Professional Documents
Culture Documents
ON
A STUDY ON COPY-MOVE FORGERY DETECTION
TECHNIQUES IN IMAGES
SUBMITTED BY
AMARTYA RANA - 2016UIT2586 ; JATIN NAGPAL - 2016UIT2567
In present day, digital media has become an important part of our lives. Also accessing and sharing
information is very easy with digital media nowadays, thus it plays a significant role in a variety of fields
and various technologies. Similarly, various image processing softwares and tools are available in the
market to enhance or edit these images to look as required. As Softwares for image editing are easily
available nowadays, image forgery has become a very common thing. And from one of the many forgeries,
Copy-Move forgery is a forgery technique in which a region of image is copied and pasted in nearly
matching area in order to hide a particular part of that image. This might be done in order to hide or add
certain feature or object in the image which changes the integrity of that image.
It is very crucial to identify Forged images from the original ones as one of the most important fields
where digital images are used as Forensic studies and Law enforcement where images are considered
as authentic proofs or have a purpose.
Image Retouching: This forgery technique is less harmful than other techniques since the final
forged image is not very different from the original one but with just minor enhancements, used to make
the object look more attractive. Thus retouching is just like polishing of an image in which a retoucher
focuses on adjusting various other elements of an image .
Copy-Move Attack: This forgery technique is a difficult and a commonly used forgery technique. The
image is forged by extracting a specific region of the original image and using it again in nearly matching
area in order to hide a particular part of that image or add extra feature to it. Thus changes the information
conveyed by the image completely.
Active detection which requires prior information about the original image. Most of the techniques in this
type work when images are already pre-processed such as with watermarking or digital signature embedding
which limits their application.
Passive detection which doesn’t require any pre-processed image information to detect whether it’s
forged or not. It is based on assumptions that even though we cannot spot the tampered image from the
original image with naked eyes but it would likely to alter some other statistics, which then are used to identify
the real image.
In this paper, we focused on “Copy-Move forgery” (CMF) only and did a Study on different CMF detection
techniques worked upon from 2013 – 2019. Here, we answered few RQ’s by comparing different techniques
that have been used to detect CMF, state the benefits we are getting out of those techniques and find any
limitation that can’t be solved with all the available techniques and are still needed to be solved.
Here we also focused on one of the most important technique named : SIFT (Scale Invariant
Feature Transform) which was used and researched upon in various papers in last 7 years and
researched about it in depth in order to implement it and find areas where more development
could be made in this particular algorithm. This study would also then help to find and fill up gaps (if
any) in the Copy-Move forgery detection field, which then would help us to give suggestions on the areas
where further investigation can be done. In this study we also try to find information about the effect of these
researched techniques in real life and how it really help us. Thus, with this paper we can give the readers an
idea about the development in the area of “Copy-Move forgery Detection techniques” and help them
understand the topic to the best of our knowledge.
2. MOTIVATION
Nowadays, digital images are used in approximate every sector, in one way or the other. For example, images are
used in Forensic and Medical Studies, act as an evidence in court cases, used to share and store information, used
in Artificial intelligence, etc. But with the available technologies it's very easy to manipulate these images and
create forged images which then destroy the whole purpose of using digital images. Thus, it’s very important to
know the integrity of images before using it.
“Copy-Move forgery” is one of the most used forgery done on images when compared with other forgeries as a
greater number of tools are available to do it and it can totally invert the information which the images were
supposed to provide us with. In “Copy-Move forgery” a particular section of an image is copied and pasted at
some other place in the same image in order to add or hide an object presented in the image.
In this paper we studied the techniques developed to identify “Copy-Move forgery” in images in the last few
years and answered various research questions and in addition compared all those techniques on various factors
and try to provide the readers with a clearer understanding about each technique and their advantages and
disadvantages all together. This paper also helps one to see the development done in this sector in the past years
as we covered all the techniques available to us to the best of our knowledge.
