Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Contents lists available at ScienceDirect

Journal of King Saud University –

Computer and Information Sciences
journal homepage: www.sciencedirect.com

A Hybrid robust watermarking system based on discrete cosine

transform, discrete wavelet transform, and singular value decomposition
Mahbuba Begum a,⇑,1, Jannatul Ferdush b,2, Mohammad Shorif Uddin c,3
a
Department of Computer Science and Engineering, Mawlana Bhashani Science and Technology University, Tangail 1902, Bangladesh
b
Department of Computer Science and Engineering, Jashore University of Science and Technology, Jashore 7408, Bangladesh
c
Department of Computer Science and Engineering, Jahangirnagar University, Dhaka 1342, Bangladesh

a r t i c l e i n f o a b s t r a c t

Article history: Digital data sharing is increased day by day due to the rapid access of the Internet by mass people.
Received 22 January 2021 Multimedia, for example text, image, video, and audio can be easily accessed by unauthorized users.
Revised 2 July 2021 To ensure multimedia security, researchers are focusing on digital image watermarking. A robust water-
Accepted 15 July 2021
marking system must satisfy the basic design requirements like imperceptibility, robustness, security,
Available online 24 July 2021
and payload capacity. Watermarking can be embedded into a host image through different transform
operations, such as discrete cosine transform (DCT), discrete wavelet transform (DWT), and singular
Keywords:
value decomposition (SVD). However, a single transformation does not ensure all the design require-
DCT
DWT
ments simultaneously. To fill this gap, a hybrid blind digital image watermarking with a combination
SVD of DCT, DWT, and SVD is proposed in this paper. At first, the Arnold map is used to encrypt the watermark
Robustness image. Then DCT is applied to it and to the host image followed by DWT before SVD. After that, the water-
Arnold Map marked image is generated by embedding the watermark image into the host image. The performance is
evaluated under various attacks and experimental results show that our algorithm ensures improved
robustness while preserving high imperceptibility and enhanced security than the state-of-the-art
methods.
Ó 2021 The Authors. Published by Elsevier B.V. on behalf of King Saud University. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction niques that has been attracted by researchers for researching on

this domain. This technology embeds (inserts) a watermark (own-
Digital content, especially digital images have been attracted a ership information) into the host image and generates the water-
lot of attention due to its excessive use of multimedia technologies marked image. Later, the system extracts the watermark image
and the Internet. Digital image alteration, replacing, regeneration, from the generated watermarked image. Based on the embedding
and distribution are very easy with a very low cost and urgent system, the watermarking system can be classified as spatial
delivery with the Internet and multimedia technologies. Data domain and transform domain methods.
authentication protects the accuracy and consistency of the multi- The general framework of the watermark embedding process
media element, which means the integrity (D’Angelo et al., 2010). consists of the original host image and the watermark image. The
Digital image watermarking is a one kind of authentication tech- watermark image is inserted into the host image by using a water-
mark embedding algorithm, and thus, the watermarked image is
⇑ Corresponding author. generated that is sent through the communication network and
E-mail addresses: [email protected] (M. Begum), jannatulferdush@just.
affected by various attacks. Then, the system extracts the water-
edu.bd (J. Ferdush), [email protected] (M.S. Uddin). mark image from the watermarked image with the help of the
1
ORCID: 0000-0003-1206-3669. watermark extraction algorithm. The above-discussed framework
2
ORCID: 0000-0002-1676-4422. is shown in the following Fig. 1.
3
ORCID: 0000-0002-7184-2809. Imperceptibility, robustness, capacity, and security are the four
Peer review under responsibility of King Saud University.
design features of a digital image watermarking system (Zhao
et al., 2010). But, due to their limited and conflicting characteris-
tics, these requirements are not possible to meet simultaneously
(Tao et al., 2014). The balance among imperceptibility, robustness,
Production and hosting by Elsevier
and capacity normally depend on the application (Giakoumaki

https://doi.org/10.1016/j.jksuci.2021.07.012
1319-1578/Ó 2021 The Authors. Published by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Fig. 1. General outline for insertion and extraction of watermark image.

