I.J.

Image, Graphics and Signal Processing, 2024, 1, 38-60

Published Online on February 8, 2024 by MECS Press ([Link]
DOI: 10.5815/ijigsp.2024.01.04

Copy-Move Forgery Detection and Localization

Framework for Images Using Stationary Wavelet
Transform and Hybrid Dilated Adaptive VGG16
with Optimization Strategy
Prabhu Bevinamarad*
Department of Computer Science and Engineering, BLDEA’s V.P. Dr. P.G. Halakatti College of Engineering and
Technology (Affiliated to Visvesvaraya Technological University, Belagavi-590018, Karnataka), Vijayapura,
Karnataka 586103, India
E-mail: prabhubev@[Link]
ORCID iD: [Link]
*Corresponding Author

Prakash Unki
Department of Information Science and Engineering, BLDEA’s V.P. Dr. P.G. Halakatti College of Engineering and
Technology (Affiliated to Visvesvaraya Technological University, Belagavi-590018, Karnataka), Vijayapura,
Karnataka 586103, India
Email: prakashhunki@[Link]
ORCID iD: [Link]

Padmaraj Nidagundi
Computer Science and Information Technology, Riga Technical University, Riga, Latvia
Email: padmaraj.nidagundi_1@[Link]
ORCID iD: [Link]

Received: 24 July, 2023; Revised: 26 August, 2023; Accepted: 10 September, 2023; Published: 08 February, 2024

Abstract: Due to the availability of low-cost electronic devices and advanced image editing tools, changing the
semantic meaning of a particular image has become straightforward by employing various image manipulation
techniques like image copy-move, image splicing and removal operations. The tampered images with this sophisticated
software are rich in visualization, making the modifications invisible to the naked eye. Detecting these image alterations
is laborious, time-consuming, and often yields inappropriate results. The current techniques use conventional square,
slide regular, and artifacts procedures to identify image deviations to combat image forgery practices. Still, these
techniques exhibit problems related to generalization, training and testing, and model complexity. So, in this paper, a
novel image forgery detection and localization framework is implemented using stationary wavelet transform (SWT),
and a Hybrid Dilated Adaptive VGG16 model with optimization is introduced to classify forgery images and localize
the forgery regions present in an image. Initially, the proposed framework processes the input image with SWT to
decompose an image into different subband and further divide it into patches. After that, the hybrid dilated adaptive
VGG16 Network (HDA-VGG16Net) is built to extract the deep image features from the patches. Later, the Hybridized
Tuna Swarm with Bald Eagle Search Optimization (HTS-BESO) technique is applied to optimize the VGG16
parameters. Finally, feature matching is formed using multi-similarity searching to recognize whether the input image is
forged or original by locating forgery regions. The evaluation results are compared with existing forgery detection
approaches to ensure the efficiency of the developed model by considering multiple performance measures.

Index Terms: Copy-move forgery; Stationary wavelet transform; Dilated CNN; Tuna swarm optimization; Bald eagle
search optimization

Abbreviations
DWT Discrete Wavelet Transform
DCT Discrete Cosine Transform
SWT Stationary Wavelet Transform

This work is open access and licensed under the Creative Commons CC BY 4.0 License. Volume 16 (2024), Issue 1
 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

SVD Singular Value Decomposition

FMT Fourier-Mellin Transform
BRISK Binary Robust Invariant Scalable Keypoints
SURF Speeded-Up Robust Feature
SIFT Scale Invariant Features Transform
CNN Convolution Neural Network
LIPO Local Intensity Order Pattern
GAN Generative Adversarial Network
LSTM Long-Short Term Memory
PRNU Photo-Response Non-Uniformity
BDP Boundary to Pixel Direction
A-EHO Autoregressive Elephant Herding Optimization based Generative Adversarial Network

1. Introduction

Introducing the latest image technologies in digital processing and hardware capabilities caused many individuals
to apply the changes in the original images and tamper without leaving visual clues[1]. When produced as authentic
documents or evidence, the tampered images may cause various problems in the medical domain and courtrooms
believing that the photos tampered with using low-cost, sophisticated image editing tools will leave some visual clues
surrounding the tampered regions, several authentication techniques are employed to determine the authenticity of the
images. These authentication techniques are divided into two categories: active and passive[2]. In the former method,
the signature code is provided in picture capturing and used to test modified images. In this case, the multimedia
information has a digital mark on it[3]. As a result, modifications to multimedia content are easily detected. However,
only some image-capturing gadgets have a digital signature code. Therefore this approach is only sometimes
appropriate. Latter, also called forensics, do not require additional data related to the image. Thus, their objective is to
determine an image’s authenticity by studying it and looking for signs of any special processing it may have undergone.
The image forgeries[4] can be done in different ways, such as copy-move image forgery, in which copying and pasting
one or more regions of an image into the same image at various locations are performed to significantly alter the
semantic content of the target image and concealing the information or duplicating objects and people. In contrast to
copy-move forgeries, this one uses portions of other photos that have been taken out to create the pasted regions and
objects. To hide some content or provide a fictitious context, splicing forgeries might be performed. A portion or "hole"
in the image is filled with believable content in this type of attack. Usually, in-painting is used to repair damaged areas
in pictures. However, potential attackers may use it maliciously to remove a visible watermark or conceal information
from an image. Furthermore these tampering techniques use post processed by applying different operations such as
contrast enhancement, blurring, and compression to make it hard for the detection systems.
Researchers have made several efforts in recent years to combat forgery practices to detect image forgeries by
incorporating conventional forgery detection techniques and deep learning. The section-II will discuss the most current
and prominent approaches that have evolved during recent years
The proposed deep learning-based forgery detection mechanism comprises the following key contributions

1. The suggested approach employs state-of-the-art deep networks and parameter optimization mechanisms to
prevent image forgery and establish a cost-effective tool for forgery detection.
2. Using SWT for image decomposition helps to identify the most similar and discriminative image
characteristics. Also, optimizing parameters like start level, wavelet type, and norm using HTS-BESO
algorithm optimization technique enhance decomposition performance and improve the performance of the
Adaptive Forgery Identification and Localization Framework.
3. The Hybrid dilated VGG16 network model employs convolution kernels with varying dilation rates across its
layers. This design guarantees full coverage of a square area, eliminating gaps or missing information through
convolution operations. Consequently, the model efficiently captures the most pertinent features from image
patches, enhancing accuracy during training and testing while reducing time consumption.
4. We conducted a comparative analysis to assess the effectiveness of our proposed method, which utilizes the
HTS-BESO-HDA-VGG16 Net-based forgery detection scheme. We evaluated its performance against various
existing algorithms and recently developed techniques.

Here is a brief explanation of the paper’s reminder section. Section 2 covers the associated research for the forgery
detection techniques. Section 3 details the proposed adaptive forgery detection and localization framework. Section 4
discusses the experimental setup, evaluation parameters, results and discussions of the proposed forgery detection
method. Section 5 concludes the proposed HTS-BESO-HDA-VGG16Net-based forgery detection technique.