While reading about CMFD techniques one needed to visit a lot of sites and read various papers to get basic
information on the topic and thus we aim to provide brief information on everything available in this single paper
in order to save time and efforts by answering basic questions that arises while reading about the topic in genera
3. METHODOLOGY
In this paper we studied various CMFD techniques with the help of research papers written in the last few years
and to complete the paper we followed the steps showed in following figure:
1. Since Copy-Move forgery is a very crucial problem nowadays as most of the things are turning digital now
and thus makes its detection a very important topic to be researched upon. And for that a thorough research
was done to see what we need to do to in this paper to make our contribution towards the field and this was
achieved by reading various research papers and articles and then providing people with better understanding
on the topic and new suggestions or improvements in the field of forgery detection.
2. After getting the motivation to work on the topic, we decided upon the research question (RQ’s) which will
be answered in the paper for the readers and for that all the doubts and questions that rose in our minds while
studying about the topic were noted down and then cumulated. The most important and informative RQ’s
were selected to be researched upon and later answered in the paper
3. Researching for RQ’s was a difficult task as we downloaded and read all the papers available on the topic
“ Copy-Move forgery detection techniques” on various sites like: IEEE Xplore, ScienceDirect, Google
scholar, etc. from which we gathered data that we required for each of our RQ.
4. There were various papers with similar information in them so we selected only those research papers which
provided us with new information or used a different technique to identify CMF and rest all papers were
excluded for researching.
5. After gathering data for each RQ everything is drafted correctly in a way it explains the respective topic in
an easy language so that a beginner can also understand about it.
What are the techniques that can be Tells which all different techniques
RQ3 implemented if the image is rotated? available can be applied on images
which are rotated and forgered later.
1. Which data sets are used for copy move forgery detection?
2. What are the performance measures used in different studies?
3. In what are the publications rates were high.
4. What are techniques that can be implemented if the image is rotated?
5. Compare SIFT and SURF technique.
6. Run a performance analysis on key points and block based tech
7. What are the techniques that are most us
5. EXPERIMENTAL RESULTS
In this section we represent the data we obtained for our RQ’s and answer each of them.
5.1 What different data sets are used for “Copy-Move Forgery detection”?
N
DATASET NAME IMAGE SIZE TYPES OF ATTACKS
O.
GRIP CMF13,
5.2 What are the different performance measures used in different papers?
CMF1 , CMF6 ,
Recall CMF10,CMF12,CMF15,CMF19,CMF20,CMF21,CMF22
Precision CMF8,CMF15,CMF19,CMF20,CMF22,CMF23
Sensitivity CMF4,CMF11
5.3 What are the techniques that can be implemented if the image is rotated?
A. Speeded-Up Robust Features extraction (SURF)
SURF initially was presented by “Bay et al” [18]. It is used to identify images with rotation and
scale invariant forgeries and thus had respective detectors and a descriptor. SURF is considered to
be better method in terms of robustness, distinctiveness and repeatability. SURF is popular because
it has faster computation time as compared to other schemes. It works by detection of interest points
from the image along with interest point descriptors. [ACMF1]
In this section, we are showing a new algorithm which is a combination of SURF with DWT. The
algorithm is as follows:
• The decomposed image is divided into four sub-parts: HL, LL, HH, LH.
• The LL part of the image has most of the image information in it.
• Then SURF is then applied on the LL part in order to obtain its descriptors and key points.
• Finally we find matches between descriptors. It can be obtained by using the method named as
Best Bin First Method.
• In this method single linkage clustering is used, along with ward linkage and average linkage in
order to get the forged area.
There are many positive results obtained using SURF with DWT when compared to simple SURF
technique alone. Thus we decided to replace DWT with DyWT . Thus a new algorithm is proposed
in this section where SURF will be combined with DyWT. The process is somehow similar to the
above algorithm except the fact that DyWT is used instead of DWT. Based on the literature
reviews:
• Then SURF is applied, which is used to extract corresponding descriptors and key points.
• Finally we find matches between these descriptor vectors in order to identify the duplicate region.
• The working of the process is same as compared to the process where SURF was combined with
DWT. The only difference is that here we are using DyWT in place of DWT and rest remains the
same.