et al., 2006; Lei et al., 2012, 2013, 2014, 2015). Besides, the existing 2018). On the other hand, SVD preserves the host image quality
literature revealed that spatial domain methods are less robust and also ensures robustness against noise, sharpening, cropping,
compared to transform domain methods (Wang et al., 2001; filtering, and JPEG compression attacks. It also solves the FPD prob-
Mukherjee et al., 2004). Spatial domain methods directly work lem of the watermark image (Chang et al., 2005; Vaishnavia and
with the image pixel values. Hence, the watermark can be embed- Subashini, 2014; Ali et al., 2016). Hence the hybrid transform
ded into the host image by manipulating these pixel values domain methods take the advantages of each transformation and
(Olanrewaju, 2011). These spatial domain techniques are suitable meet the basic design requirements simultaneously.
only for a noise-free image. The watermark image can be excluded Our contributions are as follows:
by cropping attacks which pose a major drawback of the spatial
domain watermarking methods. The watermark embedding capac-
We implement a hybrid method that ensures improved imper-
ity to the host image is high. But, the inserted watermark image ceptibility, robustness, and security simultaneously.
can be easily detected by several attacks (Celik et al., 2002;
We encrypt the watermark image with Arnold map for ensuring
Mukherjee, 2004; Nikolaidis and Pitas, 1998). Although spatial security.
domain techniques are simple and require less implementation
We embed an invisible image as watermark to the host image.
time, but they are less imperceptible and robust against various
We design a blind watermarking system so that the host image
attacks. Because of the above-mentioned limitations of the spatial is not required to extract the watermark image from the water-
domain techniques, researchers are now concentrated on the marked image.
transform domain methods. The transform domain techniques
work with signals rather than time. The images are converted to The rest of this research is organized as follows: Section 2
the frequency domain by using the predefined transformation. describes the existing literature, and Section 3 describes the theo-
Then, the watermark image is embedded into the original host retical background. Our proposed methodology is described in Sec-
image by altering the transform domain coefficients of the host tion 4. The experimental results with comparison are described in
image through various transformations like DCT, discrete Fourier Section 5, and finally, the last section concludes the study along
transform (DFT), DWT, and SVD. Finally, the watermark image is with a new research direction.
extracted by using the inverse transformation (Begum and Uddin,
2020). The conventional spatial domain methods have a low peak
signal-to-noise ratio (PSNR) values than the transform domain 2. Related literature
methods (Megalingam et al., 2010). Transform domain methods
ensure enhanced security, better imperceptibility, and high robust- The transform domain algorithms have improved performance
ness against various attacks. But a single spatial or transform than the spatial domain algorithms. It is very important to main-
domain method does not ensure all the basic design requirements tain a trade-off among the design requirements simultaneously.
simultaneously. For addressing this issue, the concept of hybrid Now, the hybrid domain methods are the appropriate domain for
transform domain methods has been arisen. maintaining this trade-off by combining two or more transform
Hybrid domain methods are usually considered as a combina- domain algorithms (Begum and Uddin, 2020). The DWT-based
tion of these two domains and designed to ensure enhanced algorithm works both for the time and frequency domain, but in
robustness along with better imperceptibility with high security some cases, DWT provides unsatisfactory results. For overcoming
(Begum and Uddin, 2020). Also, hybrid transform domain methods the traditional limitations of the wavelet-based watermarking
combine more than one transformation for improving the perfor- algorithm, a hybrid method is designed based on lifting wavelet
mance of the system. For example, DCT ensures better impercepti- transform (LWT), DCT, and SVD (Takore et al., 2018). At this point,
bility and is robust against noise, filtering, JPEG compression, and the Canny edge detector is used for selecting the suitable region of
geometric attacks. Also, DCT embeds more watermark bits to the binary watermark embedding. For maintaining the trade-off
host image. This technique also solves the false positive detection among imperceptibility and robustness, multiple scaling factor
(FPD) problem of the watermark image (Pun, 2009; Loani et al., (MSF) has been used during watermark embedding and particle
2018; Patra et al., 2010; Laouamer and Tayan, 2015; Singh and swarm optimization (PSO) has been used to achieve optimized
Bhatnagar, 2018). Besides, DWT ensures high imperceptibility MSF. But the method is not robust against print/scan, hybrid, and
and robustness against JPEG compression, salt-and-pepper noise desynchronization attacks. The method does not consider the secu-
(SPN), Gaussian noise (GN), and statistical attacks (Chen et al., rity of the watermark image. Also, the watermark embedding
2015; Haribabu et al., 2016; Jia et al., 2017; Hannoun et al., capacity is not calculated. Imperceptibility and robustness are
5857
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

the two significant requirements for copyright protection applica- (RDWT), and SVD is proposed for ensuring imperceptibility,
tions. With this view, Lei et al. (2019) proposed a multipurpose robustness, and security. The two dimensional (2D) logistic chaotic
watermarking method based on integer DWT (IDWT) and chaotic map provides the enhanced security to the generated watermarked
map for protecting copyright information and content authentica- medical image. The method is imperceptible, robust, secure, and
tion. For inserting copyright information as a robust watermark to used for medical applications. But the method does not consider
the host image, the IDWT is used. For inserting the logo informa- the watermark embedding capacity to the host image. Besides, a
tion as a fragile watermark into the host image, the high- blind efficient image watermarking scheme based on integer
frequency components (HH) of IDWT are used in case of content wavelet transform (IWT)-DCT-SVD is proposed in the scheme
authentication. For achieving an optimal balance among the three (Khare et al., 2020). At first, the host image is applied one-level
conflicted and limited requirements like imperceptibility, robust- IWT, and a low-low (LL) sub-band is selected. Then, the watermark
ness, and capacity, an optimization algorithm such as artificial is inserted into the singular values of the LL sub-band of the host
bee colony (ABC) is used. The method ensures better imperceptibil- image. The method is robust against scaling, HE, rotation, SPN,
ity and robustness against JPEG compression, GN, SPN, median fil- GN, MF, and translation. But the security and watermark embed-
ter (MF), cropping, rotation, translation, and resizing attacks. Their ding capacity of the proposal is not analyzed. Hence, for improving
proposed method ensures high security by introducing the Cheby- the watermark embedding capacity to the host image, a robust
shev chaotic sequence. But the method does not calculate the watermarking system is designed in the wavelet domain based
watermark embedding capacity. on NSCT, DCT, and multi-resolution SVD (Singh et al., 2020). The
In another scheme (Zhang et al., 2019), various levels of DWT imperceptibility of the watermarked image and embedding capac-
are applied on the cover (or host) image, and then DCT is merged ity are improved by using the wavelet transforms. The method is
with DWT. Here, for embedding the watermark, a spread transform robust against geometric and common image processing attacks.
quadrature index modulation (QIM) algorithm is used. An orthog- But the method does not consider the security of the watermark
onal matching pursuit compression reconstruction algorithm is image and calculate the computational complexity. On the other
used for optimizing the performance of the watermarking system. hand, they suggest that the performance of the proposal can be
But the method is less robust against print/scan, Wiener filter, improved by using the optimization algorithm. Therefore, a blind
Poisson noise (PN), speckle noise (SN), hybrid, and desynchroniza- image watermarking scheme is proposed based on principal com-
tion attacks. Also, the security of this proposal is not considered. ponent analysis (PCA) optimization in the R-DW (redundant dis-
For ensuring the security of the watermark image, a hybrid robust crete wavelet) domain (Rajani and Kumar, 2020). For improving
watermarking method is proposed based on color by combining the performance of this PCA-R-DW scheme, an improved grey-
DCT and DWT (Abdulrahman and Ozturk, 2019). The RGB color wolf optimizer (IGWO) algorithm is applied to the watermarking
image is divided into three color channels: red, green, and blue. system for controlling the trade-off among conflicting characteris-
Each color component is modified by DCT and DWT, respectively. tics robustness and imperceptibility along with capacity. The
The watermark image is a grayscale that is encrypted by Arnold’s method ensures better imperceptibility than existing watermark-
map. Then, DCT divides the encrypted watermark image into smal- ing methods. The method is robust against GN, resizing, rotation,
ler parts of equal size. The DCT coefficients of each watermarked compression, HE, MF, and sharpening attacks. But this method
image are embedded into the host image. The system ensures does not use any security method to encrypt the watermark image.
robustness against various attacks by the experiment. But the They claimed their method improves the watermark embedding
method is less robust against histogram equalization (HE), low capacity, but they do not calculate this capacity.
pass Gaussian filter (LPGF), cropping, and hybrid attacks. Also, For ensuring security, another researcher proposed a robust and
there exists no idea about the watermark embedding capacity into secure color image watermarking method that works on the
the host image in the proposal. Moreover, the limitation is that wavelet-packet domain (Al-Otum, 2020). In this case, the binary
their proposed method is non-blind as it requires original host watermark logo is inserted into the safe location of the color host
image to extract the watermark image. For minimizing this limita- image. Here, Arnold transform is used for ensuring the watermark
tion, a robust invisible digital image watermarking method based image security. The method ensures better imperceptibility of the
on blind (inner method) and non-blind watermarking techniques watermarked image. Also, the method provides improved robust-
is proposed in (Savakar and Ghuli, 2019). In the blind scheme, a ness against noise, filtering, JPEG 2000, and cropping attacks. The
secret binary image as watermark is embedded into the internal method is designed for protecting copyright applications. But, the
host image using DWT. The scheme generates the inner water- watermark embedding capacity is not good for this method. In
marked image that is inserted into the outer host image by using another research, a blind color image watermarking method is pro-
DWT and SVD with the help of the non-blind watermarking tech- posed in the DCT domain based on just noticeable distortion (JND)
nique. Thus, the hybrid watermarked image is generated. The (Wang et al., 2020). Here, the binary watermark logo image is
method provides better robustness against JPEG compression, encrypted by Arnold transformation and then embedded into the
Gaussian noise, rotation, SPN, PN, and SN. But the method is not color host image. The method ensures strong robustness against
robust against MF, resizing, Wiener filter, print/scan, and hybrid JPEG compression, GN, SPN, MF, Gaussian filter (GF), and rotation
attacks. Moreover, the method has not discussed the security issue attacks. Also, the method guarantees a better trade-off between
of the watermark image. Also, the watermark embedding capacity imperceptibility and robustness. But, the method does not consider
to the host image is not observed. the watermark embedding capacity to the host image.
A method is designed in the scheme (Zheng and Zhang, 2020) A binary watermark image is embedded into the DCT coeffi-
for ensuring high watermark embedding capacity. Here, a hybrid cients of the host image (Sunesh and Kishore, 2020). This method
robust digital image watermarking technique based on transform embeds the same watermark bits to the two different positions
domains like DWT, DCT, and SVD is proposed for resisting rota- of the host image for ensuring security. This blind image water-
tional attacks. The method ensures better imperceptibility and marking method provides better imperceptibility and improved
improved robustness of the watermarking system. But the security robustness against HE and GN attacks. Also, the method has less
is not considered by this as the watermark image is not encrypted computational complexity in terms of time. But, the method does
by any encryption technique. Therefore, in Thakur et al. (2020), a not analyze the watermark embedding capacity. Hence, a Walsh
chaotic hybrid medical image watermarking technique based on Hadamard Transform (WHT) based blind color image watermark-
non-sub-sampled contourlet transform (NSCT), redundant DWT ing method is proposed for ensuring high watermark embedding
5858
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