Volume 16 (2024), Issue 1 39

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

2. Literature Survey

Over the years, researchers have proposed various techniques for determining whether an input image has been
forged, emphasizing forged areas. According to the literature review, the first block-based method for identifying
forgery images was put forth in 2003[5]. They later developed various strategies using different feature extraction and
matching methodologies. They most recently applied deep learning models to enhance detection outcomes and lessen
the computational burden of detecting fake photos. Therefore this section summarizes the most recent methods based on
conventional feature extraction and matching techniques and the deep learning models developed in recent years.
2.1 Conventional forgery detection techniques
These methods are primarily classified as block-oriented and key-point-oriented. In the former, an entire image is
divided into irregular or regular overlapping shaped blocks and processed to extract significant features from each
image block, followed by some block matching algorithm to isolate the forged blocks. In the latter, correlated pixels are
discovered by removing the image’s key points of significance.
In the paper, [6] proposed a technique based on Gaussian-Hermite Moments (GHM), which divides an input image
into constant-sized overlapping blocks, and then the Gaussian-Hermite moments are extracted from each block. Later,
the lexicographical sorting and matching process is used to find similar blocks. The forgery detection technique
proposed in the paper[7] combines SURF and BRISK descriptors. A similar work in the paper[8] combines block-based
and keypoint-based techniques using FMT and SIFT for forgery detection. The paper[9] introduces a method where
features are extracted and compressed using a combination of DCT and SVD techniques. These features are then used
for training support vector machine (SVM), enabling the identification of forgery regions within an image by applying
the K-means machine learning technique. The work referenced in [10] identifies both types of forgeries, such as
splicing and copy-move simultaneously, by extracting features using DCT, DWT, and Ensemble classifier to classify
the input images as forged or authentic.
Further, it utilizes a key point-based method to detect the forgery regions. The paper in [11] presents a
comprehensive approach for investigating JPEG compressed test images that are suspected of being altered through
splicing or copy-move forgery and localizing the tampered region in forged images. The paper [12] proposes a
technique that includes double matching and region-localizing processes for filtering isolated keypoint pairs, finding
approximate suspicious parts, and localizing copy-move areas. The paper [13] introduces a frequency-domain image
manipulation method that utilizes DWT and inverse DWT to identify the region within the host image that will be
manipulated.
These techniques effectively locate the tampered areas even though the tampered regions are post-processed.
However, the algorithm did not do any further robustness tests, and the size of the feature vectors increased the
computational complexity. So, deep learning frameworks are evolved to overcome the difficulties identified in the
conventional forgery detection techniques to achieve higher classification and detection accuracy rates
2.2 Deep learning models
The prior research has looked at residual patterns, wavelet transform, statistical properties, image pixel information,
and other image attributes to verify the authenticity of an image. Due to the growing popularity of deep neural network
frameworks recently, some efforts have been made to incorporate CNN to design deep learning models to improve
forgery image detection. In paper,[14]the present paper investigates copy-move forgery detection with fusion
comprising of a deep convolutional and an adversarial model. The proposed algorithm enables constantly update its
learning via training data to differentiate areas from forged ones. The paper[15] presents a novel approach for detecting
and locating non-aligned JPEG forgery. The method utilizes a deep neural network to perform semantic pixel-wise
segmentation of JPEG blocks. The paper, [16]A high-confidence tampering localization structure is proposed by
utilizing Long Short-Term Memory (LSTM) cells and an encoder-decoder network to distinguish tampered regions
from un-tampered ones based on resembling features. In 2020, Lin and Li[17] investigated a segmentation-based
forgery detection technique. This method focuses on the local uniformity of visually hidden clues, which helps
overcome the limitations of existing segmentation methods that rely solely on visually perceptible content. Additionally,
they proposed a forgery localization method based on [Link] the paper[18], CNN is used to train hierarchical features
represented from an input image for identifying altered and original images. In the paper[19], proposed two approaches,
a model using a custom architecture and a model using transfer learning to distinguish between altered and original
images. The paper[20], proposes a method based on color illumination, deep CNN, and semantic segmentation is
employed to detect and localize image forgeries. The paper[21] introduces a deep learning approach with a dual-branch
CNN to detect passive copy-move forgery by extracting multi-scale features using various kernel sizes. An automated
deep learning-based fusion model is implemented in[22] by combining GANs and densely connected network
(DenseNets) models to create a layer for encoding the input vectors with the initial layer of the extreme learning
machines(ELM) classifier along with artificial fish swarm technique to adjust parameters to distinguish between the
input and target areas in a fake image. In 2022, Ganeshanet al.[23]Offered a GAN based on Autoregressive Elephant
Herding Optimization (A-EHO-based GAN) for copy move forgery detection. The paper [24]presents a lightweight

40 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

model constructed using mask R-CNN with MobileNet V1 for detecting and identifying copy move and image splicing
forgeries present in an image along with corresponding percentages. In 2022,Koulet al.[24], proposed convolutional
neural network with 3-Layered CNN algorithm for feature extraction and classification. Table 1 summarizes the
methodology, features and future challenges of the current forgery detection techniques.

Table 1. Summarization of methodology, features and challenges of previous forgery detection systems
Author
Methodology Features Challenges
[citation]
DWT,  Identify single and multiple tampered regions
 Complex post-processing attacks such as large
M. Bilal et al. [7] SURF and that are processed with post processing
scaling, smoothening, and brightness change.
BRISK operations.
 Capable of detecting forged areas scaled by a
K. B. Meena and FMT and  Unable to detect copy-move forgery regions
factor of 50% to 200% and subjected to JPEG
V. Tyagi[8] SIFT present in videos.
compression.
 Difficult in the detection of small-sized forged
regions.
G. S. Priyanka DCT and  Detect copy-move forgery and provides better
 Appropriate clustering by the K-means and
and K. Singh[9] SVD results against post-processed images.
selecting optimal threshold values pose
significant challenges.
 Can detect both spliced and copy-move
 Hard to detect images with varying
S. P. Jaiprakash DCT and forgeries simultaneously, exhibits high
resolutions, blurring attacks, and multiple
et al.[10] DWT accuracy across various image formats and
forgery regions.
demonstrates generalization capability.
 Precisely identify duplicated regions within  Experiences slightly higher rates of false
uncompressed, compressed, and doubly detections, mainly when there is a change in
Dua, J. Singh,
compressed images of various formats. illumination and blurring
and H. DCT
Parthasarathy[11]  Locate single and multiple manipulated areas  Extracting phase congruency features from
with precision, even when similar objects and multiple orientations in the covariance matrix
regions are present. poses a challenge.
 Precision decreases due to the extended
LIOP  Achieves high recall with acceptable triangles and increased number of keypoint
Q. Lyu et al.[12]
keypoints precision, and robustness. pairs resulting from the double-matching
process.
 Detect copy-move forgery regions using  The method fails to detect when the patch is
DWT and
T. Qazi, et al.[13] DWT and IDWT features not taken from the host image and high JPEG
Inverse DWT
compression.
Fusion of
Y. Abdalla et  Detect image with different sized forgeries
CNN and  Experiences more training time
al.[14] than the ones used during training.
GAN.
 Classify the different types tampering such as  Missing useful information during the splicing
Bappyet al.[16] CNN-LSTM object removal, copy-move and splicing and of large dimensional datasets and affected by
Handles the large dimensional dataset noise.
 It highly exploits the local homogeneity from  Need to investigate potential of image
X. Lin and C. T. indiscernible clues and able to detect and segmentation in other forensic detectors and
PRNU
Li [17] localize the object insert and object removal the combination of different techniques in
forgeries future work.
M. A. Elaskily et  Able to classify original and forgery images  There is a need to address the detection and
CNN
al[18] accurately within no time. localization of other digital image forgeries.
 There is a need for further exploration in
Custom expanding the training dataset, understanding
 Addressed the issue of generalization by
model and the influence of hyper-parameters on classifier
Y. Rodriguez- training the architecture with only one dataset
VGG16 with performance, and investigating hybrid
Ortega[19] and evaluated on a variety of datasets instead
transfer techniques that integrate deep learning-based
of training on an extensive dataset.
learning feature extraction with domain
transformation.
 Able to accurately detect multiple, rotated,  Detection of small altered regions needs to be
and scaling forgeries with large-scale scaling improved which exhibits sharp edges and
super‑ BPD forgeries and small manipulated image areas. background shadow.
N. Jindal[20]
segmentation  Incorporates local information features from
and deep the shallow network outputs and employs the  The SD-Net’s segmentation module and dual-
CNN Atrous spatial pyramid pooling (ASPP) layer branch structure have complicated the
to create a feature pyramid to improve the method.
overall performance.
 Forgery region detection is need to be
dual branch addressed in the model. It is important to
convolutional focus on improving the model's generalization
N. Goel et al.[21]  Classify the image is original or forged.
neural capabilities, particularly in terms of handling
network varying image sizes and different types of
forgeries.
 Unfortunately, training a GAN becomes
 Classify the input image is forged or original
N. Krishnaraj et GANs and challenging when the generator and
al.[22] DenseNets  Achieved higher training and validation
discriminator are highly proficient, as GANs
accuracy
typically require significant training time.

Volume 16 (2024), Issue 1 41

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

 It accurately captures the copied image portion

 Currently, the model does not address the
in the host image.
Ganeshanet A-EHO based detection of forgery regions, and achieving
al.[23] GAN  Employing the separation operator with clan
generalization of the model remains a
updating with regressive terms successfully
challenge.
achieves global convergence.
 It provides improved forgery detection
 It loses some information during merging.
Koulet al.[24] CNN accuracy than the other methods and alleviate
 It is affected by white Gaussian noise.
the derelictions.