B. BRISK Features
This feature detects, then describes all the key points. After description then it matches various key
points without complete information on camera dimensions and scene. In case of SURF technique
and SIFT technique (methods which are known and established) algorithms, the BRISK feature
technique proves to be faster in execution of matching patterns. This feature is dependant on
modifiable sampling pattern for high performance. It is used for less computational power
applications or various real-time constraints [6].
It uses interest point detectors [12][13] which is used in computer vision security applications and
the image forgery detection system[14]. These features can withstand different transformations
and changes in an image.
.
D. KAZE Features
It is a 2D feature description and detection method. For identification of feature points, Hessian
matrix is usually used for locating those feature points. It operates on a non linear space. If the
point which is detected has a value greater than others then it is taken as a key point. Then vector
description is done and in this step we require proper orientation. Proper orientation is done by
local structuring of feature points. Then the final key points are computed in order to obtain a
rotation and scale invariant descriptor. In the end forgery is detected in the image [10][11].
E. SIFT Features
The full form of SIFT is “Scale Invariant Feature Transform”. It works as an image descriptor and
helps in image-based matching and image recognition. And was developed by David Lowe
(1999, 2004). SIFT and similar image descriptors are used for various other purposes in computer
vision when it comes to the view-based object recognition and point matching between different
views in a 3-D scene. And the SIFT descriptor remains same even if rotations, translations and
scaling transformations are applied on an image. The SIFT descriptor is claimed to be very useful
for and object recognition and image matching under real-world conditions.
• The publications were surveyed on CMFD from the year 2015 to 2019.
• This paper has addressed the numerous forgery techniques, limitations, important results
and tells the direction in which further development can be done in future.
• It focuses on image’s, copy-move forgery datasets. It shows the comparison in the
performance of different datasets in different papers.
• Some of the issues are being identified in CMFD for further research.
Publications
57
43
41
32
30
20
9
5 6 5
Journals Conference
Feature matching and detection are used in object retrieval, object tracking and image registration.
There are various approaches which are used to match and detect features such as SURF
(Speeded up Robust Feature), SIFT (Scale Invariant Feature Transform), ORB and FAST. SURF
and SIFT are most useful approaches to match and detect features. In this paper, the comparison
between SURF and SIFT are discussed. The following points demonstrate the comparison:
• SURF is better in blur, rotation invariant and warp transform as compared to SIFT.
• SIFT is better in different scale images as compared to SURF.
• SURF is 3 times faster than SIFT because of the usage of box filter and integral image.
• SIFT and SURF both are good in illuminating changes in the images.
Comparing SIFT and SURF on the basis of time taken by both techniques in identifying that the
images considered are similar. Following are the images:
d
. Blur image e. Hue image f. Saturated image
g, Wrapped image h. Noise image
SIFT and SURF can also be compared on various other factors as shown in table below: -
Algorithm Rotation Illumination Scale Blur Warp RGB Noise Time Cost
SURF [31] key point features and SIFT introduced by Lowe [28] are used in various CMFD. The following
steps are followed for SIFT computation [31]:
• Key point descriptors.
• Orientation Assignment
• Key point Localisation.
• Scale-space extrema detection
Matching speed is faster in SURF technique when compared to SIFT due to integral image and Hessian
matrix’s approximation. An image is formulated by the addition of the values between the origin and points.
Key points from SIFT are derived and used for CMFD[17].Various images were collected from
various sources (majorly internet) for this experiment. SIFT descriptors are first derived and then
Euclidean distance is calculated between the descriptor vectors for matching extracted SIFT
descriptors. It is then observed that due to properties of SIFT (of rotation and scale invariant), the
method is also robust for forgeries in which scaling and rotation was done . Although it still isn’t
robust for detecting forgeries against small-size smooth regions and low signal-to-noise ratio and
needs to be improved.
.
In [48], “SURF” we first extract descriptors from the duplicate or tampered image. Then the
descriptor’s subsets were matched. It was noticed that this method can prove to be fast and reliable in
case of images that are small in size. Although localization of forgery in the image is not done.