capacity to the host image in the scheme (Prabha and Sam, 2020). where N is the total number of input data samples: x(0), . . ., x
The RGB color image is divided into 4  4 non-overlapping blocks. (N1), x(n) is the given data sample, y(k) is the output DCT coeffi-
WHT transforms these blocks. Here, the color watermark image is cient, and a(k) is the scaling factor.
embedded into the third and fourth row WHT coefficients. The
method ensures better visual quality. The method has strong 3.2. Discrete wavelet transform (DWT)
robustness against various image manipulations, hybrid and lossy
attacks. The security is ensured by reformatting the watermark A DWT is a wavelet transform that decomposes the host signal
pixel values. Also, the method requires less time to simulate than into wavelets that are discontinuous. Its temporal resolution
the existing methods. But, the space complexity is not considered makes it more attractive over Fourier transform because it contains
in this method. Kazemi et al. proposed a neural network-based more information both in time and frequency (Kehtarnavaz, 2008).
color image watermarking method where the binary watermark A signal x[n], whose DWT is defined by the following equations:
logo is embedded into the edge of the color host image in some dif-
1 X
ferent directions (Kazemi et al., 2020). The contourlet transforms W / ½j0 ; k ¼ pffiffiffiffiffi n
x½n/j0;k ½n ð4Þ
with multilayer perceptron is used to extract the watermark logo M
image. The method is effective by various analyses. But there exists
1 X
no analysis for security or watermark embedding capacity. Hence, W w ½j; k ¼ pffiffiffiffiffi n
x½nwj;k ½n; for j  j0 ð5Þ
in another research (Niu et al., 2020), the watermark embedding M
capacity is determined. The undecimated dual-tree complex wave- where, W / ½j0 ; k and W w ½j; k are the approximation coefficients and
let transform (UDTCWT) domain with Weibull mixtures-based detail coefficients, respectively, and the inverse DWT is given by
vector hidden Markov trees (HMT) is used. In this case, the binary
1 X 1 XX
J
watermark image is inserted into the UDTCWT coefficient magni- x½n ¼ pffiffiffiffiffi W / ½j0 ; k/j0;k ½n þ pffiffiffiffiffi W w ½j; kwj;k ½n ð6Þ
tudes. For extracting the watermark, UDTCWT coefficient magni- k k
M M j¼j0
tudes are used with HMT. The method outperforms better in
terms of robustness and imperceptibility compared to the existing With
state-of-the-art methods. But the watermark image security is not n ¼ 0; 1; 2;    ; M  1; j
observed.
j
From the above discussions, we can say that the existing ¼ 0; 1; 2;    ; J  1; k ¼ 0; 1; 2;    ; 2  1 ð7Þ
works cannot meet the basic design requirements like imper-
where M is the number of transferred samples = 2 J, J is the numbers
ceptibility, robustness, security, and capacity simultaneously
of levels of transfer, f/j;k ½ngand wj;k ½n are two basic functions, /[n]
with perfection. Being motivated by the above-mentioned limi-
tations, we have designed a hybrid blind robust watermarking and w[n] are the scaling and wavelet functions, respectively.
system by combining DCT, DWT, and SVD. Here, DWT is used
for increasing the imperceptibility of the watermarked image. 3.3. Singular value decomposition (SVD)
Besides, SVD is used for increasing more robustness against dif-
ferent attacks. A real matrix (R) can be decomposed into three matrices and
matrix diagonalization can be done by SVD. The three matrices
are two orthogonal matrices (U and V) and one diagonal matrix
3. Theoretical background
(S). In SVD, the matrix S contains the singular values that are
arranged in decreasing order. The real matrix R can be written by
DCT, DWT, and SVD are the most popular transform domain
the following equation (Chung et al., 2007):
techniques for embedding and extracting the watermark into the
host image and from the generated watermarked image, respec- R ¼ U R SR V TR ð8Þ
tively. However, this sub-section keeps an eye on the theoretical
background of these transformed domain techniques along with where, V TR is the transpose matrix of VR.
Arnold transform.
3.4. Arnold transform
3.1. Discrete cosine transform (DCT)
Arnold transform scrambles a digital image to enhance image
DCT converts an image into a transform domain by manipulat- security and reduces the amount of memory storage. This tech-
ing its frequency components. This DCT is a sum of cosine func- nique scrambles the pixel position of an image iteratively (Sui
tions and here, only real values are used (Begum and Uddin, and Gao, 2013; Chen et al., 2013). Let, the size of the host image
2020). DCT is an orthogonal transformation that means it has data is N  N and the pixel co-ordinate is F ¼ fðx; yÞjx; y ¼ 0;
decorrelation and energy compaction properties (Khayam, 2003). It 1; 2;       ::; N  1g. Then, the Arnold transform is given by the fol-
is used for JPEG compression. The one dimensional (1D) DCT can be lowing equation:
calculated by the following Eq. (1) (Meyer-Baese and Schmid,     
xn 1 a xn1
2014): ¼ modN ð9Þ
X   yn b ab þ 1 yn1
N 1
pð2n þ 1Þk
yðkÞ ¼ aðkÞ xðnÞ cos ; k ¼ 0; 1;    ; N  1 ð1Þ
n¼0
2N Here, (xn, yn) is the transformed co-ordinate after n-iterations,
two positive integers are a and b, and N is the height or width of
and the inverse DCT is given by the host image.
X  
N1
pð2n þ 1Þk
xðnÞ ¼ aðkÞyðnÞcos ; n ¼ 0; 1;    ; N  1 ð2Þ 4. Proposed methodology
k¼0
2N

