This work is licensed under a Creative Commons Attribution 4.0 International License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
Com puter Modeling in
Engineering & Sciences ech
A Thorough Investigation on Image Forgery Detection
Anjani Kumar Rai*and Subodh Srivastava
Department of Electronics & Communication Engineering, NIT, Patna, India
*Corresponding Author: Anjani Kumar Rai. Email: firstname.lastname@example.org
Received: 19 December 2021 Accepted: 11 May 2022
Image forging is the alteration of a digital image to conceal some of the necessary or helpful information. It cannot be
easy to distinguish the modied region from the original image in some circumstances. The demand for authenticity
and the integrity of the image drive the detection of a fabricated image. There have been cases of ownership
infringements or fraudulent actions by counterfeiting multimedia les, including re-sampling or copy-moving.
This work presents a high-level view of the forensics of digital images and their possible detection approaches. This
work presents a thorough analysis of digital image forgery detection techniques with their steps and eectiveness.
nd the best forgery detection technique based on the dierent environments. It also shows the current issues in
other methods, which can help researchers nd future scope for further research in this eld.
Forgery detection; digital forgery; image forgery localization; image segmentation; image forensics; multimedia
Millions of digital documents are created and circulated daily through newspapers, magazines,
websites, and television. Images are an excellent tool for communication in any of these information
channels. These images, along with video and audio, can be easily collected using various devices
or software. DIs serve as evidence or proof against crimes. Unfortunately, manipulating images
with computer graphics and image processing tools is not difficult. The way we deal with photo
modification raises many legal and ethical issues that need to be addressed . However, before
considering what action to take in response to a problematic image, one must first determine it has
been altered. There are various ways to modify the content of an image, such as compression, splicing,
copy-move, and retouching techniques [2–4]. One of the most typical picture alteration processes is
image composition (or splicing) and retouching. Various applications can perform manipulations
in images that humans cannot detect by just looking at them once. Therefore, there is a need for
automated digital image forgery detection (DIFDs) techniques to perform these tasks very fast and
effective. Image forgery detection has developed as a fantastic study in various DIPs applications
(digital image processing), image forensics, criminal investigation, biomedical technology, computer
1490 CMES, 2023, vol.134, no.3
vision [5–7]. There are various digital image forgery detection techniques used to detect the different
kinds of forgery. DIFDs can be broadly classified passively or actively , as depicted in Fig. 1.An
active forgery detection technique requires pre-extracted or pre-embedded information.
Figure 1: Categorization of image forgery detection techniques
1.1 Categorization of DIFDs
Active DIFDs necessitate the preparation of digital images, including watermarking, embedding
and signature creations, limiting their practicality in applications. Passive DIFDs, unlike watermarks
 and signatures , do not generate digital signatures or embedding from watermarks. DIFDs can
be classified into five groups , as shown in Fig. 2. A few select DIFDs are presented for detecting
passive image forgeries .
Figure 2: Digital image forgery detection techniques
Pixel-based DIFDs: Pixel-based approaches process DI pixels to statistically identify abnormalities
that arise in image pixels due to tampering. The approaches also consider spatial or altered domain
correlations between pixels that happen in the tampering of images. Copy-move, re-samples, re-
touches, and image splicing are some of the approaches under this category. These are the most
frequently used methods in DIFDs .
Format-based DIFDs: Forged images can also be detected based on their formats. They are
used mainly in JPEG formats. Identifying fraud in compressed images is a complex task. However,
structure-based approaches detect forgery even in compressed images. A modified, fabricated image
that has been compressed (JPEG) makes forgery detections extremely difficult. Forensics investiga-
tions, however, use many characteristics of JPEG compression to discover manipulations. Format
based DIFDs can be found on one of JPEG quantization , double JPEG compressions [15,16],
multiple JPEG compressions [2,17], and JPEG blocks.
CMES, 2023, vol.134, no.3 1491
Camera-based DIFDs: A digital camera transfers acquired images from sensors to memory. They
are quantized, color correlated, gamma-corrected, white balanced, filtered, and JPEG compressed.
These processes are based on the different camera models and artefacts. Image captures undergo
multiple stages of processing. Light enters the camera’s lens, followed by the sensors’traversals through
CFAs (Color Filter Arrays). The sensor’s Photodetectors capture this incidental light and convert it to
a voltage form, converted into digital data using A/D (Analog-to-Digital) conversions. Most current
cameras use CMOS (Complementary Metal-Oxide Semiconductors), while a few use CCDs (Charged
Coupled Devices). CFAs acquire color images from these sensors, which capture one color. At the same
time, balance colors are computed using interpolations correlations that could be used to identify
manipulations. Image Enhancement techniques are used to improve image qualities before storage.
Artefacts generated at various phases of image generations can also be used to identify manipulations.
These artefacts are generally estimated using chromatic aberrations, source camera identifications,
CFAs, interpolations, and sensor noise errors where discrepancies indicate manipulations.
Physical environment-based DIFDs: Assuming fakes of two Hollywood stars rumored to be
romantically connected through their beach images. Two individual photos can be superimposed
to create a combined image but fail in mimicking the lighting effects of original images captured
separately. Differences in illuminations throughout an image are proof of manipulations. These
approaches are based on the lighting conditions of the captured original idea, as lighting is crucial to
photography. They can be categorized into three groups. Discrepancies in light sources between certain
items in the images assist in detecting manipulations in physics-based methods . This approach was
first developed by Kee et al.  and used three-dimensional surface geometries.
Geometry-based DIFDs: Geometry-based approaches measure object locations in the environ-
ment of the camera. Geometries can be faked using two sub-divided intrinsic camera parameters-based
techniques, including multi-view geometries, focal lengths, primary points, aspect ratios, skews, and
metrics. The main points or optical axis intersections and planes lie in genuine image centres. When
tiny portions of images are moved or translated, copy-moves or images are merged, called spliced. It
is not easy to retain main image points from the same perspectives .
DIP approaches have traditionally focused on altering image pixels. Hence, pixel-based DIFDs
have been extensively utilized as the most basic and widely used forgery techniques. These approaches
analyse inter-pixel correlations caused by direct/indirect image tampering. Copy-moves, Image splices,
re-samples, and retouches are the most prevalent pixel-based DIFDs, as indicated explained in the
introductory part of this work. In copy-moves, image parts are duplicated and/or within images,
resulting in significant correlations in these areas, which can be utilized as evidence for DIFDs.
However, developing effective characteristics or matching algorithms for DIFDs is a big issue.
DIFDs generally include pre-processing, feature extraction, and classification/detection tech-
niques. These DIFD processes are examined by others researchers in previous works. The review con-
sists of three sub-sections. Where the first section discusses processing methods. While subsequently,
reviews of feature extraction approaches have been done. The third section review of classification and
deep learning methods with its descriptions.
2.1 Preprocessing Methods
Forensic research mainly focuses on strong algorithmic creations to identify where altered image
elements are present. Hence, image noise eliminations are the first step for clearly identifying forgeries.
1492 CMES, 2023, vol.134, no.3
A multitude of approaches exist in pre-processing, like color conversions, image smoothing, CEs
(Contrast Enhancements), HEs (Histogram Equalizations), and MFs (Median Filters). In image
forensics, it is critical to look at these processing processes, which are detailed below. Pre-processing
step typically begins with color to grey scale conversions before the subsequent stage of feature
DIFDs for detecting copy-moves were proposed by Panzade et al. . The scheme called
CMFDs (Copy-Move Forgery Detections) converted RGB (Red Green Blue) images into HSVs
(Hue Saturation Values) in their representations from faked photos. The study used SIFTs (Scale
Invariant Feature Transforms) to extract key points and match them where they were clustered for
final detection. Their comprehensive experimental findings demonstrate detection of cloned areas
successfully. Their scheme also gave good results to geometrically transformed or multi-cloned images.
Many pre-processing methods were combined by Lonnie et al.  in their proposed scheme.
Their pre-processing included HEs and filters (median, Gaussian and sharpening), processed using
SIFTs. The study decreased false matches in identifying forged areas. SIFT key points such as identified
counts, count of matches, and inaccurate match counts were displayed in the survey. Their scheme
coupled SIFTs with pre-processing methods to reduce false matches in 30 tampered photos divided
into three categories. Their optimum pre-processing strategy produced minimal false matches when
tested on image databases.
A unique pre-processing method was proposed by Kuznetsov et al. . Their work used hashing
for the detection of copy-moves and could work on duplicated images. The work used initial image
transformations to integrate modifications. In the second step, Image intensity range reductions,
gradient computations, orthonormal expansions, Alces (adaptive linear contrast enhancements) and
LBPs (local binary patterns) were compared during their tests on their ability to detect DIFDs.