2.3 Forgery detection challenges in existing systems

With widespread digital devices and freely available open-source and commercial image editing tools, image
modification poses a significant challenge in visual media. Due to this, an authenticity of an image becomes
questionable. Moreover, it can lead to misleading testimonies within a court of law and poses a challenge because
human eyes cannot discern alterations made to an original image due to its tiny nature. Also, the localization process is
crucial for locating the places that have been altered. As a result, several conventional and deep learning methods are
evolved to authenticate and detect forgery regions but exhibit several limitations. The preceding techniques carry a
more significant burden of computational complexity and yield imprecise outcomes when manipulated images are
subjected to a range of post-processing maneuvers, including substantial and subtle scaling, smoothening, and
brightness adjustments.
Furthermore, this approach needs to be revised in identifying instances where the altered patch originates from a
different source image, and it struggles in cases of high-level JPEG compression, rendering it unsuitable for real-time
applications. In contrast, deep learning models focus solely on discriminating between forged and authentic images.
Nonetheless, these models necessitate extensive training data and heightened computational capabilities. Despite these
requirements, achieving effective model generalization remains an ongoing challenge.

3. Proposed Adaptive Forgery Detection and Localization Framework

Although many conventional and deep learning approaches mentioned above have demonstrated promising results
in classifying images as forged or authentic and locating forgery regions, still, they have some drawbacks as mentioned
in section 2.3. Furthermore, these current methods are susceptible to fuzzy attacks and necessitate usage in a distinct
space, making distinguishing between innocent and malicious retouching difficult. Hence, they need to be revised,
including imprecise results, low resilience, and a high rate of false alarms. Therefore, we propose a model called HTS-
BESO-HDA-VGG16Net-based forgery detection that offers a solution to improve the training and testing performance
and generalize the model for detecting forgeries to overcome some of these constraints. The suggested model can be
used by various web applications, government agencies and social media platforms as the back end to verify the
legitimacy of each image data provided during transactions. Fig. 1 depicts the visual representation of suggested forgery
detection approach.

Fig. 1. Visual representation of the proposed forgery detection technique.

42 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

In the first step, we gathered the required images from two datasets. Then SWT is applied to decompose each
image by optimizing the parameters like start level, wavelet type and norm with the help of the suggested HTS-BESO
algorithm. Before the images are decomposed using SWT, all images are preprocessed to reduce the size to 400×400
[Link] decomposed images split into several patches. Later the significant image features are extracted from the
patch using hybrid dilated adaptive VGG16. The parameters from the VGG16, like hidden neuron count and epochs, are
optimized to maximize the accuracy and precision values. Finally, keeping one patch as a constant and a matching
process is performed to determine the similarity value for the remaining patches. If the similarity value exceeds 0.7, we
consider the image forged and mark the forgery regions in the image; otherwise, we believe it is authentic. Finally, the
final results are contrasted with the existing algorithms to evaluate the effectiveness of the offered HTS-BESO-HDA-
VGG16Net-based forgery detection scheme. Comprehensive explanations of each step are explained in the following
sections, encompassing all essential particulars.
3.1 Image Decomposition using Adaptive SWT

The input images collected from the datasets are indicate lncol
a , wherea  1, 2, 3, ....., A as the total number of images.
In the first step, each input image is decomposed by applying adaptive SWT[25]to divide an input image into a few
frequency bands. The choice of adaptive Stationary Wavelet Transform (SWT) for image decomposition is founded on
its ability to capture both high-frequency and low-frequency details within an image effectively. The adaptive SWT
adjusts the filter lengths depending on the scale of analysis, unlike conventional SWT, which employs preset filter
banks for decomposition. This versatility enables SWT to capture local image characteristics, improving the
representation of structures at different scales and ensuring that essential features are not lost throughout the
decomposition process. In the proposed approach, the SWT parameters such as norm, start level, decomposition level,
and wavelet type are optimized using HTS-BESO, enabling the retention of significant image details for extracting
discriminative features and achieving high forgery detection accuracy and precision. By incorporating the adaptive
SWT-based decomposition, we greatly enhance the detection reliability of the proposed model. The adaptive SWT
image decomposition process and corresponding mathematical expressions are depicted and illustrated in Fig. 2 and
equation (1-6), respectively.

dk ,l   y ( m) 'k ,l ( m) (1)

mX

Here, discrete wavelet is represented as  'k ,l (m) and its value is determined using equation (2).

'k ,l (m)  2( k / 2) 0,0 (2k (m  l )), (3)

db1.l (m)   i1(m   ) y( )

de1.l (m)   h1(m   ) y( ) (4)

The above equations are generalized and written in equation (5).

 
dbk ,l (m)   2k 1[i1 ]  dbk 1.l   i k (m   )dbk 1,l ( )
 
dek ,l (m)   2k 1[h1 ]  dek 1.l (l )   hk (m   )bek 1,l ( ) (5)

Here, dbk ,l and dek ,l are the approximate and detailed coefficients, respectively. These coefficients are generated
with the help of signal sequence and are represented as y(m) , i1 and j1 is the adaptive size of the high and low pass
filter respectively, over sampling of high pass and low pass filter at coefficient values of i k1 (m) and h k1 (m) is indicated
as  2k 1[i1 ]  i k (m) and  2k 1[h1 ]  hk (m) the expression is used to find the values and it is written in equation (6).

 j j 1( m )
h (2 m )  h

 h j (2 m 1) 0
(6)
 j j 1 ( m )
i (2 m ) i

 i j (2 m 1) 0

Volume 16 (2024), Issue 1 43

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

The above equation gives very accurate output signals by utilizing adaptive SWT method. The output obtained
from this process is denoted as DbSWT .

Fig. 2. Illustration of adaptive SWT-based image decomposition

3.2 Optimization Technique Adopted in Image Decomposition and VGG16

We have implemented the HTS-BESO algorithm in a proposed forgery detection scheme. This algorithm aims to
optimize the parameters of SWT and VGG16 to enhance the detection performance. The proposed algorithm
demonstrates an advantage over alternative swarm-based metaheuristic algorithms by effectively balancing exploitation
and exploration to conduct a thorough search and identify the best answer. It also ensures the capacity to converge
quickly to optimal or almost optimal solutions. The information-sharing and adaption mechanisms of the algorithm
allow it to focus on promising areas of the search space efficiently. In contrast, the BESO algorithm reduces costs and
provides high comfort.
Consequently, the HTS-BESO algorithm addresses the limitations of existing algorithms by significantly
improving accuracy and maximizing precision in detecting forgeries. To implement the suggested HTS-BESO
algorithm, we incorporate an updated uniform random number based on the fitness value. This allows us to obtain the
best solution by adjusting the random parameters. The key requirement is to update the candidates' positions, as
specified in equation (7).

Bf
if c  (7)
Wf

𝐵𝑓 and 𝑊𝑓 stand for the best and worst fitness values, respectively, and c is a uniform random number. The
conventional method’s c weight varies between [0-1]. The suggested strategy provides a highly optimized solution in
the problem space under the abovementioned criterion. The HTS-BESO algorithm also enhances the proposed system’s
rate of convergence.
 Tuna Swarm(TSO)
TSO[26] algorithm is designed based on individual movement and interaction principles within a swarm, enabling
it to tackle complex optimization problems effectively. The marine predatory fish known as tuna is officially known as
Thunnini. Tuna comes in a variety of species, and they come in a wide range of sizes. Tuna are top marine predators
that eat various surface and midwater fish. The fishtail form, which tunas use to swim, is a unique and effective

44 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

swimming style that keeps the body tight while allowing the long, thin tail to swing quickly. Tunas are continuous
swimmers. Despite swimming rapidly, the lone tuna takes longer to react than the nimble little fish. So the tuna will
hunt in groups, using the “group travel” strategy. To locate and catch their prey, they employ intellect. Numerous
valuable and clever foraging techniques have evolved in these species.
The first technique is Spiral foraging. To lure their prey into shallow water, where they may be more readily
attacked, tuna swim in a spiral shape when feeding. Parabolic foraging is the second tactic. Each tuna form a parabolic
shape as it swims behind the one before it to encircle its prey. The two approaches mentioned above are effective for
tuna foraging. Following is a description of the TSO's mathematical model. Like most swarm-based metaheuristics,
TSO initiates the optimization process by uniformly generating beginning populations at random in the search space
and it is given in equation (8).