With This section mainly describes the proposed algorithm. The

qffiffiffi qffiffiffi
detailed watermark embedding and extraction algorithms are
að0Þ ¼ N1 ; k ¼ 0 and aðkÞ ¼ N2 ; 1  k  N  1 (3)
explained in Section 4.1 and Section 4.2, respectively.

5859
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

4.1. Proposed algorithm for watermark embedding For the host image, at first, 2D DCT (block size ½8  8) is applied
to the host image and get Dc . Then, the system applies 3L DWT to
At first, the watermark image is preprocessed that consists of the non-overlapping blocks of the host image to extract the high-
securing it by the Arnold map, then applying DCT, DWT, and SVD frequency sub-band and get HH3Dc .
to the scrambled watermark image. Security is required to protect
½LLDc ; LHDc ; HLDc ; HHDc  ¼ DWTðDc Þ ð14Þ
the watermark image from unauthorized access. The Arnold map
shuffles the pixel value, and it depends on the iteration time. So,
½LL2Dc ; LH2Dc ; HL2Dc ; HH2Dc  ¼ DWTðLLDc Þ ð15Þ
we kept this iteration number as a key that the sender and receiver
can only share. In Arnold’s map, the old pixel location ½i; j is chan-
0 0 ½LL3Dc ; LH3Dc ; HL3Dc ; HH3Dc  ¼ DWTðLL2Dc Þ ð16Þ
ged to a new location ½i ; j  by the following equation (10). Then, it
shuffles all pixel locations by n times. Here, n is the key of the algo- Then, the SVD is applied on HH3Dc . That is,
rithm needed by the receiver to decrypt the original watermark ½uc ; sc ; v c  ¼ SVDðHH3Dc Þ. The s component of SVD of the host image
image. is sc . This sc is used for embedding the watermark image. These
steps are the same as for the watermark image. Because, we want
 0 0
i ; j ¼ ½1 1; 1 2 ½i; j; ð10Þ to follow the same procedure to both host and watermark image to
generate the watermarked image.
Then, the two-dimensional (2D) DCT is applied to the encrypted The watermark embedding function is simple. The whole sc is
watermark, ew. We know that DCT is a block-based algorithm. So, divided into ½8  8 blocks. So, if the size of sc is½512  512, then
it requires block to compute, and we choose ½8  8 block to per- the number of blocks will be 4096. Then, for each block, the [8,8]
form the DCT operation. After performing DCT on ew, we get a pixel of the host image is changed by the watermark pixel, sw
new image dew, which is now on the frequency domain. But for and alpha. Then, we get a new pixel value for each [8,8] pixel. After
better imperceptibility, we apply three-level (3L) DWT on dew. then, by combining all blocks, we get a new image which is called
The 3L DWT consists of the following equations: by SSc :Alpha is the watermark strength, and it requires to be opti-
mal for ensuring better imperceptibility. Here, alpha must be
½LLdew ; LHdew ; HLdew ; HHdew  ¼ DWTðdewÞ ð11Þ 0 < alpha < 1. We select the value of alpha randomly for every
image. This relation is defined by equation (17).
½LL2dew ; LH2dew ; HL2dew ; HH2dew  ¼ DWTðLLdew Þ ð12Þ
SSc ¼ alpha sw ð17Þ
Then, the inverse SVD (ISVD), inverse DWT (IDWT), and inverse
½LL3dew ; LH3dew ; HL3dew ; HH3dew  ¼ DWTðLL2dew Þ ð13Þ
DCT (IDCT) are applied to the embedded host image. Thus, the sys-
DWT is better for imperceptibility and the number of levels is tem generates the watermarked image.
also important because more level means a little change on the Fig. 2 shows the proposed method for watermark embedding.
host image is occurred by the watermarking algorithm. So, we There exist some methods that work jointly on DCT, DWT, and
choose 3L DWT, and after then, we apply SVD on HH3dew . SVD. But we have embedded the watermark image on the 3-level
By SVD decomposition, we get three matrices. That means, of DWT and also applied the Arnold map on it. On the other hand,
½uw ; sw ; v w  ¼ SVDðHH3dew Þ. sw is the s component of SVD of the our embedding function is different from other existing methods.
watermark image. We choose sw for watermark embedding because We have chosen the watermark embedding strength randomly,
sw has only diagonal entries, and other entries are zero. It has a which is the strong point for our method. Here, we use a blind
good impact on the watermark embedding algorithm. watermarking method and an invisible watermark image for

