size Segmentation
performance like we can find the more forgery region. In
the process of this method finding the initial size of the Computation
Input block size: S
super pixels in SLIC method is very difficult. Basically in
the copy-move forgery detection the sizes and content will
be different when it is compared with the host images. In
Fig.4: The adaptive over-segmentation flowchart.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1571
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

The proposed Adaptive Over-Segmentation method can D. Forgery Region Extraction Algorithm
divide the host image into blocks with adaptive initial sizes After gathering the LFP (labeled feature points) then we
according to the given host images, with which each image have to locate the forgery region. The extracted LFP
can be determined to be an appropriate block initial size to regions are taken as the forgery regions. To get the more
enhance the forgery detection results. accuracy of the forgery region forgery region algorithm is
used. The LFP region is replaced with the small super
B. Block Feature Extraction Algorithm pixels to get the suspected regions (SR). The superpixels
are segmented from the host image. To get the more
In the adaptive over segmentation the host image is precision and recall rates the LFP of the super pixels that
divided into image blocks after that the block features are are neighbors of the suspected regions (SR) are also
extracted from the image blocks by using the block feature measured.
extraction algorithm. The image block content is mainly
affected by these feature extraction and these features are
non-resistance to the various image transformation
techniques. So in this method the feature points are
extracted from the each image block as block features and
these are not affected by image scaling, rotation, and JPEG
compression. The main feature point extraction methods Input : Label Feature
Output: Detected
which are used in this proposed method are SIFT and Points (LFP)
Forgery Regions
SURF. Compared to the previous techniques this method is
robust. When the experiment results are compared SIFT
will give better performance compared to the other feature Small super-pixels Morphology
extraction methods. So in this every block will contain the Replacement Processing
irregular block region information and the extracted SIFT
feature points.

C. Block Feature Matching Algorithm

In the existing system if there are so many other matching

pairs in one mutual position then only the block matching
process will give the specific block pair as a output and if Suspected Merged Regions
there are same shift vectors then the threshold value is Regions (SR) (MR)
calculated then the matched blocks are which specified to
the same shift vector are identified as the regions of copied
or moved. In the proposed algorithm we are considering Local Colour
the block features which consist of different feature points Feature
so a different method is used to locate these matched
blocks.
Algorithm: Block Feature Matching algorithm
Input: Block Features (BF);
Fig. 5: Flow chart of the Forgery Region Extraction
Output: Labeled Feature Points (LFP).
STEP-1:.After generating blocks load the features of algorithm
blocksBF={BF1,BF2,.BFN} where N is denoted as
number of image blocks. And find correlation between
the blocks. If the LFP is same as suspected regions then the super
STEP-2: By distribution correlation coefficient TRb is pixels which are neighbor are merged into the
corresponding suspected region. These results are
calculated for block matching threshold.
considered as the merged regions (MR). the final step is
STEP-3: With help of TRb threshold matched blocks
(MB)are located. detecting the copy-move forgery regions this step was
done by applying morphological operations on the merged
STEP-4: suspected forgery regions are indicated in
regions. The final flowchart of the forgery region
the matched blocks (MB) by labeling the matched
features points. extraction algorithm is shown in fig.4

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1572
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

Algorithm: Forgery Region Extraction Fig 6: original image

STEP-1: Load the Labeled Feature Points (LFP),
apply the SLIC calculation with the underlying size
S to the host picture to section it into small
superpixels as highlight squares, and supplant each
named include point with its relating highlight
piece, in this manner producing the Suspected
Regions (SR).
STEP-2: Measure the local color feature of the
superpixels neighbor to the SR, called neighbor
blocks; when their color feature is similar to that of
the suspected regions, we merge the neighbor
blocks into the corresponding SR, therefore
creating the merged regions (MR).
STEP-3: Apply the morphological close operation
into MR to finally generate the detected forgery
regions.

5. EXPERIMENTAL RESULTS

1) JPEG compression: the JPEG compressed images are

the forgery images. The compression can be with a quality Fig 7: forged image
factor varying from 100 to 20,in steps of -10. So here we
have to test the total of 48 9=432 images.
2) Rotation: the regions which are copied are rotated by
the rotated angle varying from 2 to 10, in steps of
2,and the rotation angles are about 20, 60 and 180 as
well. So here we have to test the total of 48 8=384 images.
3) Scaling or Noise: The regions which are copied are
scaled by using the scale factor varying from 91% to
109% in steps of 2%, and the scale factor is about 50%,
80%, 120%, and 200%. As well. So here we have to test
the total of 48 14=672 images.
4) Median filter: Total 48 forged host images are present
in the dataset. These images are scaled down from 90% to
10% in steps of 20%. So here we have to test the total of
48 5=240 images.
In this section, a series of experiments square measure
conducted to evaluate the effectiveness and lustiness of
the planned image forgery detection theme exploitation
adaptive over-segmentation and have purpose matching.
Within the following experiments, the image dataset in
[22] is employed to check the proposed methodology. This
dataset is made supported forty eight high-resolution
uncompressed PNG true color pictures, and the average
size of the pictures is 1500*1500. within the dataset, the
copied regions square measure from the classes of living,
nature, man-made and mixed, and that they vary from to a
fault swish to highly textured; the copy-move forgeries
square measure created by copying, scaling and rotating
semantically substantive image regions.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1573
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

Fig 8: Adaptive over segmentation Fig 12: Analysis of F1 Measure (a) JPEG Compression,
(b) Rotation, (c) Noise Scale and (d) Median filter.