Y Ij  rand (ua  lb), j  1, 2, 3, ........, MP (8)

The rand is used to represent a uniformly distributed random vector and has a range of [0-1]. The upper and lower
boundaries of MP are given as 𝑢𝑎 𝑎𝑛𝑑 𝑙𝑏, respectively. MP is the number of Tuna populations. The initial individual in
the j th direction is denoted by Y Ij J th ,
Sardines, herring, and other small schooling fish establish a thick configuration when they come into contact with
predators, which makes it harder for the predator to lock on a victim. The tuna school forms a tight spiral formation. It
begins chasing the prey at this point, even though most of the fish in the school has a poor sense of direction. When a
small number swims steadily in a particular sequence, the neighboring fish gradually alter their focus until they
eventually form a large group with the same aim and begin to hunt. Tuna schools communicate with one another in
addition to spiraling after their prey. Since each tuna follows the one before, information can be shared between nearby
tuna. Based on the above principles, the mathematical formula for the spiral foraging strategy is as follows: equation (9).

  u u 
 1 . Y premi   . Y premi Yiu   2 . Y u
j , j 1,
u 1   
Yj  (9)
  . Y u u u u
 1  premi   . Y premi Yi   2 . Y j 1 , j 2,3,..., MP
  

u
1  b  (1  b). ,
umax
u
2  (1b)  (1b). ,
umax
cm
 e .cos(2 c),
3cos((u max 1/ u )1) )
m e ,

Where Y ju 1 j th individual of the u  1 iteration, the weight coefficients 1 and  2 govern how inclined
people are to travel in the direction of the ideal person and the preceding person, respectively. Parameter b determines
how closely the tuna follow the ideal person and the preceding individual in the first phase. The current iteration is
u
denoted by u, while the u max represents the maximum iteration, the current optimal individual is indicated as Y premi ,
the parameter c denotes uniform random number and its ranges in between [0,1] .
All tuna can utilize the search space surrounding the food when they forage spirally around it. However,
unthinkingly following the ideal individual to feed is not advantageous for group foraging when that person cannot find
food. As a result, we consider creating a random coordinate in the search space as a starting point for the spiral search to
enable everyone to look for a larger area and offer TSO the power to explore the entire world. The details of the specific
mathematical model are as follows in equation (10).

Volume 16 (2024), Issue 1 45

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

  u u u 
 1 .  Yrand   . Yrand Yi 
  
   . Y u , j 1,
 2 (10)
Y j u 1  
j
  . Y u u
  . Yrand Yiu 

 1  rand 

  2 . Y u , j  2,3,..., MP
 j 1

u
As a result, the random generated reference point in the search space is denoted by Yrand .
Metaheuristic algorithms usually start by engaging in sizeable global exploration before progressively shifting to
focused local exploitation. As a result, TSO switches the spiral foraging reference points from random individuals to
optimal individuals with each subsequent iteration. The final mathematical model summarizes the spiral foraging
strategy illustrated in equation (11) as follows.

  u u u 
 1 .  Yrand   . Yrand Yi 
  
  . Y u , j 1,
 2 j

 . Y  u u 
  . Yrand Yiu 
 1  rand  u
 if rand 
  2 . Y u
j 1 , j  2,3,..., MP
u max (11)

Y j u 1   u
  . Y u u  if rand 
  . Y premi Yiu
 1  premi  u max


  2 . Y u , j 1
 j
  u u 
 1 .  Y premi   . Y premi Yiu 
  
 u
  2 . Y j 1 , j  2,3,..., MP


Tunas also engage in cooperative parabolic feeding in addition to spiral formation feeding. Tuna use food as a
point of reference to construct an illustrative structure, using their surroundings as a search area and hunting for food by
searching around themselves. To perform these two approaches simultaneously, we assume a selection probability of
50%. And we describe the specific mathematical model as in equation (12) as follows.

Y u u
 rand .(Y premi Y uj )
 premi
 if rand 0.5,
Y ju 1UG . q 2 .(Y premi
u Y uj ), (12)

 UG. q 2 .Y uj if rand 0.5,

(u / umax )
 u 
q 1 
 umax 

Here, UG is a random number and the value is between [1,1] .

Tuna uses two foraging techniques to locate their prey cooperatively. The initial population is generated at random
in the search space for the TSO optimization procedure. Each person determines whether to regenerate their location in
the search space based on probability ‘x’ or selects one of the two foraging strategies to use randomly in each iteration.
The value of the parameter ‘x’ is determined using parameter setting simulation experiments for parameter setting. All
TSO members are continually updated and calculated throughout the optimization procedure until the end condition is
satisfied. The ideal person is then given back along with the associated fitness value.
 Bald Eagle Search Optimization (BESO)
The BESO[27] algorithm, inspired by the hunting behaviors of bald eagles, captures the sequential nature of each
hunting phase. This algorithm can be divided into three stages: search space selection, exploration of the selected search
space, and decisive swooping.

46 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

Select stage: During the selection stage, bald eagles exhibit a strategic behavior of choosing the optimal hunting
location within a defined search region based on the abundance of available food. This behavior is mathematically
represented by equation (13).

SWnew ( j )SWbest  . rand .( SWmean SW ( j )) (13)

Where  is a control parameter for regulating positional changes and ranges between [1.5, 2] and rand is a random
number with a range of values between 0 and 1. The term SWbest refers to the area of the search that bald eagles have
chosen based on the best location they discovered during their previous quest. The SWnew ( j ) signifies new position and
Eagles randomly search all locations near the pre-selected search zone. While doing so, SWmean shows that these
eagles have consumed all of the information from the earlier points.
Search stage: Bald eagles maneuver in a spiral pattern within the chosen search zone during the search stage to
speed up their quest for prey. Equation 14 mathematically expresses the ideal swoop position.

SW ( j )  SW ( j )  o( j ).( SW ( j )  SW ( j  1))  n( j ).( SW ( j )  SW ) (14)

np m

Here, directional coordinates for j th position is signified by o , m parameters.

n rand ( j )
n( j )  ; nrand ( j )  rand ( j ). sin( ( j ))
max(| n rand |)
o rand ( j )
o( j )  ; o rand ( j )  rand ( j ). cos( ( j )) (15)
max(| o rand |)
 ( j )  b. . rand ; rand ( j )   ( j ). S. rand

S is a parameter that accepts a value between 0.5 and 2 to represent the number of search cycles, and b is a
parameter that takes a value between 5 and 10 to represent the corner between the point searches in the central point.
Swooping stage: Bald eagles swing to their intended prey in the swooping stage from the best location in the
search area. Every point moves in the direction of the ideal time. This behavior is represented quantitatively in equation
(16).

SWnp ( j )  rand . SWpremi  n1( j ).(SW ( j )  d1. SWm )  o1( j ) .(SW ( j )  d 2. SWpremi ) (16)

In this context, the random number is denoted as d1 and d2, respectively, while the directional coordinates are
indicated as n1 and o1. Finally, the values of the directional coordinates are calculated using equation (17).

n rand ( j )
n1( j )  ; nrand ( j )  rand ( j ). sin i ( ( j ))
max(| n rand |)
o rand ( j )
o1 ( j )  ; o rand ( j )  rand ( j ). cos i ( ( j )) (17)
max(| o rand |)
 ( j )  b. . rand ; rand ( j )   ( j )

Here, directional coordinates are denoted by m and o , respectively, and b is signifies a control parameter.
Pseudocode 1 provides a visual representation of the algorithmic steps employed in the proposed HTS-BESO method.

Pseudocode 1: HTS-BESO algorithmic steps

Set the initial population size as 𝑁𝑝
Set maximum iteration value 𝑈𝑚𝑎𝑥
Compute the value of 𝐵𝑓 and 𝑇𝑓 then
Calculate the value of uniform random number 𝐶
for t=1 to 𝑈𝑚𝑎𝑥
for j=1 to 𝑁𝑝
𝑖𝑓𝑐 < 𝐵𝑓 ⁄𝑊𝑓

Volume 16 (2024), Issue 1 47

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

𝑢
Discovery the value of 𝑌𝑟𝑎𝑛𝑑
Update the position using the TSO method mentioned in equation
(10).
Else
Find the value of best position
Update the position using BESO by equation (13)
end if
end for
end for