Fig. 2. A proposed method for watermark embedding.

5860
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

embedding to the host image. The watermark image is extracted 4.2. Proposed algorithm for watermark extraction
from the watermarked image without using the host image, which
makes this proposed algorithm unique from existing hybrid digital The watermark extraction algorithm is an inverse process of the
image watermarking algorithms. The whole process of watermark watermark embedding algorithm. The proposed watermark extrac-
embedding is shown in Algorithm 1. tion algorithm is shown by Algorithm 2. By this algorithm, we can
Algorithm 1. Proposed watermark embedding method. extract the watermark image from the generated watermarked image.

5861
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Algorithm 2. Proposed watermark extraction method.

5.2. Robustness analysis

In our proposed algorithm, we have combined three different
methods and encrypted the watermark image with the Arnold The performance of the watermarking algorithm is defined by
map. Hence, it increases the computational costs of our method. benchmarks (Voloshynovskiy et al., 2001). One of the popular
It is very difficult to find out the imperceptibility and robustness benchmark packages is the Stirmark benchmark for testing the
at the same time by using a single transformation method. By com- robustness of the watermarking algorithms. There exist various
bining multiple transformation methods, the imperceptibility and attacks like geometric, removal, cryptographic, and protocol
robustness of our method are improved than existing state-of- attacks. We have analyzed filter, noise, geometric, and row/column
the-art methods, which are shown in Section 5. delete attacks in our proposed method. We have tested the robust-
ness of our proposed method based on the second-generation
benchmarking tool (Voloshynovskiy et al., 2001). To evaluate the
performance of the system, we have used CPSNR, WPSNR (PSNR
5. Experimental results and analysis
of the watermark image), the bit error rate of watermark image
(WBER), and normalized correlation of watermark image (WNC).
This section analyses the proposed method in terms of imper-
We have used various filters with different window sizes. These fil-
ceptibility and robustness. Also, we have analyzed the security of
ters include median, average, wiener, range, and standard devia-
our method. Then, we have compared our method with existing
tion filters. The graph for wiener, average, and median filters
recent methods. Here, noted that we have conducted our experi-
with different window sizes is shown in Fig. 3. From this Fig. 3,
ment with MATLAB R2016a.
we see that the value of CPSNR is greater than 30 dB except for
the average filter (50  50). For range filter (3  3) and standard
deviation filter (3  3), CPSNR values are very low. Hence, the
5.1. Imperceptibility analysis imperceptibility of our proposed method is decreased after stan-
dard deviation and range filters. But the proposed method ensures
For evaluating the imperceptibility of our proposed method better imperceptibility even after wiener, average, and median fil-
without attack, CPSNR (PSNR of the cover image) in dB and CSSIM ter attacks. For filtering attacks, WPSNR > 30 dB. This means the
(structural similarity index measure (SSIM) of the cover image) are original watermark and extracted watermark seem to same from
used. The values of CPSNR and CSSIM for our proposed method are a human visual perspective. Also, our method is robust against
57.6303 dB and 0.9984, respectively.
5862
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Fig. 3. Filter attack.

Fig. 4. SPN attack.

Fig. 5. Rotation attack.

5863
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Fig. 6. Watermarked image and extracted watermark image after attacks (a) Median filter (3  3), (b) SPN (0.005), and (c) Rotate (900).