CEs in DIPs can improve the dynamic range of image pixel values, as Chakraverti et al. showed
in their study for detecting copy-moves. Their ORB (Oriented FAST and Rotated BRIEF) approach
was combined with modified Local CLAHEs (Contrast Limited Adaptive Histogram Equalizations),
an alternative to SIFTs. Their experimental results revealed the success and betterment of their
proposal when compared with other techniques in terms of FPRs (False Positive Rates) and TPRs
Cao et al.  proposed two new methods using CEs for digital images that were modified. Their
scheme concentrated on detecting global CEs, which were applied to JPEGs. Theoretically, histogram
peak/gap artefacts caused by JPEG compressions or translation of pixels were examined, and zero-
height gap fingerprints were identified for differentiation. It was followed by a novel method that
detected composite images generated by enforcing contrast adjustments on source image regions. For
detecting applied CEs in source areas, they were identified using block-wise peak/gap groupings. Image
forgeries were detected by evaluating composition borders and the consistency of regional artefacts.
Their extensive tests showed the efficacy of the proposed approaches. However, when CEs was the
final post-processing step and failed on image compressions after executing CEs, it worked efficiently.
Yuan  proposed grey level cumulative distributions of image HEs. The distributions were
represented as discrete identity functions on inherent fingerprints created by global HEs. The study
has observed cumulative distributions matched well with their model. Their classifications recognized
usage of global HEs. Compared to prior approaches, their proposed method differentiated global
HEs from other types of CEs accurately. The proposal’s effectiveness was exhaustively evaluated in
identifying image HEs and resistance against attacks.
CMES, 2023, vol.134, no.3 1493
Powerful MFs were presented by Kang et al.  in their study where residuals of MFs (differences
between original and filtered image versions) were examined statistically. The study fitted their MF
residuals into an AR (Autoregressive) model for capturing statistical characteristics. Their model used
AR coefficients as characteristics for detecting MFs, followed by a series of tests that evaluated the
efficiency of their proposed MF based detections. Their results demonstrated that their proposed
forensic methodology outperformed previous techniques, shallow FPRs and limiting characteristic
MFs for DIFDs were also proposed by Gao et al.  using CFDIs (combined differences
image characteristics). Their CFDIs were a combination of first-order JCPDFs (Joint Conditional
Probability Density Functions) and Second-Order Difference Images. Dimensionality reductions were
executed using PCAs (Principal Component Analyses) for obtaining final features for given threshold
values. The study’s experiments on single/compound databases showed that their proposed scheme
outperformed other approaches on uncompressed/compressed image datasets, mainly on solid JPEG
compressions and low-resolution images.
A different approach for detecting global CEs and copy-paste forgeries was proposed by
Charpe et al. . The study used contrast computations for detecting CEs in images. The proposed
method was resistant to post-processing JPEG compressions, and features from the pictures in copy-
paste duplicates were extracted using DCTs (Discrete Cosine Transforms). The scheme could detect
tiny or medium, or large areas of fabrications in forged images with ease.
Sensor’s pattern noise was used for DIFDs by Chierchia et al.  in their study. Their scheme
recast forgery issue as a Bayesian estimation issue, used appropriate MRFs (Markov Random
Fields) to describe the source’s spatial solid dependencies, and judged each image pixel collectively.
Subsequently, convex optimizations were used to produce globally optimum solutions, and non-local
denoising enhanced PRNUs (Photo-Response Non-Uniformities) for estimations. Their simulation
of genuine forgeries indicated that their proposed approach improved existing approaches successfully
in a wide range of practical instances.
CNNs (Convolution Neural Networks) were used by Chen et al. , where the study presented
convolution filters with an isotropic design. The scheme is minimized the number of CNN parameters
and their proposed filter, rotation-invariant features using equal weights for image forensics. Their
experiments demonstrated that their new rotation-invariant CNNs achieved significantly higher
performances and fewer parameters, improving 13% in Gamma corrective forensics DIFDs. When
compared to the popular BayarNet, it also produced considerably better generalization results on
diverse databases, in addition to its resilience against JPEG compressions.
It may be noted that CLAHEs responsible for RGBs-HSVs conversions play a critical role in
DIFDs pre-processing stages.
Picture enhancement entails methods that increase image quality, allowing for more accurate
visuals for analyses. It is extensively utilized in various applications because it can overcome some
image capture systems’ constraints . Image improvement processes include deburring, noise
reduction, and contrast enhancement. CLAHEs [32,33] are standard approaches for enhancing local
CEs that have proven practical and helpful in various applications [34–36]. CLAHE is a technique for
increasing the visibility of a hazy picture or video.
CLAHE is an AHE variation that minimizes noise amplification. CLAHE has also been proven
to be unsuitable for digital images with fine details. They combined global histogram modification
with CLAHE in Histogram Modified (HM-CLAHE) . Along with the usual CLAHE, local CEs
1494 CMES, 2023, vol.134, no.3
emphasized the small features buried in images and an enhancement parameter to adjust the amount
of enhancement. As a result, combining Local Contrast Modification (LCM) with CLAHE yields
optimum contrast enhancement with all local information of pictures that standard CLAHE may
The authors propose the LCM-CLAHE algorithm, which contains various steps: First, take the
input image. LCM is provided with the original picture and the enhancement parameter as input. We
alter the photo in LCM to generate the more delicate features concealed in the mammography image
and then send that output image to CLAHE, enhancing image quality .
The initial step for image enhancements is an application of CEs to images where both global and
local information are considered for image improvements. Local knowledge of images is captured by a
window defined to the network’s pixel widths. An equation can be used to express the transformation
function Eqs. (1)–(2).
E-parameter for enhancement, M–input image’s global mean, g-enhanced image, f-input image,
m-input image’s local mean, σlocal SD (standard deviation), E–constant in the interval [0, 1]. Eqs. (3)–
(4) are used to calculate m and σfor a user-defined window width,
Averages of windows are computed from the obtained standard deviation and local mean values
utilized in Eqs. (1) and (2). The more delicate features of mammography pictures will be highlighted
with this approach. This approach produces an improved image that is sent into CLAHE.
A picture received in RGB space is transformed into a color space with a brightness (Y) and two
chrominance components (Cb, Cr) as shown in Eq. (5),
65.481 128.553 24.966
−37.797 −74.203 112.000
112.000 −93.786 −18.214⎤
The two chrominance channels are separated, and the number of rectangular contextual tiles
into which the image is split is determined for each chrominance channel. The best value for this
is determined via experimentation. The contrast transform function is built based on a uniform
ic_out =[ic_max −ic_min]∗Fkicin +ic_min (6)
Assuming ic_min is minimum allowed intensity, ic_max stands for maximum allowed intensity and
optimal clipping limit is fixed and if Fk(ic_in)represents cumulative distribution function for inputs ic_in .,
then Eq. (6) mathematically depicts modified chrominance with uniform distributions. Inferences of
pre-processing methods for forgery detection are discussed in Table 1.
CMES, 2023, vol.134, no.3 1495
Table 1: Inferences of preprocessing methods
Author Algorithm Merits Demerits Analysis and
Panzade et al.  RGBs to HSVs Successful
detection of cloned
regions via color
Noise present in
the image is not
removed. It may
quality of the
Lionnie et al.  Histogram
with median and
Gaussian filter and
methods to reduce
This system does
not find suitable
for some tampered
et al. 
LCM-CLAHE The study’s scheme
results in terms of
FPRs and TPRs
are not applied to
Yuan et al.  Histogram
HEs and resilience
to attacks was
Noise present in
the image is not
removal is not
Kang et al.  Median Filter
reduced FPRs and
The work does not
easily apply to
1496 CMES, 2023, vol.134, no.3
Table 1 (continued)
Author Algorithm Merits Demerits Analysis and
Gao et al. MFsbasedon
median filtering in
images becomes a
with QF =70-
Charpe et al.  Global contrast
The technique can
the small, medium
and large size
regions in the
It could not see
low, middle and
Chierchia et al.  Photo-Response
Fails to improve
which help detect
Note: The introduced method uses local CEs for highlighting minute image details and a parameter for controlling enhancements along
with CLAHEs. This usage of LCMs (Local Contrast Modifications) with CLAHEs produced optimal CEs as more local image information
was visible compared to normal CLAHEs.
As a result of the preceding, a variety of techniques to improve detection performance have
been developed. The only difference between the following works is the characteristics utilized in the
forgery detection method. Here, three main categories are introduced to categories these algorithms:
Techniques based on space, transforms, and hybrid techniques (seeFig. 3).
CMES, 2023, vol.134, no.3 1497
Figure 3: Classifications of feature extraction methods for copy-move forgery detection techniques
2.1.1 Spatial Domain Methods
The pixel location directly describes the content of a picture in a spatial feature space, where
the energy is distributed evenly, and nearby pixels are highly associated. As a result, the matching
procedure is highly computationally intensive. In spatial feature spaces, copy-moves can be based on
moments or intensities or key points or textures.
Moment-based methods: Copy-move DIFDs can be assessed using Hue, blur invariance and
Zernike moments. Liu et al.  detected duplicate locations in forged images by rotating the photos
using circle block and Hu moments in their scheme. Their proposed method worked well against
notions with noises, blurs, JPEG compressions and rotations.
Ryu et a l.  proposed copy-move DIFDs detections using Zernike moments to locate duplicated
areas. The presented approach detected forged areas amidst rotations as Zernike moment amplitudes
show invariance in rotations. The proposed method is also resistant to deliberate distortions, including
blurs, JPEG compressions, and the addition of white Gaussian noises. Their experiments showed that
their technique was effective in identifying the faked region in copy-rotate-move forging.
