Performance evaluation of RDWT-SVD and

DWT-SVD watermarking schemes

Cite as: AIP Conference Proceedings 1774, 050021 (2016); https://doi.org/10.1063/1.4965108
Published Online: 19 October 2016

Taha H. Rassem, Nasrin M. Makbol, and Bee Ee Khoo

ARTICLES YOU MAY BE INTERESTED IN

Robust image watermarking using DWT and SVD for copyright protection
AIP Conference Proceedings 1813, 040003 (2017); https://doi.org/10.1063/1.4975968

Watermarking scheme based on singular value decomposition and homomorphic transform

AIP Conference Proceedings 1897, 020036 (2017); https://doi.org/10.1063/1.5008715

A robust watermarking scheme using lifting wavelet transform and singular value
decomposition
AIP Conference Proceedings 1802, 020002 (2017); https://doi.org/10.1063/1.4973252

AIP Conference Proceedings 1774, 050021 (2016); https://doi.org/10.1063/1.4965108 1774, 050021

© 2016 Author(s).
 Performance Evaluation of RDWT-SVD and DWT-SVD
Watermarking schemes
Taha H. Rassem1, a), Nasrin M. Makbol2, b), Bee Ee Khoo2, c)
1
Faculty of Computer Systems and Software Engineering, Universiti Malaysia Pahang, Kuantan, Malaysia.
2
School of Electrical and Electronic Engineering, Universiti Sains Malaysia, Penang, Malaysia.
Corresponding authors: a ) tahahussein@ump.edu.my , b) Nasrin_makbol@usm.my
c)
Beekhoo@usm.my

Abstract. Digital image watermarking protects content by embedding a signal (i.e., owner information) into the host
image without noticeable degradation in visual quality. To develop any image watermarking scheme, there some
important requirements should be achieved such as imperceptibly, robustness, capacity, security, and, etc. Generally, the
watermarking scheme based on wavelet transform domain shows an advantage in human perception and good
imperceptibility and robustness. Due to this fact, this paper presents two blind image watermarking schemes based on
DWT-SVD and RDWT-SVD. To evaluate their performance, these schemes are exposed to different geometric and non-
geometric attacks. Although, DWT-SVD and RDWT-SVD showed robust against all attacks, RDWT-SVD is better than
DWT-SVD, especially for geometrical attacks.

Keywords: Digital watermarking, Discrete Wavelet Transform (DWT), Redundant Discrete Wavelet Transform
(RDWT), Singular Value Decomposition (SVD).

INTRODUCTION
Protecting digital multimedia data on the world web is attracting more and more attention for a long time. The
rapid development in computer technology and the easy availability of tools enabled malicious parties to produce
illegal copies of copyrighted multimedia by removing or replacing the owner information. A lot of technologies are
developed to protect those copyrighted multimedia data. They are stenography techniques, encryption techniques,
information hiding techniques and watermarking techniques. Digital watermarking represents the ideal solution into
achieving the required security to preserve that data. Due to that, many digital watermarking techniques are
available [1-2].
Digital watermarking refers to the process of embedding secret digital information (such as image, text or number)
called a watermark into the digital content (such as an image, video, audio, etc.). The watermark can be classified
into visible and invisible watermarks. Embedding watermarks into media serve various purposes such as copyright
protection, fingerprinting, authentication, copyright control, etc. Each application, in which the watermark will be
applied, should satisfy some properties. The most important of them are robustness, imperceptibility, security and
capacity. If the watermark can be detected after geometric or non-geometric attacks refer as the robustness. If the
watermark cannot be seen by the human visual system, then it is imperceptible. Security is the ability of the
watermark to resist attacks those want to remove it. Finally, the capacity is how much information can be embedded
into the digital media. Each of these properties should be considered when developing any digital watermarking
technique.
The digital watermarking techniques are categorized into spatial domain techniques and frequency-domain
techniques. In the spatial domain watermarking techniques, the watermark will slightly modify the selected host

International Conference on Advanced Science, Engineering and Technology (ICASET) 2015