Table 1
Row/column delete attack.
these filter attacks as WNC = 1. WBER is close to 0 for all of these
filters. It proves the strong robustness of the system against these Parameter CPSNR (dB) WNC
filter attacks. Delete 100 rows 14.2461 1
Next, we have analyzed SPN with different noise density values Delete 200 rows 13.0414 0.9999
Delete 100 columns 12.6250 1
and evaluated the performance by CPSNR, WPSNR, WNC, and
Delete 200 columns 12.3356 1
WBER. This is shown by Fig. 4. In this case, the CPSNR value is less
than 30 dB except noise density 0.001, 0.002, and 0.003. Hence, our
method does not ensure better imperceptibility after this attack.
In our algorithm, we have used block-based DCT. The main
Also, this Fig. 4 shows that, WPSNR < 30 dB, which means low
problem of block-based transformation techniques is that the
imperceptibility of the watermark image after this noise attack.
embedded watermark can be destroyed by removing some lines
Moreover, WNC is close to 1, which means our method can extract
or columns in the corners. We can say it is a row/column delete
the watermark image even after this attack with different noise
attack. For our experiment, we have deleted some rows/columns
density values. The system overcomes the noise attack with differ-
from the last corner. The result of the row/column delete attack
ent noise density values, as WBER values are close to 0.
is shown in Table 1. From this table, we can say that the impercep-
After then, we have analyzed rotation attacks (both clockwise
tibility of the host image is decreased as the values of CPSNR are
and counterclockwise) for different degrees, which is shown in
less than 30 dB. But still, our proposed method can extract the
Fig. 5. In this case, we see that CPSNR < 30 dB for both clockwise
watermark image from the generated watermarked image as the
and counterclockwise rotation. Hence, our method does not ensure
values of WNC are close to 1. Hence, our method is also robust
better imperceptibility after this rotation attack. The impercepti-
against row/column delete attack.
bility of the watermark image is decreased after both clockwise
and counterclockwise rotation as the value of WPSNR is less than
30 dB except for counterclockwise rotation 10, 50, 900, and clock- 5.3. Security analysis
wise rotation for 10 and 50. But we can say that, still, the proposed
method can extract the watermark image from the generated In the proposed method, the watermark image is secured by the
watermarked image as the values of WNC are close to 1. The pro- Arnold map. Arnold map is a well-known security method that is
posed method is also robust against rotation attack as the values of used in watermarking research because of its lightweight. In our
WBER are close to 0. method, we have used the number of iteration as the secret key.
Hence, our method is robust against the filter, SPN, and rotation In the following Fig. 7, it is shown that the appropriate secret key
attacks. correctly decrypts the watermark image. We have also changed
Fig. 6 shows the output of the generated watermarked images the key a little bit, and then we have seen that the watermark
and extracted watermark images from the watermarked images image is wrongly decrypted. The value of SSIM of these two
for different attacks like median filter (3  3), SPN (0.005), and decrypted images is 0.0721, which means there exists no similarity
counterclockwise rotation for 900. between these two images. The value of SSIM of the original water-
5864
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Fig. 7. (a) Original watermark (b) Encrypted watermark (key = 55) (c) Decrypted watermark (key = 55) (d) Decrypted watermark (key = 56).

Table 2
Comparison of imperceptibility of existing methods with different parameters.

Parameter Al-Otum (2020) Wang et al. Sunesh and Kishore Prabha and Sam Kazemi et al. (2020) Niu et al. (2020) Proposed Method
(2020) (2020) (2020)
Method Inter Layer Wavelet DCT DCT WHT Contourlet UDTCWT + HMT DCT + DWT + SVD
Packets Transformation
CPSNR (dB) 62.37 45.4018 47.6637 49.2075 48.81 50.3151 57.6303
CSSIM – 0.9872 – 0.9948 – – 0.9984
Watermark Binary Binary Binary Color Binary Binary Grayscale
Type
Watermark – 64  64 32  32 90  90 – 32  32 64  64
size
Security Arnold Map Arnold Map – Shuffling Arnold Map – Arnold Map

mark and the decrypted watermark (right key) is 1, which means attack. Hence, we can say that our proposed method ensures better
these two images are identical. imperceptibility than the existing recent methods.

5.4. Comparison with existing methods

5.4.2. Robustness comparison
Table 3 compares the robustness of the proposed method with
In this Section, we have compared the imperceptibility and
the recent state-of-the-art methods for various attacks by WNC.
robustness of our proposed method with existing state-of-the-art
From this table, we can say that, for the median filter with different
methods.
window sizes, WNC = 1, which is greater than the existing meth-
ods. Hence, our method is robust against filter attack. For SPN with
5.4.1. Imperceptibility comparison a noise density of 0.005, the WNC value is greater than the method
We have compared our proposed method with existing recent (Wang et al., 2020). SPN with other noise densities, methods
methods by different parameters, such as CPSNR, CSSIM, water- (Wang et al., 2020; Kazemi et al., 2020) perform better than our
mark type, watermark size, and security techniques. The results method. But still, the proposed method can extract the watermark
are shown in Table 2. From this table, we can see that the proposed image as the WNC values are close to 1. For counterclockwise rota-
method’s CSSIM is the highest and the CPSNR is higher than the tion with different degrees, the WNC values are close to 1, which
existing methods except for method (Al-Otum, 2020) without are greater than the recent methods. Hence, our method can

Table 3
Comparison of the robustness of the proposed method with existing methods after attacks by WNC.

Attacks Wang et al. (2020) Prabha and Sam (2020) Kazemi et al. (2020) Proposed Method
Median Filter (3  3) 0.8731 0.8371 0.9609 1
Median Filter (5  5) – 0.7250 0.7402 1
Median Filter (7  7) – 0.6909 – 1
Median Filter (9  9) – 0.6719 – 1
SPN (0.001) – – 1 0.9992
SPN (0.005) 0.9744 – – 0.9956
SPN (0.01) 0.9971 – 0.9785 0.9912
SPN (0.02) 0.9938 – – 0.9831
SPN (0.03) 0.9908 – – 0.9762
SPN (0.04) 0.9870 – – 0.9699
Rotate (100) – 0.9378 – 0.9993
Rotate (200) – 0.8920 – 0.9989
Rotate (300) 0.9963 0.8662 – 0.9988
Rotate (400) – 0.8560 – 0.9987
Rotate (600) 0.9967 – – 0.9988
Rotate (900) 1 – – 1
Rotate (1200) 0.9963 – – 0.9988