Ryu et a l.  suggested a forensic approach using Zernike moments of tiny picture blocks to locate
duplicated image areas. To dependably reveal repeated measurements following arbitrary rotations, use
rotation invariance characteristics. The block matching process is based on locality-sensitive hashing.
It reduces false positives by looking at the phase of the moments, by using signal properties and
differentiating between “textured” and “smooth” duplicated areas. The proposed approach beats prior
art, especially when the repeated measurements are smooth. Experiments show that the system is
resistant to JPEG compressions, blurs, additive white Gaussian noises, and modest scaling.
Singh et al.  proposed the No-Reference Image Quality Assessment (NR-IQA) technique,
which employs simple spatial filtering operations and is computationally efficient. Laws’ filters, which
are effective in texture analysis, are used to calculate the features. A primary Generalized Regression
Neural Network is used to predict the picture quality score (GRNN). Because of its low computational
complexity, the proposed technique could be used in real-time applications. The proposed method
produced good results amidst distortions with reduced computational cost when compared to most
Intensity-based methods: Images were divided into distinct sub-blocks for computing their energies
using various intensity-based copy-move DIFDs . These techniques support both color and
1498 CMES, 2023, vol.134, no.3
grayscale pictures. However, all of these methods anticipated that the duplicated areas would not
be subjected to post-processing, including scaling or rotations or JPEG compressions. The features
mentioned above’ rotation invariant characteristic has also been proved in a recent study . Bravo-
Solorio et al. [5,43] suggested methods for detecting reflections and rotation invariance in copy-
moves. Pixel’s Log polar transforms from overlapping picture blocks were used to determine the
Key point-based methods: These techniques use high entropy areas to focus the entire picture.
Das et al.  proposed a fast and robust detection method for this type of picture fraud. The image
is first turned to grayscale. The grayscale picture is then divided into four parts using two-level SWTs
(Stationary Wavelet Transforms). The key points were extracted by approximating components of a
decomposed image using SIFTs. The matching pairs of key-points are then discovered. Then, using a
variety of linking techniques, matching pairs of key-points are grouped.
SIFTs and reduced LBPs based histograms were used in Park et al.  method. The 256-level LBP
values collected from local windows centered on key-points were then minimized. A 138-dimensional
is created for each key point to detect copy-move fraud. While comparing the detection accuracy of
the proposed algorithm with current techniques on many image datasets, findings showed that their
proposed method outperformed other copy-move DIFDs in evaluations. Their scheme also exhibited
uniform detections on a multitude of test datasets.
For feature extractions, Amerini et al.  proposed SIFTs that determined whether a copy-
move attack has happened and retrieved the geometric transformation utilized for cloning. Extensive
experimental data shows that the approach can accurately identify the changed region and estimate
the geometric transformation parameters with high accuracy. This approach also handles multiple
Texture-based methods: The human visual system primarily interprets images through texture,
divided into spatial texture and spectral texture properties. The pixel statistics are used to derive texture
characteristics in the spatial domain. They may be computed from any data and are usually noise
Li et al.  exploited textural characteristics using LBPs to manage geometrical changes. Images
were low-pass filtered and subsequently divided into overlapped circular blocks, and finally, uniform
rotation invariance LBPs were employed for extracting features in the pre-processing stage. The study
was resistant to rotations, flips, noises, JPEG compressions, and blurs.
By collecting LBPs based Histogram Fourier Features of blocks, Soni et al.  detected copy-
moves by proposing block-based blind DIFDs. The proposed approach is evaluated using the
CoMoFoD dataset as a benchmark. Experiments indicate that the proposed technique decreases the
time complexity of tamper detections and shows resistance to post-processing assaults such as blurs,
brightness alterations, and contrast adjustments.
Kalsi et al.  proposed the Approximation Image Local Binary Pattern (AILBP) technique for
feature extraction. A typical solitary picture is used to start the number of trials. The experiments
show that the proposed system can provide exemplary performance in terms of speed and accuracy.
Darmet et al.  proposed a method to disentangle source and target areas in copy-move based on
local statistical model of image patches. Zhang et al.  proposed an end-to-end deep learning model
for robust smooth filtering to identify multiple filtering operations simultaneously.
Yang et a l . [ 52] proposed LBPs for rotation invariance in their DIFDs for detecting copy-moves.
The study initially filtered and divided images into constant sized chunks that overlapped. LBPs
CMES, 2023, vol.134, no.3 1499
then extracted image characteristics from blocks and stored them as sorted feature vectors. Euclidean
distance computations between blocks found block pairs. The study’s shift-vector counter C identified
tampered areas, thus demonstrating their DIFDs even on multiple copy-moves while resisting JPEG
compressions, noises, blur rotations, and f lips.
Adding an LBP histogram-based descriptor to the CMFD technique improves it . LBPs for
pixels centered on in 16 ×16 window key points were computed. The histogram of the 256 LBP levels
is utilized as a new descriptor, and the histogram is decreased to 10 levels. A reduced LBPs histogram
was used as additional descriptors to enhance key point pair matches.
LBPs are generic ways to extract textures from images. Their computations are simple with a
higher level of discrimination. LBPs for pixels located at (p, q) can be computed using Eq. (7),
where L(p, q)-centre pixel’s LBP with (p, q) as its location and Inis the intensity of neighboring pixel
and I(p, q) represents the intensity of pixel located at (p, q), and N represents neighboring pixel count
within a defined radius. Eight-bit LBPs are obtained by using a local window centred at (p, q)) and
defined as s(x) in s(x) (8),
s(x)=1, if x ≥0
0, else (8)
Assuming (xi,yi)represents pixels within (16 ×16) local window centred with kias t6he key
point located at (xi,yi). LBPs are computed for all pixels at(p, q)∈(xi,yi), LBP.
Features obtained using SIFTs carry information suitable for CMF detections. Hence, using
complete LBPs may not be required. LBPs with a maximum of two transitions in 0→1or1→0
are categorized as uniform patterns (characterized by consecutive 1’s) where 00110000 is an example.
01010100 six transitions or is a non-uniform pattern. Amongst 256 LBPs, only 58 were found to be
uniform. Assuming Lc(p, q) (c =0, 1, 2, ..., 8) are LBPs with c consecutive 1’s, then L0(p, q) =
00000000, and L8(p, q) =11111111. For c =1, 2, ..., 7, Lc(p, q) can have 8 binary patterns or view as
rotational shifts of the single pattern. Lnon(p, q) are patterns without consecutive 1’s except for L0(p,
q). 256 level LBPs can be divided into 10 groups where Lc(p, q) total nine and Lnon(p, q).
This work uses Lc(p,q),Lnon (p,q)probabilities as CMFD descriptors. Further, the descriptor’s
rotation invariance is maintained by checking Lc(p,q)the occurrences. Values from Lnon (p,q)might
reflect variations occurring due to noises or modified backgrounds or errors in quantization within
small windows. Non-uniform patterns are checked for reducing these effects of variations. A descriptor
riwith its corresponding keypoint kiis depicted in Eq. (9),
where Rc(xi,yi)and Rnon (xi,yi)are the normalized number of occurrences of Lc(p,q)and Lnon (p,q),
respectively, in (xi,yi).Rc(xi,yi)is calculated by Eq. (10),
|(xi,yi)|,for all (p,q)∈(xi,yi)(10)
where #[Lc(p,q)]representsLc(p,q)pattern counts or cardinality in (xi,yi)and Rnon(xi,yi)can
also be found similarly. riis a ten-dimensional feature vector encompassing Lc(p,q)and Lnon (p,q)
1500 CMES, 2023, vol.134, no.3
Eq. (11) is used to get the SIFT based key point descriptor and to obtain a histogram of reduced
where girepresents CMFD’s descriptor with 138-dimensional features as riin contrast to fiis computed
from larger areas and can robustly handle minute pixel changes and errors in quantization caused by
2.1.2 Transformation Based Approaches
Coefficients have lower correlations when transformed, and only a few coefficients have most of
the energy. Hence, only these coefficients are used as features for overlapping blocks. Transformations
can be in three forms, namely Frequencies, textures and reduced dimensions.
Frequency-based transformations: DIs were validated based on pixels by Parveen et al. in
their proposed DIFDs for copy–moves. Their study used five significant steps: (1) color images were
converted to grey-scale images, (2) the converted images were then split into 8 ×8 overlapping blocks,
(3) DCTs were used for extracting features based on feature sets, (4) the blocks were clustered by K-
means clustering (5) features were matched using radix sorts. Their experimental evaluations showed
that the scheme accurately detected forged areas in DIs.
Alahmadi et al.  suggested passive DIFDs based on LBPs and DCTs. First, discriminative
localized features are extracted from the chrominance component of the input picture using 2D-DCTs.
Detections were then carried out using a support vector machine. Experiments on three picture forgery
benchmark datasets showed that their approach outperformed recently developed methods with better
Hayat et al.  proposed DIFDs based on feature reductions using the DWTs and DCTs. After
splitting the DWT images into separate blocks, the DCT is applied. The correlation coefficients are
then used to compare the blocks. A mask-based tampering mechanism is also created as part of the
studies to verify the detection approach. When compared to two other systems in the literature, the
method yields intriguing findings.