3.3 Block tilig process

Here, the decomposed image DbSWT is input to the block tiling process to divide an image into blocks. The image
block is a small rectangular box composed of several pixels to help the image processing algorithms easily carry the
feature extraction process to identify the similarity values in detecting the forged images with high accuracy and reduce
the computational complexity. Initially, we mark the seed points to divide an image into different blocks. After that, we
find the movement of this point using the slight gradient. We must increase the number of seed points to separate the
image into small patches. The decomposed images are split into nine non-overlapping patches of size 128 x 128denoted
as Pcsp .The developed forgery detection model employs a non-overlapping block tiling process to reduce the
redundancy and higher computational complexity during feature extraction and help to capture more global features and
textures in an image. Fine details can be captured from individual patches during the feature extraction process from
hybrid dilated VGG16 later. The smaller block sizes can capture finer details but might increase storage requirements
and introduce block artifacts.
3.4 Hybrid Dilated Adaptive VGG16 Network for Forgery Detection and Localization
In many traditional image classification algorithms, the conventional CNN plays a vital role. However, in these
traditional CNN models, increasing the network depth by stacking layers is often employed to achieve higher accuracy
and handle more complex scenarios. Unfortunately, this approach results in excessive computational resource
consumption and can lead to the vanishing gradient problem, where gradients become extremely small as they
propagate through the network layers. Consequently, the performance of the network may saturate or even deteriorate
significantly. To address these challenges, a dilated CNN model is introduced as an alternative to the convolution layers
of traditional CNN models. In this approach, dilated convolution layers, which include holes within the convolution
kernels, are utilized. This strategy helps reduce the computational resource requirements during feature extraction while
expanding the receptive field without increasing the number of parameters[28]. However, simply stacking dilated
convolution kernels can expedite the training process and improve training accuracy to some extent. Nevertheless, it
does not effectively enhance testing accuracy because dilated convolution kernels can result in the omission of specific
pixels, leading to the potential neglect of continuity information within the image. Additionally, when the size rate
remains constant, it becomes challenging to consider both large and small-scale information simultaneously while
extracting the image’s feature map.
Therefore, we have employed hybrid dilated convolution(HDC) layers[29] to process images and ensure that the
vital information won’t be lost to the greatest extent when extracting the feature map. In this process, the image patches
sp
Pc obtained from block tiling process are given as an input to HDC-VGG16Net to extract relevant image features and
incorporate the HTS-BESO for parameter optimization to efficiently detection the forgery images. The dilated
convolution kernels in the HDC model have different dilation rates in the various layers. Instead of using the same
dilation rate for all layers after down sampling, we use a different dilation rate for each layer. The dilation rates in the
dilated convolution kernels are set as 1, 2, and 5 in one set and 1, 2, and 5 in another set, respectively, to cover every
pixel point in the image and preserve critical information during a series of convolution operations—the Fig. 3
Illustrates dilated convolution kernel stacking effect that contributes to the calculation of the center pixel marked as a
red through three convolution layers with kernel size 3 x 3. (a) All convolution layers have a dilation rate r = 2. (b)
Subsequent convolution layers have dilation rates of r = 1, 2, 3, respectively. It can be seen that (a) always leaves some
holes between pixels, and these holes, with different dilation rates, every single step in the HDC model can fill the gaps.
Hence the top layer can access information from a broader range of pixels in the same region as the original
configuration. This process is repeated through all layers, thus making the receptive field unchanged at the top layer.
We constructed the HDC model with six dilated convolution-pooling modules, two fully connected layers, and a
softmax function. To prevent overfitting, we optimized the model using a two-layer dropout function.

48 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

Fig. 3. Dilated convolution kernel stacking effect[30]

Hence the designed HDC model effectively retrieves the most relevant features from the image patches, improves
accuracy in training and testing, and reduces time consumption. As a result, it maximizes accuracy and precision values,
thereby increasing the system’s overall effectiveness. We represent the objective function of the proposed deep
learning-based forgery detection scheme as equation (18).

 1 
F N1  arg min   (18)
  
A
 D PPV 
SW T T ,SW T L,SW T SL,SW T N ,VGG16 Hn,VGG16e

Thus, accuracy and precision is represented as D A , and PPV respectively, SW T T , SW T L, SW T SL, SW T N are the
optimized parameters of SWT and they are wavelet type and it is ranges between [0,4] , level and its value is ranges
between [5,15] , start level and the value is ranges between [1,4] , norm and its value is ranges between [0,1] and
VGG16 Hn, VGG16e are the optimized parameters of VGG16 and they are hidden neuron count its value is ranges
between [5,255] and epochs and its value is range in between [5,50] respectively. The value of accuracy and precision
also called as positive predictive value (PPV) is determined by using equation (23) and equation (24).The block
diagram of hybrid dilated adaptive VGG16-based forgery detection in Fig. 4.

Fig. 4. Block diagram of Hybrid dilated VGG16-based forgery detection process

Volume 16 (2024), Issue 1 49

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

3.5 Multi-Similarity-based Image Localization

We use a multi-similarity analysis technique to identify similar regions in images. When copy-move image
tampering occurs, the copied parts do not need to be precisely positioned where they should be, resulting in altered
blocks that do not intersect with the duplicated region. After splitting the image patches, we adopt a multi-similarity
analysis by keeping one patch constant and comparing the others to find similar image patches. If the similarity value
exceeds 0.7, we consider the image patch altered; otherwise, it belongs to the original image. So, this process allows us
to separate the forgery and original image patches precisely. Finally, to visualize the resulting image with the
discovered forgery region, we apply binary mapping techniques and additional morphological operations like opening
and closing to remove isolated image patches. Fig. 5 presents a Multi-similarity-based image localization process for
highlighting forgery regions.

Fig. 5. Multi-similarity-based image localization process

4. Experimental Setup

The proposed forgery detection model HTS-BESO-HDA-VGG16Net, was implemented using Python 3.8 with
Keras and TensorFlow as the backend toolkit for training and testing. The Keras default function is used to initialize the
parameters of all layers. We use the Adam optimizer with a learning rate 0.01 and the binary cross-entropy loss function.
The epoch and mini-batch sizes are set to 90 and 30, respectively. The proposed framework is tested on an Intel Core i7
64-bit processor with 16 GB RAM and an 8GB GPU. To assess the efficiency of the proposed method, we have set the

50 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

population and maximum iteration value as 10 and 50, respectively. To evaluate the proposed approach, we have used
two publicly available benchmark datasets employing various evaluation parameters and the results were compared
with those of existing algorithms. The comparative study was performed by considering the Jaya(JA)[31], Deer Hunting
Optimization (DHOA)[32], TSO[26], and BESO[27] algorithms and different deep learning techniques. The subsequent
subsection provides detailed information about the dataset, evaluation metrics and analysis of the results.
4.1 Dataset description and performance evaluation metrics employed
The presented approach for detecting forgeries utilizes two diverse benchmark datasets, CMFD[33]and
CoMoFoD[34], to perform an experiment to detect and localize forged content. These datasets consists of large image
samples with real-world scenarios and the diversity of forgeries encountered (plane image forgery with post-processing).
The CMFD, comprised of medium-sized BMP format images measuring 1000x700 or 700x1000 pixels, is divided into
different subsets (D0, D1, D2, and D3). D0 consists of 50 images that are exact translations of the original images and
corresponding binary masks indicating the source and destination areas of tampering. D1 contains tampered images
created by copy-pasting objects after rotation, while D2 comprises tampered images obtained through scaling. Finally,
D3 includes 50 original, untampered images. Hence, it encompasses 1,210 images of originals, tampered versions, and
tampered images subjected to post-processing attacks. On the other hand, the CoMoFoD dataset comprises 200 primary
images with a resolution of 512X512 pixels. These photographs are categorized based on four types of geometric
alterations: translation, rotation, scaling, and distortion. Each category consists of 40 pictures. Within each category, six
subcategories correspond to six different post-processing techniques that can be applied to an image: JPEG compression,
noise addition, image blurring, brightness alteration, color reduction, and contrast changes. Considering all geometric
alterations and post-processing methods, the CoMoFoD dataset encompasses 10,400 image samples. The
comprehensive description of these datasets are tabulated in the table as follows.

Table 2. Comprehensive description of these datasets

Dataset Post processing methods Parameters and corresponding values

Rotation (Range, step)= [((-25°, 25°), 5°),((0°, 360°),30°), ((-5°, 5°), 1°))]
CMFD
Scaling (range, step)=[((0.25, 2), 0.25), ((0.75, 1.25), 0.05)]
Post processing methods Parameters and corresponding values
Jpeg Compression Factors=[20,30,40,50,60,70,80,90,100]
Noise adding µ=0, σ2=[0.009, 0.005,0.0005]
CoMoFD
Image blurring Average filter=[3x3,5x5,7x7]
Brightness change (lower bound, upper bound)=[(0.01,0.95),(0.01,0.9),(0.01,0.8)]
Contrast adjustment (lower bound, upper bound)=[(0.01,0.95),(0.01,0.9),(0.01,0.8)]

The model is validated on an image level by considering both forgery images and post-processed with JPG
compression, image blurring, rotation and scaling from both datasets to practical applicability and generalization to
ensure whether or not the model can accurately classify forged and original images. Once the fake image has been
accurately identified, we use the block similarity matching algorithm to construct a bounding box around the affected
region to highlight it. The following assessment metrics stated in the equations (19-28) are formulated to assess the
proposed model’s effectiveness and compare it to other state-of-the-art methodologies.