5865
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

extract the watermark image after this rotation attack. So, our pro- D’Angelo, A., Cancelli, G., Barni, M., 2010. Watermark-based authentication. In:
Sencar, H.T., Velastin, S., Nikolaidis, N., Lian, S. (Eds.), Intelligent Multimedia
posed method is robust against median filter with different win-
Analysis for Security Applications, Studies in Computational Intelligence.
dow sizes, SPN, and rotation with different degrees compared to Springer, Berlin, Heidelberg.
the existing methods. Giakoumaki, A., Pavlopoulos, S., Koutsouris, D., 2006. Multiple image watermarking
applied to health information management. IEEE Trans. Inf. Technol. Biomed. 10
(4), 722–732.
6. Conclusions and future work Hannoun, K., Hamiche, H., Lahdir, M., Laghrouche, M., Kassim, S., 2018. A novel DWT
domain watermarking scheme based on a discrete-time chaotic system. IFAC-
Pap. Line 51 (33), 50–55.
In this research paper, we have proposed a hybrid blind robust Haribabu, M., Bindu, C.H., Swamy, K.V., 2016. A secure & invisible image
digital image watermarking method by combining DCT, three-level watermarking scheme based on wavelet transform in HSI color space. In:
(3L) DWT, and SVD. We have implemented watermark embedding Proceedings of the 6th International Conference on Advances in Computing &
Communications, pp. 462–468.
and extraction algorithms for generating watermarked and
Jia, S., Zhou, Q., Zhou, H., 2017. A novel color image watermarking scheme based on
extracted watermark images, respectively. We have performed DWT and QR decomposition. J. Appl. Sci. Eng. 20, 193–200.
extensive experiments to evaluate the performance of our pro- Kazemi, M.F., Pourmina, M.A., Mazinan, A.H., 2020. In: Analysis of watermarking
framework for color image through a neural network-based approach,”, pp.
posed method and obtained promising results. Our proposed
213–220.
method ensures better imperceptibility, which is 57.6303 dB. The Kehtarnavaz, N., 2008. Digital Signal Processing System Design. Elsevier,
method also provides improved robustness against the filter, Cambridge, MA, USA.
salt-and-pepper noise (SPN), and rotation attacks compared to Khare, P., Srivastava, V.K., 2020. An efficient image watermarking technique based
on IWT-DCT-SVD. In: Dutta, D., Kar, H., Kumar, C., Bhadauria, V. (Eds.), Advances
the existing state-of-the-art methods. For a median filter with dif- in VLSI, Communication, and Signal Processing, Lecture Notes in Electrical
ferent window sizes, our WNC value is 1, which is greater than Engineering. Springer, Singapore, pp. 841–849.
existing methods. For SPN with different noise density values, Khayam, S.A., 2003. The Discrete Cosine Transform (DCT): Theory and Application.
Michigan State University.
WNC values are close to 1, and for counter-clockwise rotation, Laouamer, L., Tayan, O., 2015. A semi-blind robust DCT watermarking approach for
our WNC values are greater than other existing methods. For row/- sensitive text images. Arab. J. Sci. Eng. 40, 1097–1109.
column delete attack, our WNC values are close to 1. The Arnold Lei, B., Soon, I., Li, Z., 2012. A robust audio watermarking scheme based on lifting
wavelet transform and singular value decomposition. In: Proceedings of
map provides enhanced security to the system. The original water- International Workshop on Watermarking and Forensics. Springer, Berlin
mark and decrypted watermark are the same for our method, as Heidelberg, pp. 86–96.
the value of SSIM of these two images is 1. In the future, we will Lei, B., Soon, I.Y., Tan, E.L., 2013. Robust SVD-based audio watermarking scheme
with differential evolution optimization. IEEE Trans. Audio Speech Lang.
try to reduce the computational complexity in terms of the time
Process. 21 (11), 2368–2378.
and space of our proposal. Also, we will extend our proposed Lei, B., Ni, D., Chen, S., Wang, T., Zhou, F., 2014. Optimal image watermarking
method with the help of genetic and neural network-based opti- scheme based on chaotic map and quaternion wavelet transform. Nonlinear
Dyn. 78 (4), 2897–2907.
mization algorithms for improving robustness against a wide range
Lei, B., Zhou, F., Tan, E.-L., Ni, D., Lei, H., Chen, S., Wang, T., 2015. Optimal and secure
of attacks. Moreover, we will extend this hybrid blind grayscale audio watermarking scheme based on self-adaptive particle swarm
image watermarking method to RGB color images. optimization and quaternion wavelet transform. Signal Process. 113, 80–94.
Lei, B., Zhao, X., Lei, H., Ni, D., Chen, S., Zhou, F., Wang, T., 2019. Multipurpose
watermarking scheme via intelligent method and chaotic map. Multimed Tools
Declaration of Competing Interest Appl. 78, 27085–27107.
Loani, N.A., Hurrahi, N.N., Parah, S.A., Lee, J.W., Sheikhi, J.A., Bhat, G.M., 2018. Secure
The authors declare that they have no known competing finan- and robust digital image watermarking using coefficient differencing and
chaotic encryption. IEEE Access 6, 19876–19897.
cial interests or personal relationships that could have appeared Megalingam, R.K., Nair, M.M., Srikumar, R., Balasubramanian, V.K., Sarma, V.S.V.,
to influence the work reported in this paper. 2010. Performance comparison of novel, robust spatial domain digital image
watermarking with the conventional frequency domain watermarking
techniques. In: 2010 International Conference on Signal Acquisition and
Acknowledgment Processing, Bangalore, India,, pp. 349–353.
Meyer-Baese, A., Schmid, V., 2014. Feature selection and extraction. In: Pattern
The authors are thankful to the Information and Communica- Recognition and Signal Analysis in Medical Imaging. Elsevier, pp. 21–69. https://
doi.org/10.1016/B978-0-12-409545-8.00002-9.
tion Technology Division of the Government of the People’s Repub-
Mukherjee, D.P., 2004. Spatial domain digital watermarking of multimedia objects
lic of Bangladesh for a Ph.D. fellowship to Mahbuba Begum. for buyer authentication. In: IEEE Trans. Multimed, pp. 1–15.
Mukherjee, D.P., Maitra, S., Acton, S.T., 2004. Spatial domain digital watermarking of
multimedia objects for buyer authentication. IEEE Trans. Multimedia 6 (1), 1–
15.
References Nikolaidis, N., Pitas, I., 1998. Robust image watermarking in the spatial domain.
Signal Process. 66 (3), 385–403.
Niu, P., Shen, X., Wei, T., Yang, H., Wang, X., 2020. In: Blind Image Watermark
Abdulrahman, A.K., Ozturk, S., 2019. A novel hybrid DCT and DWT based robust
Decoder in UDTCWT Domain Using Weibull Mixtures-Based Vector HMT, pp.
watermarking algorithm for color images. Multimedia Tools Appl. 78, 17027–
46624–46641.
17049.
Olanrewaju, R.F., 2011. Development of Intelligent Digital Watermarking via Safe
Ali, M., Ahn, C.W., Pant, M., Siarry, P., 2016. A reliable image watermarking scheme
Region Ph.D. Thesis. Kulliyyah of Engineering, International Islamic University
based on redistributed image normalization and SVD. Discret. Dyn. Nat. Soc.
Malaysia, Selangor Malaysia.
2016, 1–15.
Patra, J.C., Phua, J.E., Rajan, D., 2010. DCT domain watermarking scheme using
Al-Otum, H.M., 2020. Secure and robust host-adapted color image watermarking
chinese remainder theorem for image authentication. In: Proceedings of the
using inter-layered wavelet-packets. J. Vis. Commun. Image R. 66, 102726.
International Conference on Multimedia and Expo, pp. 111–116.
Begum, M., Uddin, M.S., 2020. Digital image watermarking techniques: a review.
Prabha, K., Sam, I.S., 2020. An effective robust and imperceptible blind color image
Information 11, 110.
watermarking using WHT. J. King Saud Univ. – Comput. Inf. Sci. https://doi.org/
Begum, M., Uddin, M.S., 2020. Analysis of digital image watermarking techniques
10.1016/j.jksuci.2020.04.003.
through hybrid methods. Adv. Multimedia 2020, 7912690.
Pun, C.M., 2009. High capacity and robust digital image watermarking. In:
Celik, M., Sharma, G., Saber, E., Murat Tekalp, A., 2002. Hierarchical watermarking
Proceedings of the 5th International Joint Conference on INC, IMS and IDC,
for secure image authentication with localization”. IEEE Trans. Image Process.
Seoul, South Korea, pp. 1457–1461.
11 (6), 585–595.
Rajani, D., Kumar, P.R., 2020. An Optimized blind watermarking scheme based on
Chang, C.-C., Tsai, P., Lin, C.-C., 2005. SVD-based digital image watermarking
principal component analysis in redundant discrete wavelet domain. Signal
scheme. J. Pattern Recognit. Lett. 26 (10), 1577–1586.
Process. 172, 107556.
Chen, Z., Chen, Y., Hu, W., Qian, D., 2015. Wavelet domain digital watermarking
Savakar, D.G., Ghuli, A., 2019. Robust invisible digital image watermarking using
algorithm based on threshold classification. In: International Conference in
hybrid scheme. Arab. J. Sci. Eng. 44, 3995–4008.
Swarm Intelligence. Springer, Cham, Switzerland, pp. 129–136.
Singh, S.P., Bhatnagar, G., 2018. A New robust watermarking system in Integer DCT
Chen, Y.-L., Yau, H.-T., Yang, G.-J., 2013. A Maximum. Entropy 15 (8), 3170–3185.
domain. J. Vis. Commun. Image Represent 53, 86–101.
Chung, K.L., Yang, W.N., Huang, Y.H., Wu, S.T., Hsu, Y.C., 2007. On SVD-based
watermarking algorithm. Appl. Math. Comput. 188 (1), 54–57.