Jwaid et al.  used DWTs and PCAs to do productive calculations in light of LBPs. Pre-
process the image to convert it from RGBs to YCbCrs (Yellow, Green, and Blue). Second, the picture
is compressed using the Discrete Wavelet Transform. The guess sub-picture comprises areas with
low recurrence and the most severe data. Covering squares divide the LLs (Low Levels) sub-images.
LBPs and PCAs matched chunks as part of the feature matching process. The final stage used SVMs
(Support Vector Machines) to classify fakes.
Thajeel et al.  created novel stage-wise CMFDs based on QPCETs (Quaternion Polar Complex
Exponential Transforms). The suspicious image is split into blocks that overlap. Second, QPCET is
used to extract invariant characteristics for each block. Finally, k-dimensional tree (kd-tree) block
matching is used to find duplicated picture blocks. Finally, a novel approach is proposed to decrease
false matches caused by flat regions. Experimental results demonstrated the proposed approach’s
exact and efficient recognition capability of copy-moves on rotated, scaled, noises added, blurred,
brightened, colors reduced, and JPEG compressed images. Furthermore, the proposed technique
solves false matches and outperforms other methods in terms of accuracy and false-positive rate. The
technique shown here might be used to conduct accurate digital image forensic investigations.
CMES, 2023, vol.134, no.3 1501
PCETs (Polar Complex Exponential Transforms) were suggested by Wo et al.  to detect copy-
moves. The study extracted rotationally invariant and multi-scale features with PCETs which used
multi-radii with graphic processing unit accelerations. For achieving coarse matches, lexicographical
order matches optimized with minimum heap were used. The radius ratio and location information
were applied to detect changes accurately. The proposed PCETs noticed forged sections created by
rotations or scaling while resisting smoothing, JPEG compressions, and noise degradations.
Soni et al.  proposed an efficient block-based copy-move DIFDs based on FWHTs (Fast
Walsh Hadamard Transforms) to reduce processing time in finding duplicated portions in a picture.
Lexicographical sorting and an effective shift-vector technique are used to detect forged areas.
Park et al.  proposed a ULPFTs (Up sampled Log-Polar Fourier Transforms) based descrip-
tors resistant to rotation, scaling, sheering, and ref lection, among other geometric changes. The
theoretical foundations of the ULPFT representation are first presented. Then, from the ULPF
representation, a feature extraction technique is shown that can extract scale-invariant features
and rotations. Analyzed common CMFDs (Copy-Move Forgery Detections) processing pipeline
and modified a section to handle various forms of tampering assaults more effectively. Simulation
findings show that the introduced feature descriptor outperforms other descriptors with established
performance guarantees. Another benefit of ULPFT is that it has low computational complexity.
A new approach for CMFD of duplicated items was given by Hosny et al. . The bounding
rectangle is designed around the identified item to create a sub-image. The morphological operator is
used to get rid of the tiny things that are not needed. For the identified objects, exact PECT moments
were employed as characteristics and items were compared using Euclidian distances and correlations
between feature vectors.
Emam et al.  used PCETs to extract block’s invariant characteristics, resulting in PCETs based
kernels representing blocks. Second, probable comparable blocks were discovered using LSH (Locality
Sensitive Hashing) and Approximate Nearest Neighbor searches. Morphological techniques are used
to eliminate the incorrectly similar blocks, making the method more resilient. The presented approach
is resilient to geometric changes with minimal computing complexity, according to experimental data.
Zhu et al.  proposed using Gaussian scales and extracting key-points quickly using ORB
features in scales. Subsequently, the input image coordinates of FAST key points were reverted, and
hamming distances matched obtained ORB features between key-points pairs. In its final part, the
scheme eliminated falsely matched key points using RANSACs (Random Sample Consensus). Their
experimentations showed that their approach was effective in detecting geometric transformations,
including rotations and scaling. Further, their system was robust to see forgeries even in Gaussian
blurred, or Gaussian white noised or JPEG recompressed images with high accuracy.
Dimensionality reduction-based methods: Chihaoui et al.  suggested automatically detecting
duplicated areas in the same picture where SIFTs identified local properties of the photographs (sites
of interest), and SVDs matched identical features. The findings demonstrate that the proposed hybrid
approach is resistant to geometrical changes and can detect duplicated areas accurately.
SWTs (Stationary Wavelet Transforms) was proposed by Dixit et al.  to detect copy-moves due
to shifting invariance of SWTs, which assist in similarity detections (matches) and dissimilar detections
(noises) due to blurs in image blocks. The study used SVDs (Singular Value Decompositions) for
deriving image features represented in image blocks. Additionally, the color-based segmentation
technique employed in this study aids in achieving blur invariance. Their experimental findings
showed the suggested technique’s effectiveness in detecting copy-moves of images using intelligent
edge blurring while outperforming most other methods in the accuracy of detections.
1502 CMES, 2023, vol.134, no.3
To describe and detect duplicated blocks in a picture, Hilal et al.  proposed an approach
that used PCAs and DCTs. The algorithm is optimized and tested on a database of forged images.
The algorithm’s f lexibility and performance are demonstrated by comparing the obtained results to a
Sunil et al.  also proposed using DCTs and PCAs to compress overlapping block feature
vectors. The down-sampling of low-frequency DCT coefficients creates features that are invariant to
local changes in intensity.
Images are initially split into overlapping square blocks by Mahmood et al. , then DCT
components were for block representations. Gaussian RBF kernel PCAs reduced the dimensionality
of the feature vectors, thus improving feature matching efficiency. Extensive tests were carried out
to compare the proposed approach with other approaches. Experimental findings show that the
proposed method accurately estimated CMFDs even when pictures were polluted using blurs or noises
or compressions and could identify numerous CMFDs. As a result, the proposed methodology was
computationally efficient and reliable for copy-move DIFDs and enhanced the trust of evidence-
based applications. However, compared to linear PCAs, both KPCAs and SVDs were computationally
inefficient. In contrast, methods described above effectively expressed 2nd order data, forgeries based
on altering high-order statistics, complex to detect.
Several authors have used PCETs, and as a result, PHTs (Polar harmonic transforms) are a
type of complex exponential transforms [6–71]. It is a signal representation technique that uses a
superposition. Harmonics to represent a signal [72–74]. PCETs are beneficial tools for characterizing
In polar co-ordinates (r, θ), the function Hnm(r, θ), includes radial basis function Rn(r) and
angular function exp(jmθ)byEq. (12),
Hnm (r,θ)=Rn(r)exp (jmθ)(12)
where Rn(r)=exp (j2nπr2),n,m=−∞,···,0,···,+∞,0≤r≤1, 0 ≤θ≤2π. Rn(r) is orthogonal
in a unit circle, as in Eq. (13),
where δnnis the Kronecker delta and R∗
n(r)is the conjugate of Rn(r). The function set Hnm(r, θ)is
orthogonal in the unit circle as in Eq. (14),
where πis the normalization factor; δnm ,δnmare Kronecker deltas; and H∗
nm(r, θ) denotes the
conjugate of Hnm(r,θ). n order PCETs with repetition m are as per Eq. (15),
nm (r,θ)rdrd θ
f(r,θ)exp −j2nπr2exp (−jmθ)rdrdθ(15)
CMES, 2023, vol.134, no.3 1503
Based on the theory of complete orthogonal function sets, images can be reconstructed by PCET
coefficient’s infinite orders (|n|≤nmax,|m|≤mmax ).
2.1.3 Hybrid Methods
Many new approaches based on merged methods have recently been introduced to improve
the performances of copy-moves in images. Yang et al.  proposed KAZE, a robust interest
point detector that may be used in conjunction with SIFTs to extract additional feature points
where enhanced matches are employed to deal with numerous duplications, and n-best matches in
features can be discovered. Then, to eliminate false conflicts, a practical filtering step based on
picture segmentation is performed. Furthermore, an iterative approach for estimating transformation
matrices and determining the presence of forgeries is devised. The duplicated areas may be found at pix
using these matrices. According to experimental results, the presented approach accurately detected
duplicate areas even after distortions, including rotations, JPEG compressions, noise additions, and
Lin et al.  proposed a hybrid feature and evaluation clustering-based region duplication
detection technique. The proposed system is broken down into two stages: rough matching and precise
matching. Rough matching begins with the extraction of hybrid key points from the input picture, then
is characterized by unified descriptors. Second, the Neural Network (NN) approach matches those key
points. Third, the introduced clustering based on evaluation groups those matching key points. Fourth,
affine transformations between these groups are approximated, and Bag of Word (BoW) is utilized to
filter inaccurate affine transformations to enhance pixel-level results. When no affine transformation
can be found, each suspicious region is addressed independently in precise matching. Under various
situations, their suggested approach outperformed most other methods.
CMFDs were proposed by Tinnathi et al.  based on both block and key point techniques.