In [34], SURF can also be combined with Discrete Cosine Transform (DCT) for “CMFD”. This
method is then applied on “Uncompressed color images database” (UCID). Co-efficients of DCT are
then checked for double JPEG compression effect. The extraction of features is done in the same
manner by using SURF and it was noticed that this method also is able to locate and detect the
forgeries in the tampered image. Although, the experiment was performed on a particular dataset
containing very few images only.
In [58], SURF was combined with Fourier–Mellin Transform (FMT) for CMFD. In this method
image is divided into flat regions and non-flat regions. Then, afterwards, in flat regions FMT is
applied and in non-flat regions SURF is applied to find duplicate regions. It was then observed that
the method cannot be used for images that are large in size as time taken by FMT for processing is
greater when compared to other techniques.
2. Block-based “CMFD” techniques
A coefficient based quantized DCT process is shown in [46]. In this report, forged regions
are identified by sorting the DCT coefficients lexicographically and by using shift vectors.
Due to the “DCT coefficient matching”, this process becomes robust against retouching,
compression and noise. However, for images that are tempered/forged by Rotating or
Scaling the copied blocks, aren’t detected by this method.
FMT for forgery detection is stated and used in [10]. Writers of the report claim that this
algorithm (based on FMT) can easily identify forgery in the tempered images even if the
tempered image is highly compressed, scaled or rotated and it also takes less time in
detection since instead of lexicographic sorting, counting bloom filters are used.
Paper [45] used two robust features based on kernel PCA and DWT. Various natural
images were used in order to test this method. From the quantitative analysis it was
observed that in the uncompressed domain and noiseless domain, features based on PCA/
kernel PCA (KPCA) are outperformed by DWT in terms of recall and average precision.
Features based on KPCA, perform well in JPEG compression and noise.
In [15], extraction of singular valued features is done with the help of SVD by applying it
on each overlapping block. Also, K-d tree approach is used for feature matching.
However, this method isn’t suitable for images with JPEG compression because it results
in loss of image details in SVD.
In [48], geometric and algebraic invariant feature vectors are also obtained by using SVD.
For identifying similarities between each block pair, “Chebyshev” distance measure is
used. It was then concluded, that this method proved to be robust against Gaussian noise
(AWGN), additive white, blur and JPEG compression. However, rotation and scaling
operations were not detected by using this method.
In [50], for dimensional reduction DWT was applied in the image. And then, SVD is
applied to the fixed-sized overlapping blocks of a low-frequency wavelet
In [45],from each block Texture features are found and matched . In terms of precision
versus execution time it is noticed that the best way is using statistical texture descriptor.
Also, in the case of images with forgery in small regions, edge histogram performed
better.
In [15], in order to produce the Gabor feature descriptor of each block, a Gabor filter with
different rotation angles, scaling factors and frequencies is used. However, tempered
images in which forged regions have undergone different geometric transformations,
resulted in under-performance of this method.
Moment invariant methods are set of features that are invariant to a geometric
transformation. A blur moment invariant is used in [51]. In the method used in the report,
the image is first separated into overlapping blocks and then 24 blur invariants are used to
represent each block up to seventh order, for creating feature vectors of 72 dimension. To
reduce the dimension of the block features, Principal component transformation (PCT) is
introduced in this process. Finally, with k-d tree, matching among the blocks is
performed. This method is robust against JPEG compression, noise and presence of blur
in images that are smaller in size. This is because of presence of PCT and blur moments.
However, due to large number of block features in images which are larger in size,
computational complexity increase when we use this method.
Due to sensitivity to various transformations “block-based CMFD techniques” hold a major drawback.
Also, due to the separation of the images into various non-overlapping and overlapping blocks, high
computational time is taken by the techniques. Generally, there is participation of each and every block in
extraction of features and phase matching. Size of the image affect these techniques ability to
detect/identify forgeries significantly. The size of the block and offset chosen for dividing an image also
brings the same problem.