AIP Conf. Proc. 1774, 050021-1–050021-9; doi: 10.1063/1.4965108
Published by AIP Publishing. 978-0-7354-1432-7/$30.00

050021-1
image pixels. The frequency-domain techniques which called transform domain techniques; the host media (such as
an image) will be transformed to frequency coefficients. Then the watermark will be embedded into the host image
by altering its frequency coefficients. The spatial domain techniques provide simple embedding and less robust. In
contrast, the frequency domain techniques (such as DCT [3], DFT [4], DWT [2] [5], RDWT [6-7], LWT [8], etc.), is
superior on the spatial domain techniques. They have a good imperceptibility and good robustness, especially
against geometric attacks. Hybrid transform domain schemes are developed to increase the robustness and the
transparency. Many of the recently hybrid transform techniques are combined SVD transform with another
transform such as DCT, DWT, RDWT, etc. SVD is newly used for digital image watermarking applications. It has a
lot of properties to support the transforms that will be joined with them [9]. It will increase the most important
requirements for any watermarking scheme, which are robust and imperceptibility.
The digital watermarking schemes can be classified into blind, semi-blind and non-blind schemes. This is
based on the information that needed during the extracting process. To establish a secure robust system, blind issue
should be considered. In this paper, two hybrid blind, robust and imperceptible image watermarking schemes are
presented. They are DWT-SVD and RDWT-SVD schemes. For each algorithm, the first step is to decompose the
host image into four sub bands using DWT or RDWT respectively. Then SVD will be applied to the LL sub-band
(DWT, RDWT). Finally, the gray image watermark (Doll image) will be embedded into the singular values of the
LL sub-band. The results show that both schemes are robust against all attacks. To be more precise, DWT-SVD is
more robust against non-geometric attacks while RDWT-SVD scheme is robust against most of the geometric
attacks.
The paper will be organized as follows: the wavelets transforms are presented in section 2. The embedding and
extracting steps of the two proposed methods will be shown in section 3. In section 4, experimental results will be
discussed. Finally, conclusions will be presented in section 5.

TRANSFORM DOMAIN SCHEMES

Discrete Wavelet Transform (DWT)

DWT represented one of the frequency-domain schemes that widely used for digital watermarking. Many levels of
DWT can be done on the image. In each level of DWT, the image will pass through a series of filters to decompose
the image into four sub-bands denoted as LL, LH, HL and HH. The size of each sub-band will be half of the host
image size. The decomposition processes will be continued until reaching the desired number of levels needed by
the proposed method. One -level DWT-SVD scheme is used in this paper. The decomposition process to one-level
DWT can be shown in Fig.1 (a) and Fig 1(b). Among long time, DWT schemes showed a good performance in
digital watermarking, but they are suffering from some drawbacks. The main drawback is the shift variant
characteristic which arises from using the down-sampling operation. RDWT has overcome the shift problem.

(a) (b)
FIGURE 1. (a) One level DWT decomposition into four sub-bands (LL, HL, LH, and HH). (b) One level Haar DWT (The
bridge image is 512 x 512 pixels and each sub-band is 256 x 256 pixels). The Bridge image, courtesy of pixabay.com (public
domain), image ID: 1660417.

050021-2
 Redundant Discrete Wavelet Transforms (RDWT)
RDWT is a shift invariance property. Many schemes based on RDWT have been proposed to overcome the shift
variant problem of DWT. Since, it removes the down-sampling operation from DWT to produce an over-complete
representation of the frequency coefficients [6]. As in DWT, applying RDWT on a digital image (such as Bridge
image) will decompose it into four sub-bands. Each sub-band size is the same as that of the host image which
achieved more robustness. Since the redundancy in the transformed domain is more robust in carrying watermarking
information [10]. Furthermore, the decomposition process can be performed along several levels as needed. One-
level RDWT is used in this paper. The one-level RDWT decomposition is shown in Fig.2 (a) and the result of it on
Bridge image will be shown in Fig. 2(b).