Accuracy (DA)
The accuracy rate represents the ratio of accurately detected forged images among all the images in the dataset. It
is computed using,

T p T n
Accuracy ( D A)  (19)
T T n F p  F n
p

True positive rate (TPR)

This is defined as the ratio of correctly detected forged images. The TPR is calculated by,

Tp
TPR  (20)
T F n
p

True Negative rate (TNR)

It indicates the negative event rates i.e. it defines the rate of the identified forged images. The TNR is calculated by,

Volume 16 (2024), Issue 1 51

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

Tn
TNR  (21)
T n F n

Positive predictive value (PPV)

It defines the ratio of correctly forged images and total predicted forged images. The PPV rate indicates that the
detected forged images are true. It is calculated by

Tp
Pr ecision( PPV )  (22)
T F p
p

False positive rate (FPR)

It indicates the original images which are not detected as not forged. The FPR rate is calculated by,

Fp
FPR  (23)
F T n
p

False negative rate (FNR)

It indicates the missing detection rate of the forged images. i.e., the detection system failed to detect the forged
images is shown as FNR. Getting lower values on this metric is much better for improving the performance. The FNR is
calculated by,

Fn
FNR  (24)
F nT p

Negative Predictive Value (NPV)

Negative Predictive Value (NPV) is the proportion of −1 instances correctly classified by the ML classifier.

1
NPV  (25)
(1 PPV ) TPR TNR
1 X X
PPV (1TPR ) 1TNR
False discovery rate (FDR)
Defined as the number of images that are forged, but are identified as original, divided by the total number of
forged images. It is calculated by,

Fp
FPR  (26)
F p T p

F1-score
It integrates the PPV and TPR rates and gives the single value and it is given by,

 PPV *TPR 
F1  Score  2 X   (27)
 PPV TPR 

Matthews Correlation Coefficient (MCC)

It includes all the parameters and it is defined as the geometric mean of the regression coefficients of the problem
and it’s dual. It can be also formulated as follows,

p n
1 F F
MCC  T p T n (28)
( PPV *TPR*TNR*NPV )

Where the parameters T p, T n, F p, and F n signifies true positive, true negative, false positive and false negative.

52 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

4.2 Performance Evaluation and results discussions

Convergence analysis of proposed system is performed to optimize the training procedure and ensure model
achieves the best possible performance on unseen data. According to the analysis’s findings, the proposed deep
network-based forgery detection system delivers more accuracy and PPV because its cost function converges at the 15th
iteration 40% more than that of JA-DA-VGG16Net, 50% more than TSO-DA-VGG16Net, and 60% more than BESO-
DA-VGG16Net. The analysis’s findings indicate that the proposed method’s convergence rate is higher when compared
to the current techniques. Fig. 6 depicts the result of convergence analysis for both CMFD and CoMoFoD datasets
comparison with the existing methods. The performance analysis in terms of accuracy, FNR, PPV and MCC for both
datasets are depicted in Fig. 7 and Fig. 8 and Fig. 9 and Fig. 10, respectively, along with the comparison with existing
techniques. The comparison reveals that none of the current algorithms achieve satisfactory results in detecting
forgeries. The x-axis represents the linear stage, and the performance analysis considers the Linear, Sigmoid, Tanh, and
Relu activation functions, enabling clear visualization of the suggested model’s performance. Remarkably, the proposed
HTS-BESO-HDA-VGG16Net-based forgery detection scheme outperforms JA-DA-VGG16Net, DHOA-DA-
VGG16Net, TSO-DA-VGG16Net, and BESO-DA-VGG16Net by 9.60%, 6.36%, 5.40%, and 2.87%, respectively, at
the linear stage. Consequently, the output demonstrates that the proposed method achieves a high false negative rate
(FNR) for the CMFD dataset. On the other hand, the proposed forgery detection system surpasses that of JA-DA-
VGG16Net by 35.2%, TSO-DA-VGG16Net by 28.5%, and BESO-DA-VGG16Net by 14.34%. The analysis indicates
an increased convergence rate of the proposed method compared to the existing algorithms, as evidenced by the output.

(a) (b)
Fig. 6. Convergence analysis of proposed technique and existing methods for (a) CMFD and (b) CoMoFoD datasets

(a) (b)

Volume 16 (2024), Issue 1 53

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

(c) (d)
Fig. 7. Performance estimation of the proposed technique in terms of (a) FNR, (b) Accuracy, (c) PPV, (d) MCC with existing algorithms for CMFD
dataset

(a) (b)

(c) (d)
Fig. 8. Performance estimation of the proposed technique in terms of (a) FNR, (b) Accuracy, (c) PPV, (d) MCC with existing techniques for CMFD
dataset

54 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

(a) (b)

(c) (d)
Fig. 9. Performance estimation of the proposed technique in terms of (a) FNR, (b) Accuracy, (c) PPV, (d) MCC with existing algorithms for
CoMoFoD dataset

(a) (b)

Volume 16 (2024), Issue 1 55

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

(c) (d)
Fig. 10. Performance estimation of the proposed technique in terms of (a) FNR, (b) Accuracy, (c) PPV, (d) MCC with existing techniques for
CoMoFoD dataset

4.3 Overall analysis of the suggested model

The overall analysis of the proposed model is tabulated in Table 3 and Table 4. The accuracy and PPV values are
used to find out the fitness value. Fitness is the main parameter, and it is used to find the value of a random number. The
proposed HTS-BESO-HDA-VGG16Net-based forgery detection scheme obtained a TPR value of 8.08% more than JA,
6% more than DHOA, 4.24% more than TSO, and 1.84% more than BESO. Therefore, the analysis results show that the
proposed HTS-BESO-HDA-VGG16Net-based forgery detection scheme attains better than the presented algorithms.
Fig.12 and Fig. 13 depict the suggested forgery detection scheme results, showing forgery detection and the highlighted
regions. Fig. 12 includes the plane forgery image (rows 2 and 5) and the forgery image post-processed with rotation and
scaling (rows 1, 3 and 4). Fig. 13 includes the plane forgery image (rows 1 and 2) and the forgery image post-processed
with JPG compression and image blurring (rows 3 and 4). To comprehend the suggested model’s effectiveness, we
compared the proposed model and a pre-existing metaheuristic algorithm for image forgery detection. The results in
Table 5 and Fig. 11 indicate that the presented model outperforms the current metaheuristic optimization methods.
Table 3. Comparison of results obtained for proposed HTS-BESO-HDA-VGG16Net-based forgery detection scheme with existing algorithms for
target datasets
Algorithm Comparison
CMFD dataset CoMoFoD dataset
Assessment HTS- HTS-
JA[31] DHOA[32] TSO[26] BESO[27] JA[31] DHOA[32] TSO[26] BESO[27]
metrics BESO BESO
Accuracy 87.28261 88.58696 90.54348 92.5 95.0117 91.58621 93.24138 94.62069 96.27586 97.8753
TPR 86.9936 88.69936 90.1919 92.32409 94.0301 91.57303 93.25843 94.66292 96.34831 98.1219
TNR 87.58315 88.47007 90.90909 92.68293 93.0133 91.59892 93.22493 94.57995 96.20596 98.2881
PPV 87.93103 88.88889 91.16379 92.91845 94.3204 91.31653 92.9972 94.39776 96.07843 97.1771
FPR 12.41685 11.52993 9.090909 7.317073 6.97661 8.401084 6.775068 5.420054 3.794038 2.82001
FNR 13.0064 11.30064 9.808102 7.675906 4.87015 8.426966 6.741573 5.337079 3.651685 2.45309
NPV 86.62281 88.27434 89.91228 92.07048 94.7054 91.84783 93.47826 94.83696 96.46739 98.2213
FDR 12.06897 11.11111 8.836207 7.081545 4.65913 8.683473 7.002801 5.602241 3.921569 2.90212
F1-Score 87.45981 88.79402 90.67524 92.62032 95.1211 91.4446 93.12763 94.53015 96.21318 97.8841
MCC 0.745653 0.771663 0.810885 0.84998 0.89019 0.831682 0.864794 0.892388 0.9255 0.95557

Table 4. Classification results comparison of proposed HTS-BESO-HDA-VGG16Net-based forgery detection scheme with existing techniques for
target datasets
Classifier Comparison
CMFD dataset CoMoFoD dataset
Assessment TCN- HTS- TCN- HTS-
CNN[35] LSTM[36] TCN[37] CNN[35] LSTM[36] TCN[37]
metrics LSTM[36] BESO LSTM[36] BESO
Accuracy 87.93478 89.13043 91.19565 93.36957 95.0114 91.72414 92.96552 94.2069 95.31034 97.6773
TPR 88.0597 88.91258 91.04478 93.60341 95.0275 91.57303 92.69663 94.10112 95.22472 97.5791
TNR 87.80488 89.35698 91.35255 93.12639 94.0113 91.86992 93.22493 94.30894 95.39295 97.2797
PPV 88.24786 89.67742 91.6309 93.40426 94.1307 91.57303 92.95775 94.10112 95.22472 97.1889
FPR 12.19512 10.64302 8.64745 6.873614 5.98219 8.130081 6.775068 5.691057 4.607046 2.72027
FNR 11.9403 11.08742 8.955224 6.396588 4.77019 8.426966 7.303371 5.898876 4.775281 2.52809
NPV 87.61062 88.57143 90.7489 93.33333 94.1071 91.86992 92.97297 94.30894 95.39295 98.5543