Adaptive watershed segmentation is utilized to split the forged picture into non-overlapping segments,
and adaptive H-minima transform retrieves the markers. In addition, a AGSOs (Adaptive Galactic
Swarm Optimizations) to find the best gap value when picking the tags can improve segmentation
performance by eliminating unwanted regional minima. After that, using HWHTs (Hybrid Wavelet
Hadamard Transforms), the features from each segment are retrieved. Adaptive thresholding was
then used to accomplish feature matches. RANSAC’s (Random Sample Consensus) eliminated false
matches or outliers. Finally, the FREA (Forgery Region Extraction Algorithm) was used to detect
the duplicated region from the host picture. The presented technique successfully detects the picture
forgery region, according to the results of the experiments.
Sunitha et al.  proposed CMFDs using key-points, combined feature extractions, and hierar-
chical clustering to identify forgeries. According to their experimental results, their suggested DIFDs
achieved considerably higher performances when compared to other methods, according to their
A machine learning classification approach was presented by Jaiswal et al. . Spliced and non-
spliced pictures were divided into two categories using logistic regression. For this, a feature vector
was created using a mixture of four handmade features taken from photos. Then, using a logistic
regression classification model, these feature vectors are trained. The outcome was evaluated using a
ten-fold cross-validation test assessment technique. Finally, the study’s comparisons of their suggested
approach with other methods on three publicly available datasets discovered that the acquired findings
outperformed other techniques.
1504 CMES, 2023, vol.134, no.3
FMTs (Fourier-Mellin Transforms) along with SIFTs were used by Meena et al.  in their hybrid
approach for DIFDs. The study separated smooth and rough parts of images followed by key point’s
extractions from image textures using SIFTs. FMTs were applied on softer image parts for extracting
their features which were then compared for detecting forgeries. Their scheme outperformed CMFD
algorithms when tested with post-processing procedures and geometric transformations within an
acceptable amount of time. In Tab le 2 , the inferences of feature extraction approaches for forgery
detection are well described.
Table 2: Inferences of feature extraction methods
Author Algorithm Merits Demerits Dataset and analysis
et al. 
Zernike moments Detected
Extended dataset from
et al. 
Zernike moments Reduce false
Does not focus on
Pixel Detection Accuracy
(PDA)-99.93%, Pixel False
Positive (PFP) rate-19.8%
True Positive Rate
et al. 
SWTs and SIFTs Improved accurate
with lesser FPRs
features from the
FPR =4%, FNR =10%,
et al. 
SIFTs and reduced
accuracy in CMF
Even after feature
could not be
MICC-F220, CMH, D,
and COVERAGE, TPRs,
FPRs, and accuracies
CMES, 2023, vol.134, no.3 1505
Table 2 (continued)
Author Algorithm Merits Demerits Dataset and analysis
et al. 
LBPs Robust LBPs in
selection of the
dimension of the
features to make
the method strong
Color Image Database
Correct detection ratio,
False detection ratio
True detection ratio of 0.8
with <0.1 false detection
et al. 
terms of accuracy
It does not extend
to detecting forged
various types of
Synthetic dataset, House
Image 300 ×300-execution
et al. 
(LBPs with DWTs The scheme
matched blocks in
applicable to the
large image size
size-23 ×23 =98.035%,
et al. 
Reduces the flat
Quick and robust
CMFD database, CDRs
(correct detection ratios)
=98.0% and FDRs (false
detection ratios) =2.3%
against blurring attacks
against brightness change
against color reduction
et al. 
PCETs Detected copied
PCETs are highly
complex in their
use of multi-radius
dataset (IMD), precision
(95.2%), recall (51.9%) and
et al. 
PCETs a robust approach
holes and isolated
pixels but could
Precision, recall, F1-Score
1506 CMES, 2023, vol.134, no.3
Table 2 (continued)
Author Algorithm Merits Demerits Dataset and analysis
et al. 
Copied areas are
detected in images
affected by attacks
MICC-F600 dataset, and
Bench mark dataset
It illustrates AGSOs ability
to identify forged images
with Precision =95.31%;
Recall =96.89% and F1
=95.28% on MICC-F600
Precision =99.08%; Recall
=98.42% and F1 =98.33%
on Bench mark dataset
under plain copy move at
et al. 
using lines without
MICC-F220 dataset EKP-
et al. 
FMTs and key
of the image
Dataset (IMD), GRIP
dataset, Precision, Recall,
recall =95.83% and
recall =97.92% and
recall =100% and
recall =68.75% and
CMES, 2023, vol.134, no.3 1507
2.2 Review of Classification and Deep Learning Methods
The job of detecting picture fraud among legitimate and counterfeit photos is a binary classifica-
tion example [79–83]. This part will go through the specifics of detecting forgeries using conventional
matching and deep learning approaches.
Cozzolino et al.  suggested CMFDs and localization based on dense nearest-neighbor field
computations in a short amount of time. The study’s Patch Match was a recursive randomized method
for nearest-neighbor search that uses image regularities to converge to a near-optimal and smooth
field swiftly is utilized to do. Modify the fundamental algorithm to make it more resistant to rotations
while maintaining its computational efficiency. Experiments demonstrate that the presented approach
outperforms all evaluated reference procedures in terms of accuracy and speed nearly equally.
Copy–moves were detected in DIs by Bi et al.  with their proposed MLDDs (Multi-Level
Dense Descriptors) based on Hierarchical Feature Matches. The study’s descriptors (Color Textures
and Invariant Moments) were extracted at several levels with MLDDs. Their Hierarchical Feature
Matched identified forged areas in DIs after computing MLDDs for pixels. The pixels with comparable
color textures were sorted into different neighbor pixel groups and geometric invariant moments of
pixels identified a pixel’s corresponding neighbors. Subsequently, Adaptive Distances and Orientations
are used for Filtering superfluous pixels based on generated/matched pixel pairs. The study used mor-
phology for their final outputs, where forged areas were identified. Their experiments demonstrated
the scheme’s sturdiness even under demanding conditions, including geometric transformations, JPEG
compressions, noise additions, and down sampling compared to other CMFDs.
Key points identified forged smooth areas in DIs by Wang et al.  in their scheme based
on CMFDs. Their scheme divided tampered DIs into non-overlapping/irregular super pixels before
categorizing super pixels from smoothened, normal textured and strongly textured backgrounds based
on local information entropies. Subsequently, adaptive feature point detectors extracted DIs key points
from super pixels belonging to different textures and generated local visual characteristics (moment
magnitudes) for these super pixel key points. A reversed generalized dual Nearest Neighbor Algorithm
discovered key point’s matches quickly. In the final stage, erroneous key-points were eliminated by
random sample consensus and forged areas were normalized using zero-mean normalized cross-
correlations. The scheme showed its superiority compared to other CMFDs in detecting copy-moves
in DIs generated with geometric transformations, JPEG compressions, and additive white Gaussian
Bi et al.  segmented host images into irregular but non-overlapping patches using multiple
scales. The study used SIFTs to extract multi-scale feature points from these patches. Subsequently,
APMs (Adaptive Patch Matches) identified suspicious/forged areas for the used scales. Matched Key
points were merged, and suspect regions of multiple scales were combined in the final stage to identify
forgeries. Their experimental evaluations outperformed other CMFDs in DIFDs on images forged
with geometric transformations, JPEG compressions, and specific noise additions, including multiple
copies and down sampling.
Li et al.  developed a quick and efficient CMFD method using hierarchical feature point
matches. First, prove that reducing contrast thresholds and image rescales required key-points even
in tiny or smoothened areas. Then, to handle more key point counts in matches, a unique hierarchical
matching technique was devised. Unique iterative localizations decreased false alarm rates and pre-
cisely located tampered regions with the assistance of key-points’ resilient features, including leading
orientations, size information, and color information. Extensive experimental data are presented to
illustrate the suggested scheme’s improved performance in terms of efficiency and accuracy.
1508 CMES, 2023, vol.134, no.3
Wang et al .  proposed color invariance as the base for their DIFDs where SIFER (Scale-
Invariant Feature Detector with Error Resilience) and FQRHFMs (Fast Quaternion Radial Harmonic
Fourier Moments) were used for detecting copy-moves. Their scheme derived adaptively stable key
points from super pixels by integrating block contents with SIFER color invariance after segmenting
DIs into non-overlapping uniform super pixel blocks. The extracted key-points and local image
features were used to build Delaunay triangles using FQRHFMs and gradient entropies. The proposal
matched Delaunay triangles using CSHs (Coherency Sensitive Hashing) followed by DLFs (Dense
Linear Fittings). Errors in Delaunay triangle matches were eliminated by localizing forged areas using
ZNCCs (Zero Mean Normalized Cross-Correlations). The study’s extensive tests assessed the scheme’s
efficacy in copy-move forgeries with positive finds.
Copy-moves were also detected by Yang et al.  in their study using multi granular super
pixel matches. Their approach combined key point and block-based features for their DIFDs. The
scheme extracted stable key-points in DIs from coarse granular super pixels after dividing DIs into
non-overlapping/irregular coarse granular super pixel-based blocks. Each of these extracted super
pixel’s features was considered quaternion exponent moment magnitudes and used for rough granular
super pixel matches where E2LSHs (Exact Euclidean Locality Sensitive Hashing algorithms) quickly
identified forged areas. In the final step, finely granulated super pixels were separated and replaced
by their key point matches followed by morphological operations on delicate granular super pixel
neighboring areas, which were then combined to yield the identified forgery areas. When tested on
publicly available online datasets, extensive experimental findings showed that the presented method
performs well under a range of demanding situations compared to other techniques.