(a) F1 measure for JPEG compression effect

Fig 9: label feature points

(b) F1 measure for rotation effect

Fig 10: Morphed detected forgery region

(c)F1 measure for noise effect

Fig 11: final detected forged region.

(d) F1 measure for median filter effect

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1574
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

Fig 13: Analysis of Precision and Recall at the pixel

level (a) JPEG Compression, (b) Rotation, (c) Noise
Scale and (d) Median filter.

(b) Rotation effect

(a)JPEG compression effect

(d)Median filter effect.

(c) Noise effect.

Fig 8: Processing time SIFT and SURF

6. CONCLUSION

This work proposes for Image forgery detection using To detect the forgery regions more accurately, and the
adaptive over segmentation and feature point matching. In forgery region extraction algorithm, which replaces the
forgery detection method proposes block based and key feature points with small super pixels as feature blocks
points integrates scheme, first the proposed adaptive over and then merges the neigh boring blocks that have similar
segmentation algorithm segments the host image into non local color features into the feature blocks to generate the
overlapping and irregular blocks adaptively. Then, the merged regions.
feature points are extracted from each block as block
features, and the block features are matched with one
another to locate the labeled feature points; this procedure
can approximately indicate the suspected forgery regions.
2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1575
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072

REFERENCES Processing, 2006. Icassp 2006 Proceedings. 2006 Ieee

International Conference On, 2006, Pp. Ii-Ii.
[1] Q.-C. Yang And C.-L. Huang, "Copymove Forgery [11] M.-J. Tsai And C.-S. Wang, "Adaptive Feature Selection
Detection In Digital Image," In Advances In Multimedia For Digital Camera Source Identification," In Circuits And
Information Processing-Pcm 2009, Ed: Springer, 2009, Pp. Systems, 2008. Iscas 2008. Ieee International Symposium
816-825. On, 2008, Pp. 412-415.
[2] B. Mahdian And S. Saic, "Blind Methods For Detecting [12] Y. Sutcu, S. Bayram, H. T. Sencar, And N. Memon,
Image Fakery," Ieee Aerospace And Electronic Systems "Improvements On Sensor Noise Based Source Camera
Magazine, Vol. 25, Pp. 18-24, 2010. Identification," In Multimedia And Expo, 2007 Ieee
[3] B. Shivakumar And L. D. S. Santhosh Baboo, "Detecting International Conference On, 2007, Pp. 24-27. [13
Copy-Move Forgery In Digital Images: A Survey And [13] T. Chen, J. Wang, And Y. Zhou, "Combined Digital
Analysis Of Current Methods," Global Journal Of Computer Signature And Digital Watermark Scheme For Image
Science And Technology, Vol. 10, 2010. Authentication," In Info-Tech And Info- yesrNet, 2001.
[4] K. N. Qureshi And A. H. Abdullah, "A Survey On Proceedings. Icii 2001-Beijing. 2001 International
Intelligent Transportation Systems," Middle East Journal Conferences On, 2001, Pp. 78-82.
Of Scientific Research, Vol. 15, 2013. BIOGRAPHIES
[5] W. Lu, W. Sun, J.-W. Huang, And H.-T. Lu, "Digital Image
Forensics Using Statistical Features And Neural Network Ch. SUDARSHAN received the
Classifier," In Machine Learning And Cybernetics, 2008 Bachelor of Technology degree in
International Conference On, 2008, Pp. 2831-2834. Computer Science engineering from
Rajive Gandhi university of
[6] S. Bayram, B. Sankur, N. Memon, And . Avciba, "Image Knowledge and Technology in 2015.
Manipulation Detection," Journal Of Electronic Imaging, Currently pursuing M.tech in
Vol. 15, Pp. 041102-041102-17, 2006. Computer Science Engineering from
[7] A. C. Popescu And H. Farid, "Exposing Digital Forgeries Acharya Nagarjuna University. His
By Detecting Traces Of Resampling," Signal Processing, areas of interest include in computer science engineering
Ieee Transactions On, Vol. 53, Pp. 758-767, 2005. specialized in image processing and Data mining.

[8] A. E. Dirik, S. Bayram, H. T. Sencar, And N. Memon, Mr. U.SATHISH KUMAR is currently
"New Features To Identify Computer Generated Images," working as assistant professor in
In Image Processing, 2007. Icip 2007. Ieee International Acharya Nagarjuna University from
Conference On, 2007, Pp. Iv-433-Iv-436. last 3 years. Currently he is pursuing
[9] M. Kharrazi, H. T. Sencar, And N. Memon, "Blind Source his Ph.D degree. His research interest
Camera Identification," In Image Processing, 2004. Icip'04. includes in Computer Science and
2004 International Conference On, 2004, Pp. 709-712. Engineering in Data mining, image
processing and Soft computing.
[10] M.-J. Tsai And G.-H. Wu, "Using Image Features To
Identify Camera Sources," In Acoustics, Speech And Signal

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1576