56 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

FDR 11.75214 10.32258 8.369099 6.595745 4.86923 8.426966 7.042254 5.898876 4.775281 2.90132
F1-Score 88.15368 89.29336 91.3369 93.50373 94.1202 91.57303 92.827 94.10112 95.22472 98.3252
MCC 0.758615 0.782592 0.823886 0.867337 0.8703 0.83443 0.859261 0.884101 0.906177 0.95327

Table 5. Comparison of results with existing metaheuristic algorithms for image forgery detection
Existing techniques Accuracy TPR TNR
Cristin R et al. [38] 95.1 95.1 93.8
S. Uma and P. D.
96.0 94.2 100.0
Sathya[39]
C. B and P. V.
91.9 92.6 92.8
Bhaskar Reddy [40]
Proposed 97.6 97.5 97.2

100 Accuracy TPR TNR

Assessment metrics

90
80
70
60
50
40
30
20
10
Cristin R et al. [37] S. Uma and P. D. C. B and P. V. Proposed
Sathya [38] Bhaskar Reddy [39]
Existing techniques

Fig. 11. Comparison of results with existing metaheuristic algorithm for image forgery detection

Forgery localization by
Sample # Original Image Forged images Ground truth image
Proposed model

(a) (b) (C) (d)

Fig. 12. Visualization of forgery detection results obtained using proposed technique for CMFD dataset

Volume 16 (2024), Issue 1 57

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

Forgery localization by
Sample # Original Image Forged images Ground truth image
Proposed model

(a) (b) (C) (d)

Fig. 13. Visualization of forgery detection results obtained using proposed technique for CoMoFoD dataset

5. Conclusion

This article has presented a deep learning-based forgery detection and localization model. The proposed technique
encompasses the fusion of HTS-BESO-HDA-VGG16Net modules. Initially, the input images are selected from two
datasets and decomposed with the help of the SWT technique. The SWT and VGG16 parameters are optimized to
reduce the model complexity. The proposed model uses dilated convolution kernels with different dilation rates in the
various layers VGG16 to ensure complete coverage of a square area without any holes or missing information during
convolution operations. As a result, it maximizes accuracy and precision values, thereby increasing the system’s overall
effectiveness. The proposed model is evaluated using two standard datasets: CMFD and CoMOFoD. As per the results
discussed in Section 5, the proposed methodology outperforms with a TPR value of 8.08% JA, 6% than DHOA, 4.24%
than TSO, and 1.84% than BESO. Therefore the analysis shows that the HTS-BESO-based forgery detection scheme
achieves maximized accuracy and precision value compared to the existing algorithms. The proposed technique also
addressed the post-processing operations (scaling and rotation). However, the forgery attacks with a blend of multiple
post-processing operations, such as JPEG compression, Noise, Texture, and a combination of these, can hide forgery
clues. In this instance, the forgery detection significantly increases its complexity and difficulty. In the future, we
extend this work to overcome all these complexities.

References

[1] “Photo tampering throughout history,” 2004. [Link] (accessed Sep. 25, 2020).
[2] M. Zanardelli, F. Guerrini, R. Leonardi, and N. Adami, “Image forgery detection: a survey of recent deep-learning approaches,”
Multimed. Tools Appl., pp. 17521–17566, 2022, doi: 10.1007/s11042-022-13797-w.
[3] M. K. Arnold, M. Schmucker, and S. D. Wolthusen, “Techniques and applications of digital watermarking and content
protection,” Ann. Phys. (N. Y)., vol. 54, no. 2, 2003, [Online]. Available: [Link]
avail/%5Cn[Link]
=Techniques+and+Applications+of+Digital+Watermarking+and+Content+Protection&amp;ots=Pz_DSY_0e-
&amp;sig=2gsD5uG.
[4] A. Piva, “An Overview on Image Forensics,” ISRN Signal Process., vol. 2013, pp. 1–22, 2013, doi: 10.1155/2013/496701.
[5] A. C. Popescu and H. Farid, “Exposing Digital Forgeries by Detecting Duplicated Image Regions,” Tech. Report, TR2004-515,
Dep. Comput. Sci. Dartmouth Coll. Hanover, New Hampsh., no. 2000, pp. 1–11, 2004, [Online]. Available:
[Link]
[6] K. B. Meena and V. Tyagi, “A copy-move image forgery detection technique based on Gaussian-Hermite moments,” Multimed.
Tools Appl., vol. 78, no. 23, pp. 33505–33526, Dec. 2019, doi: 10.1007/s11042-019-08082-2.