Liu et al.  proposed a new Key point and Patch Match-based CMFDs. To obtain trustworthy
key-points, LIOPs (Local Intensity Order Patterned), robust key point descriptors was coupled with
SIFTs. After matching the collected key points using g2NN, the redundancy matched key point pairs
were eliminated using matched key point pair descriptions and filters based on density grids. Finally, an
improved method for matching patches was used to evaluate key-point pair matches to detect forgeries
properly. According to their experimental results, their presented technique accurately saw copy-moves
in images better than existing methods and performed well even on deformed images processed by
rotations, JPEG compressions, and noise additions and scaling.
Elhaminia et al.  proposed treating CMFDs as MRFs in labelling. Pre-processing included
over segmentations to create super pixels which were treated markov network nodes for balancing
precision and speed. The maximal a posteriori labelling can accurately map the forged areas while
selecting unary and binary potentials intelligently. According to qualitative and quantitative compar-
isons with other methods utilizing public benchmarks, the presented technique can enhance accuracy
while keeping processing demands low.
Cozzolino et al.  detected copy-moves accurately using rotational invariance, where these
characteristics were computed based on localizations. The study’s Patch Match was a dense-field
approach that showed better performances when compared to key point approaches but took longer
execution time in feature matches. The study computed dense fields in DIs and handled invariant
characteristics more effectively, increasing resilience against rotation or scaling. Furthermore, based
on the output field’s smoothness, a dependable and straightforward post-processing technique was
devised. According to their experimental research on available online datasets, their approach was
accurate with greater resilience and quicker than most dense field references.
Ouyang et al.  proposed CNN based CMFDs. The proposed technique takes an already
trained model from an extensive database, such as ImageNet, and tweaks the net structure significantly
CMES, 2023, vol.134, no.3 1509
with tiny training examples. Experiments demonstrate that the proposed approach produced good
counterfeit images automatically using simple image copy-moves using computers.
Liu et al.  proposed that CMFD be performed using CKNs (Convolution Kernel Networks).
CKNs are deep convolution architectures based on data-driven local descriptors. Because of its high
discriminative capabilities, it can produce competitive results. Three essential modifications are made
to better adapt to the situation of CMFD: First and foremost, CKN has been rebuilt for use with a
Graphical Processing Unit (GPU). The GPU-based reconstruction achieves excellent efficiency and
allows hundreds of patch matching in CMFD to be applied. Second, to create homogenously dispersed
key points, a segmentation-based key point distribution technique is presented. Finally, an adaptive
over-segmentation approach is employed. Experiments on publically available datasets are carried out
to verify the proposed method’s other performance.
Thakur et al.  concentrated on efficient splicing detection and CMFD pipeline design, which
focuses on identifying the traces left by different Splicing and copy-move forgeries post-processing
activities like JPEG compressions or noises or blurs or contrast adjustments. Their use of LFRs
(Laplacian Filter Residuals) and SDMFRs (Second Difference of Median Filters) on images as one
of the residuals were introduced jointly to suppress image content and focus solely on the traces of
Agarwal et al.  developed a deep learning-based approach for identifying the CMFD. The
newly developed method uses the altered picture as the system’s first input for detecting the tampered
region. Segmentation, feature extraction, dense depth reconstruction, and ultimately seeing the
tampered areas are all part of this method. The newly developed deep learning-based method may
reduce computing time and improve the accuracy of duplicated region detection.
BRISKs (Binary Robust Invariant Scalable Key points) were used as Yang et al.  descriptors in
their study. Their approach used adaptive uniform threshold value distributions to extract key-points
of local features from DIs. The study then used the embedded random ferns approach for formulating
needed matches, thus achieving discriminative classifications. Their local descriptors based on BRISKs
matched image key-points. The study also used RANSAC’s to eliminate erroneous key point pairs
Normalized Intensity Correlations to detect tampers in DIs. Their experiments with other CMFDs
showed their scheme’s enhanced detection and localization accuracies even under adverse image
DCNNs (Deep Convolution Neural Networks) were exploited by Kao et al.  for offline hand
signature verifications. The study used novel local feature extractions using SigComp on ICDAR
(Document Analysis and Recognitions) 2011 dataset. Their training on the authenticity of signatures
was used for testing unsaved fresh author’s signatures.
Feng et al.  proposed a CNN based picture forgery detection method to achieve image pre-
processing for the Columbia University picture mosaic detection dataset. SRM and high-pass filtering
are introduced initially instead of the standard feature of extracting related features based on image
content. The CNN then completed the training and verification processing. On the classification
findings, the impacts of pre-processing and the number of convolution layers are thoroughly compared.
Experiments indicate that the CNN approach described in this paper is successful and resilient in
classifying picture forgeries.
Agarwal et al.  compared different deep learning-based forgery detection approaches with
strategies that do not utilize NN architecture for feature extraction. To identify these photos as factual
or fabricated, developing an effective image forgery detection system is necessary.
1510 CMES, 2023, vol.134, no.3
Zhong et al.  proposed a Dense-Inception Net-based image CMFD method. Dense-
InceptionNet, which are multi-dimensional DNNs (Deep Neural Networks) with dense feature
connections. Their DNNs learnt feature correlations in training and used their learning to match
forgeries. The use of PFEs (Pyramid Feature Extractors), FCMs (Feature Correlation Matches)
and HPP (Hierarchical Post-Processing) modules were a part of their Dense-Inception Nets. PFEs
extracted dense multi-dimensional/scale features where each layer was linked directly to previous
layers. FCMs learnt strong correlations amongst features for producing candidate matches as maps.
HPP in the final stage used these maps to generate cross entropies using training’s back propagations.
Their experimental results showed that their Dense-Inception Net approach produced efficient DIFDs
while proving to guard against most known assaults.
For Copy-Move Forgery Detection, Pun et al.  proposed a two-stage localization method
CMFDs. SLIC (Simple Linear Iterative Clustering) divided images into meaningful patches in the first
step, preliminary localization. The WLDs (Weber Local Descriptors) computed and extracted feature
from each super pixel is then presented. The super pixel matches are then obtained using a matching
threshold based on an experimental study. In the final step, weak super pixel’s Euclidean distances were
used to generate suspect approximations. The study’s DAFMTs (Discrete Analytic Fourier–Mellin
Transformations) extracted image characteristics at a blocking lever while localizations were given by
sliding varying radii circular blocks in suspect areas. The study’s generated candidate circular blocks
were matched by LSHs (Locality-Sensitive Hastings). Poor matches were filtered and eliminated in
identified areas to obtain final identified regions, and geometric morphological techniques were used.
The extensive experimental findings show that the proposed approach outperformed other CMFD
methods on available benchmark databases.
Silva et al.  proposed a new method to CMFD based on a digital image’s multi-scale
analysis and voting procedures. Extract interest points from a suspicious image resilient to scaling
and rotation, then look for probable correspondences between them. Based on geometric restrictions,
corresponding group points into regions. Following that, a multi-scale picture representation is built
for each scale. The produced groups are examined using a highly resilient descriptor to rotation,
scaling, and somewhat robust to compression, reducing the search space of duplicated regions yielding
a detection map. The ultimate choice is made once all detection maps have been voted. Validate the
approach using a variety of datasets that include both original and realistic picture cloning. Compare
and contrast the proposed method with 15 others found in the literature and present promising
Pun et al.  proposed an adaptive over-segmentation and feature point matching CMFD
method. The proposed technique combined forgery detection methods based on blocks and key-
points. The study’s over segmentations adaptively divided DIs into non-overlapping/irregular blocks to
extract feature points from blocks that were matched to identify labelled feature points, thus indicating
forged areas in DIs. They replaced feature points in forged areas with tiny super pixels and combined
neighboring blocks similar to feature blocks’local color characteristics, thus creating a merged region.
As the last step, morphological operations identified forged regions from the merged regions. Their
proposed scheme identified developed regions as very effective when compared to other DIFDs for
copy-move detections. Their experimental results also showed that their copy-move forgery detection
system produced considerably superior detection results under various demanding situations.
Li et al.  proposed a method for CMFD in images based on the extraction of key points. The
study’s approach differed from prior approaches in dividing DIs as semantically independent blocks
CMES, 2023, vol.134, no.3 1511
before key point extractions. Subsequently, these extracted patches were matched for identifying copy-
moves. The study’s matches used two steps where initially suspicious patch pairs included forged areas
in the first step to estimate approximate affine transform matrices. EMs (Expectation-Maximizations)
was used subsequently to refine the estimated matrices and establish their existence. Comparing
the presented method to other strategies on public datasets, experimental findings showed that it
Bi et al.  proposed a new and fast reflecting offset-guided CMFD image searching tech-
nique. The features are retrieved, and feature correspondences are randomly allocated during the
initialization step to get initial mapping offsets, while reflective offsets were computed in searches
to obtain mapping offsets as copy-move forgery mapping offsets. Then copy-move forgery mapping
offsets were disseminated to enhance mapping and ref lective offsets based on priority feature matches.