5866
M. Begum, J. Ferdush and M.S. Uddin Journal of King Saud University – Computer and Information Sciences 34 (2022) 5856–5867

Singh, R., Nigam, S., Singh, A.K., Elhoseny, M., 2020. Integration of wavelet Voloshynovskiy, S., Pereira, S., Iquise, V., Pun, T., 2001. Attack modelling: towards a
transforms for single and multiple image watermarking. In: Intelligent second generation watermarking benchmark. Signal Process. 81 (6), 1177–
Wavelet Based Techniques for Advanced Multimedia Applications. Cham, 1214.
Springer, pp. 51–63. Voloshynovskiy, S., Pereira, S., Pun, T., Eggers, J.J., Su, J.K., 2001. Attacks on digital
Sui, L., Gao, B., 2013. Color image encryption based on Gyrator transform and Arnold watermarks: classification, estimation based attacks, and benchmarks. IEEE
transform. Opt. Laser Technol. 48, 530–538. Commun. Mag. 39 (8), 118–126.
Sunesh, R., Kishore, R., 2020. A novel and efficient blind image watermarking in Wang, R.-Z., Lin, C.-F., Lin, J.-C., 2001. Image hiding by optimal LSB substitution and
transform domain. In: International Conference on Computational Intelligence genetic algorithm. Pattern Recogn. 34 (3), 671–683.
and Data Science (ICCIDS 2019), Procedia Computer Science, pp. 1505–1514. Wang, J., Wan, W.B., Li, X.X., Sun, J.D., Zhang, H.X., 2020. Color image watermarking
Takore, T.T., Kumar, P.R., Devi, G.L., 2018. A new robust and imperceptible image based on orientation diversity and color complexity. Expert Syst. Appl. 140,
watermarking scheme based on hybrid transform and PSO. I.J. Intell. Syst. Appl. 112868.
10 (11), 50–63. Zhang, Y., Li, Y., Sun, Y., 2019. Digital watermarking based on Joint DWT–DCT and
Tao, H., Chongmin, L.i., Mohamad Zain, J., Abdalla, A.N., 2014. Robust Image OMP reconstruction. Circuits, Syst., Signal Process. 38, 5135–5148.
watermarking theories and techniques: a review. J. Appl. Res. Technol. 12 (1), Zhao, X., Ho, A.T.S., 2010. An introduction to robust transform based image
122–138. watermarking techniques. In: Sencar, H., Velastin, S., Nikolaidis, N., Lian, S.
Thakur, S., Singh, A.K., Ghrera, S.P., Mohan, A., 2020. Chaotic based secure (Eds.), Intelligent Multimedia Analysis for Security Applications, Studies in
watermarking approach for medical images. Multimedia Tools Appl. 79, Computational Intelligence. Springer, Berlin Heidelberg, pp. 337–364.
4263–4276. Zheng, P., Zhang, Y., 2020. A robust image watermarking scheme in hybrid
Vaishnavia, D., Subashini, T.S., 2014. Robust and invisible image watermarking in transform domains resisting to rotation attacks. Multimedia Tools Appl. 79,
RGB color space using SVD. In: Proceedings of the International Conference on 18343–18365.
Information and Communication Technologies, Kochi India, pp. 1770–1777.

5867