58 Volume 16 (2024), Issue 1

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

[7] M. Bilal, H. A. Habib, Z. Mehmood, T. Saba, and M. Rashid, “Single and Multiple Copy–Move Forgery Detection and
Localization in Digital Images Based on the Sparsely Encoded Distinctive Features and DBSCAN Clustering,” Arab. J. Sci.
Eng., vol. 45, no. 4, pp. 2975–2992, 2020, doi: 10.1007/s13369-019-04238-2.
[8] K. B. Meena and V. Tyagi, “A copy-move image forgery detection technique based on tetrolet transform,” J. Inf. Secur. Appl.,
vol. 52, Jun. 2020, doi: 10.1016/[Link].2020.102481.
[9] G. S. Priyanka and K. Singh, “An improved block based copy-move forgery detection technique,” Multimed. Tools Appl., vol.
79, pp. 13011–13035, 2020.
[10] S. P. Jaiprakash, M. B. Desai, C. S. Prakash, V. H. Mistry, and K. L. Radadiya, “Low dimensional DCT and DWT feature
based model for detection of image splicing and copy-move forgery,” Multimed. Tools Appl., vol. 79, no. 39–40, pp. 29977–
30005, 2020, doi: 10.1007/s11042-020-09415-2.
[11] S. Dua, J. Singh, and H. Parthasarathy, “Detection and localization of forgery using statistics of DCT and Fourier components,”
Signal Process. Image Commun., vol. 82, no. July 2019, p. 115778, 2020, doi: 10.1016/[Link].2020.115778.
[12] Q. Lyu, J. Luo, K. Liu, X. Yin, J. Liu, and W. Lu, “Copy Move Forgery Detection based on double matching,” J. Vis. Commun.
Image Represent., vol. 76, no. September 2019, p. 103057, 2021, doi: 10.1016/[Link].2021.103057.
[13] T. Qazi, M. Ali, K. Hayat, and B. Magnier, “Seamless Copy–Move Replication in Digital Images,” J. Imaging, vol. 8, no. 3, pp.
1–15, 2022, doi: 10.3390/jimaging8030069.
[14] Y. Abdalla, M. Tariq Iqbal, and M. Shehata, “Copy-move forgery detection and localization using a generative adversarial
network and convolutional neural-network,” Inf., vol. 10, no. 9, 2019, doi: 10.3390/info10090286.
[15] N. Alipour and A. Behrad, “Semantic segmentation of JPEG blocks using a deep CNN for non-aligned JPEG forgery detection
and localization Content courtesy of Springer Nature , terms of use apply . Rights reserved . Content courtesy of Springer
Nature , terms of use apply . Rights re,” pp. 8249–8265, 2020.
[16] J. H. Bappy, C. Simons, L. Nataraj, B. S. Manjunath, and A. K. Roy-Chowdhury, “Hybrid LSTM and Encoder-Decoder
Architecture for Detection of Image Forgeries,” IEEE Trans. Image Process., vol. 28, no. 7, pp. 3286–3300, 2019, doi:
10.1109/TIP.2019.2895466.
[17] X. Lin and C. T. Li, “PRNU-Based Content Forgery Localization Augmented with Image Segmentation,” IEEE Access, vol. 8,
no. 2, pp. 222645–222659, 2020, doi: 10.1109/ACCESS.2020.3042780.
[18] M. A. Elaskily et al., “A novel deep learning framework for copy-moveforgery detection in images,” Multimed. Tools Appl.,
vol. 79, no. 27–28, pp. 19167–19192, 2020, doi: 10.1007/s11042-020-08751-7.
[19] Y. Rodriguez-Ortega, D. M. Ballesteros, and D. Renza, “Copy-move forgery detection (Cmfd) using deep learning for image
and video forensics,” J. Imaging, vol. 7, no. 3, 2021, doi: 10.3390/jimaging7030059.
[20] N. Jindal, “Copy move and splicing forgery detection using deep convolution neural network , and semantic segmentation
Content courtesy of Springer Nature , terms of use apply . Rights reserved . Content courtesy of Springer Nature , terms of use
apply . Rights reser,” pp. 3571–3599, 2021.
[21] N. Goel, S. Kaur, and R. Bala, “Dual branch convolutional neural network for copy move forgery detection,” IET Image
Process., vol. 15, no. 3, pp. 656–665, 2021, doi: 10.1049/ipr2.12051.
[22] N. Krishnaraj, B. Sivakumar, R. Kuppusamy, Y. Teekaraman, and A. R. Thelkar, “Design of Automated Deep Learning-Based
Fusion Model for Copy-Move Image Forgery Detection,” Comput. Intell. Neurosci., vol. 2022, 2022, doi:
10.1155/2022/8501738.
[23] R. Ganeshan, S. Muppidi, D. R. Thirupurasundari, and B. S. Kumar, “Autoregressive-Elephant Herding Optimization based
Generative Adversarial Network for copy-move forgery detection with Interval type-2 fuzzy clustering,” Signal Process. Image
Commun., vol. 108, no. November 2021, p. 116756, 2022, doi: 10.1016/[Link].2022.116756.
[24] S. Koul and M. Kumar, “An efficient approach for copy-move image forgery detection using convolution neural network
Content courtesy of Springer Nature , terms of use apply . Rights reserved . Content courtesy of Springer Nature , terms of use
apply . Rights reserved .,” pp. 11259–11277, 2022.
[25] G. P. Nason and B. W. Silverman, “The Stationary Wavelet Transform and some Statistical Applications,” pp. 281–299, 1995,
doi: 10.1007/978-1-4612-2544-7_17.
[26] L. Xie, T. Han, H. Zhou, Z. R. Zhang, B. Han, and A. Tang, “Tuna Swarm Optimization: A Novel Swarm-Based Metaheuristic
Algorithm for Global Optimization,” Comput. Intell. Neurosci., vol. 2021, 2021, doi: 10.1155/2021/9210050.
[27] S. Ferahtia, H. Rezk, M. A. Abdelkareem, and A. G. Olabi, “Optimal techno-economic energy management strategy for
building’s microgrids based bald eagle search optimization algorithm,” Appl. Energy, vol. 306, no. PB, p. 118069, 2022, doi:
10.1016/[Link].2021.118069.
[28] Y. Zhou et al., “Remote sensing scene classification based on rotation-invariant feature learning and joint decision making,”
Eurasip J. Image Video Process., vol. 2019, no. 1, 2019, doi: 10.1186/s13640-018-0398-z.
[29] X. Lei, H. Pan, and X. Huang, “A dilated cnn model for image classification,” IEEE Access, vol. 7, pp. 124087–124095, 2019,
doi: 10.1109/ACCESS.2019.2927169.
[30] P. Wang et al., “Understanding Convolution for Semantic Segmentation,” Proc. - 2018 IEEE Winter Conf. Appl. Comput.
Vision, WACV 2018, vol. 2018-Janua, no. March, pp. 1451–1460, 2018, doi: 10.1109/WACV.2018.00163.
[31] R. A. Zitar, M. A. Al-Betar, M. A. Awadallah, I. A. Doush, and K. Assaleh, “An Intensive and Comprehensive Overview of
JAYA Algorithm, its Versions and Applications,” Arch. Comput. Methods Eng., vol. 29, no. 2, pp. 763–792, 2022, doi:
10.1007/s11831-021-09585-8.
[32] G. Brammya, S. Praveena, N. S. Ninu Preetha, R. Ramya, B. R. Rajakumar, and D. Binu, “Deer Hunting Optimization
Algorithm: A New Nature-Inspired Meta-heuristic Paradigm,” Comput. J., vol. 133, no. 1, p. bxy133, 2019, doi:
10.1093/comjnl/bxy133.
[33] E. Ardizzone, A. Bruno, and G. Mazzola, “Copy-Move Forgery Detection by Matching Triangles of Keypoints,” IEEE Trans.
Inf. Forensics Secur., vol. 10, no. 10, pp. 2084–2094, 2015, doi: 10.1109/TIFS.2015.2445742.
[34] D. Tralic, I. Zupancic, S. Grgic, and M. Grgic, “CoMoFoD - New database for copy-move forgery detection,” Proc. Elmar - Int.
Symp. Electron. Mar., no. September, pp. 49–54, 2013.

Volume 16 (2024), Issue 1 59

 Copy-Move Forgery Detection and Localization Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated
Adaptive VGG16 with Optimization Strategy

[35] Q. Yang, D. Yu, Z. Zhang, Y. Yao, and L. Chen, “Spatiotemporal Trident Networks: Detection and Localization of Object
Removal Tampering in Video Passive Forensics,” IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 10, pp. 4131–4144,
2021, doi: 10.1109/TCSVT.2020.3046240.
[36] A. R. Gu, J. H. Nam, and S. C. Lee, “FBI-Net: Frequency-Based Image Forgery Localization via Multitask Learning with Self-
Attention,” IEEE Access, vol. 10, pp. 62751–62762, 2022, doi: 10.1109/ACCESS.2022.3182024.
[37] P. Hewage et al., “Temporal convolutional neural (TCN) network for an effective weather forecasting using time-series data
from the local weather station,” Soft Comput., vol. 24, no. 21, pp. 16453–16482, 2020, doi: 10.1007/s00500-020-04954-0.
[38] Cristin R, G. M. N.R, R. L, and V. S, “Image Forgery Detection Using Back Propagation Neural Network Model and Particle
Swarm Optimization Algorithm,” Multimed. Res., vol. 3, pp. 21–32, 2020.
[39] S. Uma and P. D. Sathya, “Copy-move forgery detection of digital images using football game optimization,” Aust. J. Forensic
Sci., vol. 54, no. 2, pp. 258–279, 2022, doi: 10.1080/00450618.2020.1811376.
[40] C. B and P. V. Bhaskar Reddy, “An approach for copy-move image multiple forgery detection based on an optimized pre-
trained deep learning model,” Knowledge-Based Syst., vol. 269, p. 110508, 2023, doi: 10.1016/[Link].2023.110508.

Authors’ Profiles

Mr. Prabhu Bevinamarad completed his B.E. in Information Science and Engineering and [Link]. Computer
Science and Engineering from Visvesvaraya Technological University, Belgaum, Karnataka, India, in 2008 and
2013, respectively. Presently he is pursuing a [Link]. in Computer Science and Engineering from Visvesvaraya
Technological University, Belgaum, and Karnataka, India. His research interests include Deep learning, digital
image and audio forensics.

Dr. Prakash H. Unki completed his B.E. in Electronics & Communication form Karnataka University,
Dharawad and [Link]. & Ph.D. in Computer Science and Information Sciences from Visvesvaraya
Technological University, Belgaum, Karnataka, India. He is working as Professor in Department of Information
Science and Engineering, B.L.D.E.A’s V. P. Dr. P. G. H. College of Engineering and Technology, Vijayapur.
He is having teaching experience of 26 years in an engineering college. He published 21 papers in various
international journals and conferences. Currently he is guiding two doctoral students. His research interests
include Digital Image Processing, Pattern Recognition, Machine learning, Cloud Computing, Data Science and
Big data. He is member of different professional societies.

Dr. Padmaraj Nidagundi completed a Bachelor’s degree in Information Science and Engineering from
Visvesvaraya Technological University, a Master’s degree in Computer Engineering and a Ph.D. from Riga
Technical University. Since 2010, he has been working at different software development companies. He is
currently a docent at Riga Technical University. His most successful field of activity in previous years has been
software development and testing, where he gained significant international work experience. In recent years, he
has been a teaching staff member and participated in research projects. His research interests are software
development, quantum computing, cybersecurity, and artificial intelligence.

How to cite this paper: Prabhu Bevinamarad, Prakash Unki, Padmaraj Nidagundi, "Copy-Move Forgery Detection and Localization
Framework for Images Using Stationary Wavelet Transform and Hybrid Dilated Adaptive VGG16 with Optimization Strategy",
International Journal of Image, Graphics and Signal Processing(IJIGSP), Vol.16, No.1, pp. 38-60, 2024.
DOI:10.5815/ijigsp.2024.01.04

60 Volume 16 (2024), Issue 1