When block-based techniques are compared to the key point- based techniques, it is observed that
performances of key point- based techniques are robust against various transformations and prove to be
better in computational complexity too.
In “key point-based technique” features size is relatively larger than those in “block-based techniques”,
but the key point derived from the features vector are generally small in magnitude when compared with a
block of the image. However, detection of forgeries in highly uniform areas are not possible with methods
based on “key point techniques”.
After going through all papers of 2015-2019, SIFT technique was the most applied technique
Either working as the only algorithm performing the copy move forgery detection or combining with other
algorithms like DWT or HSV etc.
Algorithm:
In this step points which can be identified from different views are located. The most efficient
way is by using a Gaussian space-scale function. The scale space is defined by (3).1 (x,y) is the
input image and G(x,y, 0") is the Gaussian function. L(x,y,O") = G(x,y, 0") * l(x,y) (3) Key-
points in the image are found by the Difference of Gaussian (DoG) technique. DoG is given by
(4). D(x,y, 0") = L(x,y, kO") - L(x,y, 0") (4) This function is compared with its eight neighbours
on the same scale and with nine neighbours up one scale and down one scale. It is a point of
extrema if it is minimum or maximum among all these values.
2) Key-point Localization
This stage eliminates those key-points which have poor contrast. This is done by applying the
Laplacian operator on the points found in the previous stage. If the value of this function lies
below a certain threshold then this value is excluded. This eliminates points of low contrast. There
is a large principle curvature across the edge but a small curvature in a perpendicular direction in
DoG function [2, 18]. The 2 x 2 Hessian matrix at the location is calculated and the ratio of the
largest to smallest Eigen vector obtained is calculated. If the difference be less than this ratio, that
particular key-point is eliminated.
3) Assignment of orientation
In this step every key-point is assigned an orientation based on the local image properties. The
steps are as follows:
a. Calculate the Gaussian smoothed image L by the scale function described above.
b. Calculate the magnitude gradient m and the orientation 8 as in (5) and (6).
m(x,y) = (L(x + I,y) -L(x _1,y))2 + (L(x,y + I) -L(x,y _1))2 (5) _ I L(X, y+ I)-L(x, y-I) 8(x, y) =
tan ) L(x+l,y)-L(x-I,y) (6)
4) Key-point Descriptor
Key-point descriptors are also created by using the local gradient data used above. The gradient
information is rotated so as to align it with the orientation of the key point and then weighted by a
Gaussian function of variance 1.5*key-point scale [13]. This is used to create a set of histograms
over a square whose center is the key point. Usually a set of 16 histograms each with 8 elements
is used by the key point descriptors. This results in the feature vectors having 128 elements. These
feature vectors are then used to didentify possible objects in the digital image.
Advantages:
1. SIFT descriptor is a classic approach, also the "original" inspiration for most of the
descriptors proposed later
Disadvantages:
1. The drawback is that it is mathematically complicated and computationally heavy.
2. SIFT is based on the Histogram of Gradients. That is, the gradients of each Pixel in the patch
need to be computed and these computations cost time.
1. It is found that only few papers have worked on the forgery having multiple forged regions and in
real, life a forger will perform multiple forgeries in a single photo which in result will fail most of
the techniques that are developed. The HSV Histogram based [cmf136] approach extracted features
successfully detects multiple duplicated region in the image but have some limitations in it which
can be worked upon. Applying SIFT only on region of interests rather than on whole image (which
will be time consuming). Reducing the computation time both for extraction the key-points as well
as to match them which is negligibly small. (cmf134)
2. SIFT and SURF are iterative methods which are itself very time consuming Hashing techniques can
be applied to get the result in less time. Like in [74] Coherency sensitive hashing is used for
identifying and matching image patches. This method is faster than other iterative methods but if
the derived level of error is still not stable enough, the entire process is implemented again. So the
researchers’ community should put more emphasis on this type of methods.