Finally, only a few iterations can completely detect the forgeries areas from the mapping offsets.
According to experimental data, the presented approach for image copy-move DIFDs decreased
computational complexities while providing higher detection results than other CMFD algorithms,
even under challenging situations.
Bi et al.  proposed a method for detecting CMFDs for accuracy and resilience. To build
feature correspondences in images, the study used an enhanced coherency sensitive hashing algorithm.
A local bidirectional coherency error factor was used in iterations for improved accuracy and advanced
feature correspondences. The iterative procedure ended when the local bidirectional coherency error
fluctuations were less than a predefined threshold, suggesting that feature correspondences were stable.
This error of each feature was used to recognize copy-moves from regular feature correspondences.
Their experimental findings demonstrated that their proposed detection technique was successful in
real-time/near real-time data and produced excellent detection results compared to other copy-move
DIFDs even under challenging situations.
CNNs were used by Rao et al.  in their study for their CMFDs, where CNNs learnt RGB color
image’s hierarchical representations from DIs. Their CNNs spliced images for detecting copy-moves.
Instead of randomizing weights, the study network’s first layer was initialized by high-pass filter sets
and residual maps computed for SRM (Spatial Rich Models). The model regularized DIs efficiently
to suppress image effects and capture artefacts introduced in image tampers. The study’s pre-trained
CNNs extracted dense features of DIs, followed by feature fusions to explore discriminative features
for classifying using SVMs. The scheme’s experimental results on multiple datasets showed that their
CNNs outperformed most other methods.
Fig. 4 depicts the suggested CNNs’ architecture, including 8 convolutions, 2 pooling, and 1 fully-
connected layer with a bi-way softmax classifier. Patches of 128×128 ×3 (128 ×128 patch, 3 color
channels) make up the CNN’s input volume. The first and second convolution layers contain 30 kernels
with a receptive field of 55, whereas the subsequent layers all have 16 kernels with a receptive field of 33.
ReLUs (Rectified Linear Units) were applied to neurons for the activation function to preferentially
react to relevant signals in the input using a size 22 filter, which resizes the input spatially and discards
75% of the activations. This is because the max-pooling process aids in the retention of additional
texture data and improves convergence performance. Local response normalization is also used to the
feature maps before the pooling layer to increase generalization, where the surrounding pixel values
normalize the center value in each neighborhood. Finally, through “dropout,” which sets the neurons
in the fully-connected layer to zero with a probability of 0.5, the recovered 400-D features (5516)
1512 CMES, 2023, vol.134, no.3
are transferred to the fully-connected layer with a bi-way softmax classifier. This usage of the fully-
connected layer at the end was different from other traditional CNNs using 2 or more fully connected
layers, leading to overfitting, especially in small training sets.
Figure 4: The architecture of the proposed 10-layer CNN
Wu et al.  also proposed end-to-end DNNs in DIFDs. The study’s CNNs extracted block fea-
tures from DIs and computed self-correlations between blocks with extracted feature points matched
to rebuild forged/masked areas using de-convolutions. In contrast to traditional approaches, which
needed multiple training and parameter tuning steps followed by post-process spans, the proposed
method eliminated multiple training/parameter adjustments. The study’s scheme was trainable as it
combined forged mask reconstructions with loss optimizations. Their experimental results showed that
their proposed scheme beat other traditional DIFDs based on their matching schemes and effectively
against assaults, including affine transforms, JPEG compressions, and blurs.
Wu et al.  proposed BusterNet, a new DNNs for CMFD images. BusterNet is an accurate,
end-to-end trainable DNNs system, unlike prior efforts. It has a two-branch design with a fusion
module in the middle. The two branches, respectively, locate possible manipulation locations (by
checking for visual artefacts) and copy-move regions (by evaluating visual similarities). This is the
first CMFD algorithm that can identify source/target areas with discernibility to the best of our
knowledge. Simple techniques for generating large-scale CMFD samples from out-of-domain datasets
are shown, as well as stage-wise BusterNet training procedures. According to extensive tests, Buster-
Net considerably beats other copy-move detection algorithms on the two publicly accessible datasets,
CASIA and CoMo-FoD, and is resistant against other known assaults. However, it is preferable to
take these features into account directly. Thus BusterNet is recommended as a two-branch DNN
Dashed blocks are only activated during branch training. Output mask of the main task, i.e.,
cis colour coded to represent pixel classes, namely pristine (blue), source copy (green), and target
copy (red). Output masks of auxiliary tasks, i.e., MX
sare binary where white pixels indicate
manipulated/similar pixels of interests, respectively. In dual branch DNN based CMFDs, dashed
blocks are active only in training. The main outputs (MX
c) are color codes representing pixel classes:
pristine (blue), source copy (green), and target copy (red) where output’s masks of auxiliary tasks
s) are white pixels in binary representing manipulated/similar pixels. An input image X’s
features are extracted using CNNs, and feature maps are scaled to original image sizes using Binary
Classifier (Mask Decoder) to create manipulation masks. CNN Feature Extractor may be used with
any CNN. Because of their simplicity, the first four blocks of the VGG16 design are employed here.
The resultant CNN feature has a resolution of 1,616,512, significantly lower than that required by the
Bunk et al.  detected altered imaging with two approaches based on a combination of
deep learning and re-samples. The study’s Radon initially transformed re-sampled images computed
from overlapped image patches converted to heat-map using deep learning classifiers and Gaussian
CMES, 2023, vol.134, no.3 1513
conditional random field models. The approach used Random Walker segmentations to identify tam-
pered areas in images. LSTMs (Long Short-Term Memories) used these overlapped images as inputs
for localizations and classifications. Their experiment results demonstrated that both approaches
successfully identified and localized forgeries in DIs in terms of detection/localization capabilities.
Qiao et al.  used a linear parametric model to examine the problem of picture resampling
detection. First, reveal the one-dimensional 1-D resampled signal’s rare artefact. The detector is
built based on the likelihood of residual noise recovered from the resampled signal using a linear
parametric model after dealing with the nuisance parameters and Bayes’ rule. After that, focus on the
characteristics of a resampled image. Meanwhile, it calculates the likelihood of pixel noise and creates
a realistic LRT (Likelihood Ratio Tests). Numerical studies indicate the significance of the presented
technique in recognizing uncompressed/compressed resampled pictures compared to another testing.
Amerini et al.  proposed a novel CMFD and localization schema based on the J-Linkage
method, achieving robust clustering in the geometric transformation space. Experiments on several
datasets demonstrate that the proposed method surpasses other comparable strategies for CMFD
reliability and accuracy in the modified patch localization.
Bayar et al.  proposed a new CNN-based technique for camera model identification that is
resampling and recompression resistant. A new low-level feature extraction method is presented that
employs both a restricted convolution layer and a nonlinear residual feature extractor in tandem. These
layers’ feature maps are then concatenated and sent on to subsequent convolution layers for feature
extraction. The presented technique improves camera model recognition performance in resampled
and recompressed pictures according to experimental data. When CNN is employed without ACFM
in experiments, we utilize the architecture which refers to as Non ACFM-based CNN. It’s worth noting
that the ACFM method may be extended and used to various types of nonlinear features to add variety
to current features and improve CNN’s resilience in real-world settings.
Non ACFM-based CNN  is for Convolution Feature Maps Augmentation; BN stands for
Batch-Normalization Layer; TanH stands for Hyperbolic Tangent Layer; ET stands for Extremely
CNNs for use in forensics was proposed by Bayar et al.  in their study. CNNs learning
for classifying features tend to be on the current state of images contents. The study defined a new
CNN layer called the restricted convolution layer, which concealed image contents during learning
and manipulations were detected adaptively to overcome this issue. The study showed that their
restricted CNNs could learn about manipulations directly in a series of tests where experimental results
outperformed current other general-purpose manipulation detectors in DIFDs. Moreover, in cases of
source camera model mismatched between training and testing data, their restricted CNN could still
identify picture alterations correctly. The proposed method is made up of four separate conceptual
blocks that may be used to: (i) use a block of 11 convolution filters to jointly suppress an image’s
content and learn prediction error features while training, (ii) extract higher-level representations
of previously learned image manipulation features, and (iii) learn new connections between feature
mappings in a deeper layer. These filters learn a linear combination of features in the exact location
but from a different feature map across channels. The classification block, which has three completely
linked layers, receives the output of the latter block. The input layer of our CNN in this study is a
grayscale picture patch with 256 ×256 pixels.
Bondi et al.  proposed a method for detecting and localizing image manipulation based
on different camera types’ distinctive imprints on pictures. The algorithm’s logic is that all pixels in
immaculate photos should be recognized as being captured by a single device. In contrast, evidence
1514 CMES, 2023, vol.134, no.3
of several devices can be identified when a picture is created by image composition. The proposed
technique uses a CNN to extract typical camera model characteristics from picture patches. These
characteristics are then examined using iterative clustering algorithms to determine if a picture is
fabricated and pinpoint the foreign location.
In low-resolution pictures, Zhang et al.  proposed a Shallow Convolution Neural Network
(SCNN) capable of identifying the borders of fabricated areas from original edges. SCNN was created
to make use of Chroma and saturation data. Two methods based on SCNN have been developed
to identify and localize picture forgery areas: Sliding Windows Detection (SWD) and Fast SCNN.
The CASIA 2.0 dataset is used to test this model. The results demonstrate that Fast SCNN operates
effectively on low-resolution pictures and outperforms the other significantly. Several works [7,118–
120] utilized this CNN. A CNN [6,121,122] is a multilayered neural network with a unique design to
detect complicated data characteristics. Pixels are the building blocks of images. A number between 0
and 255  is assigned to each pixel. A neural network based analysis [124–126] for image forensics
based on localization of features.
Introduce a Fast SCNN method that is quicker and more efficient than SWD, inspired by the Fast
RCNN. The suggested Fast SCNN computes all of the image’s CNN characteristics. After that, the
features are sent to SCNN’s fully linked layers. Table 3 clearly explains classification inferences and
deep learning approaches for forgery detection.
Table 3: Inferences of classification and deep learning methods
Author Algorithm Merits Demerits Dataset and analysis
et al. 
Patch Match and
Robust to rotations
CPU-time =54.73 s
et al. 
3-step pipeline for
et al. 
DIs could have
been hidden in
FAU, a n d GRIP,
computation cost =2760 s
CMES, 2023, vol.134, no.3 1515
Table 3 (continued)
Author Algorithm Merits Demerits Dataset and analysis
et al. 
CNN classifier Achieve better
could not detect
UCID, OXFORD, CMFD
Error =2.32%, 2.43%,
4.2% for UCID ,
OXFORD, CMFD image
et al. 
The GPU version
was robust to
dataset =55.99%, 78.25%
et al. 
small sample size
ICDAR 2011 SigComp
Testing accuracy =99.96%
Test FR R =0%,
Test FAR =0.22%
Training accuracy =100%,
Testing accuracy =90.85%
Test FR R =2.83%,
Test FAR =16.31%
et al. 
CNN classifier Effective and
robust for image
image mosaic detection
No pre-processing layer
General filtering =83.94%
Combination of SRM
filtering and general
1516 CMES, 2023, vol.134, no.3
Table 3 (continued)
Author Algorithm Merits Demerits Dataset and analysis
et al. 
attacks and could
FAU, CASIA CMFD, and
58.85%, 64.29% for
42.20%, 44.41% for FAU
et al. 
CNN classifier The advantage of
was original image
layers with more
lead to overfitting
in small training
CASIA v1.0, CASIA v2.0
and Columbia grey
Spatial Rich Model
88.24% for CASIA v1.0
97.30% for CASIA v2.0
74.67% for DVMM
et al. 
BusterNet Discernibility to
CASIA and CoMo-FoD
Image Level Evaluation
Recall, F-Score =78.22%,
73.89%, 75.98% ,
Processing Speed =0.62 s
for CASIA Dataset
F-score =83.52%, 78.75%,
80.09% for CoMo-FoD
et al. 
LSTMs Effective in
No visual changes
images fail to
perform well in
NIST Nimble 2016
Accuracy =94.86%, and
CMES, 2023, vol.134, no.3 1517
Table 3 (continued)
Author Algorithm Merits Demerits Dataset and analysis
et al. 
since it is a
solution of the
Dresden Image Database,
et al. 
Sharp edges are a
good indicator for
images; it is not
images that are
CASIA 2.0 Dataset,
Fast SCNN =85.35%
(JPEG), 82.93% (TIFF)
There is also a deep fake based technology which is emerging and hottest branch of image forgery.
In this GAN’s network is used to learn the probability distribution based on examples and generate
the images which are very similar to original images. Li et al.  proposed a method to detecting
forgery in face images generated by unseen face manipulation. In this method author used face x-ray
based on blending step. Li et al.  proposed a novel frequency-aware discriminative feature learning
framework based on single-center loss. Haliassos et al.  proposed Lip Forensics technique which
targets high-level semantic irregularities in mouth movements.
Shang et al.  proposed pixel-region relation network which exploit pixel wise and region-
wise relations for face forgery detection. It is very helpful to detect deep fake face forgery detection.
Hu et al.  proposed dynamic Inconsistency-aware Network which uses CRM to capture global
and local inter-frame inconsistencies to detect deep fake forged video.
Image inpainting forgery is task to reconstructing some regions in the image. It is the used
in applications like object removal, image manipulations. This type of forgery has been focused in
Elharrouss et al. . Where various method has been discussed based on this problem. Currently,
various work has been done to improve image inpainting. To tackle this type of forgery 
proposed method like impainting quality assessment tool using local features. Zhu et al. 
proposed a encoder-decoder network to detect patch-based inpainting operation as shown in Fig. 5.
Zhang et al.  proposed a feature pyramid network for diffusion-based image inpainting.
1518 CMES, 2023, vol.134, no.3
Figure 5: The layout, architecture and parameter settings of the CNN based on Encoder and Decoder
network for inpainting forensics
Zhang et al.  proposed a Federated Learning to face forgery video detection, which is trained
with decentralized data. Inconsistency-Capture module (ICM)  to capture the dynamic inconsis-
tencies between adjacent frames of face forgery videos. Qian et al.  proposed a novel Frequency
in Face Forgery Network taking advantages of frequency-aware decomposed image components, and
2) local frequency. Li et al.  proposed a novel feature learning framework for single-center loss
which compresses mere intra-class variations of natural faces.
There are some other model which exploits the deep learning based approach such as DLFM-
CMDFC , deep learning by recompression , copy-move image forgery , CNN by using
the architecture of ResNet50v2 .
A convolution is a two-function integration that demonstrates how one inf luences the other.
The input picture, the feature detector, and the feature map are key components in this process as
in Eq. (16),
For input images, the feature detector (Kernel) could be a 3 ×3or7×7 matrix. The kernel
multiplies matrix representations of the image elements to generate feature maps (convolved features
or activation maps) to reduce image sizes and hasten to process. Though specific image characteristics
might be lost, essential factors necessary for DIFDs are retained.
2) ReLu (Rectified Linear Unit)
This function boosts non-linearity in CNNs, and items that are not linear to each other are used
to generate images. The primary use of ReLu is to handle image classification as a non-linear issue.
The idea of spatial invariance states that the position of an object in an image has no bearing on the
neural network’s capacity to recognize its unique characteristics. CNNs Pooling (maximum/minimum)
identifies image characteristics irrespective of illumination/camera angle differences s. A 2 ×2matrix
is placed on feature maps, and the most significant value in the box is selected while pooling. This
2×2 matrix is passed over the entire feature map from left to right, choosing the maximum value
in each pass. These data are then combined to create a new matrix known as a pooled feature map.
Max pooling helps to keep the image’s key characteristics while shrinking its size. This helps avoid
CMES, 2023, vol.134, no.3 1519
overfitting, which occurs when the CNN is given too much data, especially if the information is
irrelevant to categorizing the picture.
After acquiring the pooled featured maps, they need to be flattened. The whole pooled feature
map matrix is flattened into a single column, then given to the neural network for processing.
5) Full connection
The flattened feature map is then sent through a neural network after f lattening. The input layer,
the fully linked layer, and the output layer make up this stage. In ANNs, the completely connected
layer is identical to the hidden layer, except it is fully linked in this case. The projected classes are the
output layer. The data is sent via the network, and the prediction error is computed. The mistake is
then sent back into the algorithm to enhance the prediction.
In most cases, the neural network’s final values do not add up to one. However, it is critical to
reducing these figures to integers between zero and one, which indicates each class’s likelihood. By
Eq. (17), the Softmax function plays this job,
for j =1, ..K(17)
3Challenges and Issues
The techniques above nevertheless have many drawbacks. First, rather than leveraging correlated
information across patches. Most existing pixel-wise tampering detectors employ a patch-based
approach. As a result, statistical information required for feature extraction is insufficient, particularly
at the boundary of a forged region. Alternatively, to make assessing the validity of an inquiry patch
simpler, the features of neighboring patches should be underlined.
Furthermore, the absence of statistical characteristics across flat sites (clear sky, Blue Ocean, etc.)
generates estimation uncertainty, resulting in poor detection performance. In this case, the texture of
the image content becomes an essential factor in enhancing detection accuracy. Furthermore, as image-
editing software has advanced rapidly, the leftovers of alteration operations now behave similarly to
the original (i.e., tampering traces are hard to detect). As a result, lowering the likelihood of detection
mismatches and increasing localization resolutions (determined by the smallest unit of detection)
remains an ongoing challenge.
Machine learning approaches that rely on block feature extraction, on the other hand, suffer from
the same flaws as block-based methods. Furthermore, to counter forgeries, these algorithms only learn
characteristics of a single recognized picture. Due to the lack of past information to handle forgeries
in other images, the techniques must re-initialize the model and repeat many times. These approaches
are significantly less efficient than deep learning methods.
4Conclusion and Future Work
In terms of simulation, this article has examined the numerous processes involved in forgery
detection approaches. It is clear from this paper that there would be two primary categories accessible
in forgery detection. Several authors rely heavily on the passive voice. The passive-based forgery
detection approaches are also discussed in this