3. Soft computing techniques in CMFD are used in very few papers and report. Although, SIFT with
“Particle Swarm optimization” (PSO) is presented in [49]. Parameter values used for CMFD, are
generated automatically in PSO techniques. Determining the values of the decision parameters
helps in the decision-making phase of the algorithm and that is generally based on results of
experiments or the experience against number of tempered/forged images present. Therefore, Soft
computing techniques in CMFD can be used to enhance the decision making phase of the
algorithm.
6. REFERENCES
[CMF1] Badal Soni, Pradip K. Das, Dalton Meitei Thounaojam ; CMFD: a detailed review of block
based and key feature based techniques in image copy-move forgery detection ; 29th October 2017.
[CMF2] Dijana Tralic, Ivan Zupancic, Sonja Grgic, Mislav Grgic ; CoMoFoD - New Database
for Copy-Move Forgery Detection ; University of Zagreb, Faculty of Electrical
Engineering and Computing
Unska 3, HR-10000 Zagreb, Croatia.
[CMF3] Amanpreet Kaur, Savita Walia, Krishan Kumar ; Comparitive analysis of different
keypoint based copy- move forgery detection methods ; 2018 IEEE
[CMF4] Mariam Saleem, M,Qasim Altaf-L, Qaiser Chaudry ; A Comparative Analysis on
Pixel-Based Blind Cloning Techniques ; 2014 IEEE International Conference
[CMF5] Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Stefan Winkler, Rama Ratnam ;
COMPARISON OF HUMAN AND MACHINE PERFORMANCE FOR COPY-MOVE IMAGE
FORGERY
DETECTION INVOLVING SIMILAR BUT GENUINE OBJECTS ; ©2018 IEEE
[CMF6] Mohammad Farukh Hashmi, Aaditya R. Hambarde, Avinash G. Keskar ; Copy Move
Forgery Detection using DWT and SIFT Features ; ©2013 IEEE
[CMF7] Ava Pourkashani, Asadollah Shahbahrami, Babak Abad Fomani ; Copy Move Forgery
Detection using Histogram Quantization of Cross Power Spectrum ; © 2017 IEEE
[CMF8] Ashwini V. Malviya, Siddharth A. Ladhake ; Copy Move forgery detection using low
complexity feature extraction. ; 2015 IEEE UP Section Conference
[CMF9] Neetu Yadav, Rupal Kapdi ; Copy Move Forgery Detection using SIFT and GMM ;
2015 5th Nirma University International Conference.
[CMF10] Aniket Roy, Akhil Konda, Rajat Subhra Chakraborty ; COPY MOVE FORGERY
DETECTION WITH SIMILAR BUT GENUINE OBJECTS ; © 2017 IEEE
[CMF11] Ghulam Muhammad, Muneer H. Al-Hammadi, Muhammad Hussain, Anwar M. Mirza
and George Bebis ; Copy Move Image Forgery Detection Method Using Steerable Pyramid
Transform and Texture Descriptor ; ©2013 IEEE
[CMF12] Diaa M. Uliyan, Hamid A. Jalab, Ainuddin W. Abdul Wahab ; Copy Move Image
Forgery Detection Using Hessian and Center Symmetric Local Binary Pattern ; 2015 IEEE
Conference on Open Systems. [CMF13] Davide Cozzolino, Giovanni Poggi, Luisa Verdoliva ;
COPY-MOVE FORGERY DETECTION BASED ON PATCHMATCH ; ©2014 IEEE
[CMF14] Prajwal Pralhad Panzade, Choudhary Shyam Prakash, Sushila Maheshkar ; Copy-Move
Forgery Detection by Using HSV Preprocessing and Keypoint Extraction ;
[CMF15] Atefeh Shahroudnejad, Mohammad Rahmati, Copy-Move Forgery Detection in Digital
Images Using Affine-SIFT ; ©2016 IEEE
[CMF16] Copy-move Forgery Detection in the Presence of Similar but Genuine Objects
[CMF17] Comparative Study of Copy Move Forgery Detection Techniques
[CMF18] Copy-Move Forgery Detection based on SVD in Digital Image