(a) (b)
FIGURE 2. (a) one level RDWT decomposition into four sub-bands (LL, HL, LH, and HH). (b) One level Haar RDWT (The
bridge image is 512 x 512 pixels and each sub-band is 512 x 512 pixels). The Bridge image, courtesy of pixabay.com (public
domain), image ID: 1660417.

Singular Value Decomposition (SVD)

Singular Value Decomposition is a linear algebra technique. It showed a good performance in a lot of applications,
especially digital image watermarking. Developing hybrid schemes based on SVD and other transforms is the
recently digital watermarking research area. This is due to the properties of employing the SVD technique.
Performing SVD on an image is computationally expensive, so developing hybrid DWT-SVD and RDWT-SVD
based watermarking scheme will require less computational effort to yield better performance [2]. Employing SVD
will decompose a given matrix of an image into three matrices. For the proposed DWT-SVD and RDWT-SVD, the
matrices will be the same size. The SVD of an image A of MxM is defined as:

A  USV T (1)

U and V are orthogonal matrices, S represents the diagonal matrix S  R .

nm

The properties of employing SVD in digital image watermarking are as follows [11]:
 The Singular Values of an image has very good stability. Since, if a small value is added to the host image,
this will not affect the visual perception of it.
 SVD can represent the intrinsic algebraic properties of an image, where the singular values will represent
the brightness of the image, and the singular vectors will reflect the geometrical characteristics of the
image.
 The diagonal image matrix has many small singular values compared with the first singular value. So, the
reconstructed image quality will not affect a lot if these small singular values are ignored.

050021-3
 THE PROPOSED WATERMARKING SCHEMES
The embedding and the extracting algorithms are same for DWT-SVD and RDWT-SVD schemes.

The Embedding Algorithm

 Apply the transform (DWT for DWT-SVD scheme and RDWT for RDWT-SVD scheme), onto the cover
image to decompose it into its four sub-bands, which are LL, LH, HL and HH.
 Apply SVD to the LL sub-band , as follows:

A  USV T (2)
 Modify the Singular Values S for the LL sub-band by embedding the watermark directly, and then again
apply SVD to it, respectively, as follows:

S  W  UW SW VW
T
(3)

 represents the scaling factor (  = 0:05, for watermark embedding into the LL sub-band).

 Perform the new modified coefficients.

Anew  USW V T (4)

Where Anew will indicate to the DWT new modified coefficients for DWT-SVD scheme, and the RDWT new
modified coefficients for RDWT-SVD scheme.

 Finally, apply the inverse DWT using the DWT modified coefficients to obtain the watermarked image
AW.
AW  DWT 1 (5)

For RDWT-SVD, apply the inverse RDWT using the RDWT modified coefficients to obtain the watermarked
image AW.

AW  RDWT 1 (6)

The Extracting Algorithm

The steps are as follows:
 Apply (DWT for DWT-SVD scheme or RDWT for RDWT-SVD scheme) on the watermarked image A*W
(possibly distorted) to decompose it into four sub-bands LL, LH, HL, HH.

 Perform SVD to the LL sub-band, as follows:

T
AW  U  S V  (7)

 Calculate
T
D   U W S VW (8)

050021-4
  Calculate

W   (D  S ) /  (9)

where W* is the extracted watermark from the LL sub-band.

EXPERIMENTAL RESULTS

To evaluate the proposed schemes performance, some experiments are carried out. The standard 512x512 Bridge
image was used as a cover image. The Doll image was used as the watermark image. They are illustrated in Fig. 3(a)
and Fig 3(b). The size of watermark image is 512x512 for RDWT-SVD scheme, while 256x256 for DWT-SVD
scheme. In embedding process, the scaling factor that used is 0.05. Fig. 3(c) and Fig 3(d) show the watermarked
images after embedding process using DWT and RDWT respectively.

(a) (b)

(c) (d)
FIGURE. 3.(a) Host image (512x512 Bridge). (b) Gray watermark (512x512 Doll image). The Doll image, courtesy of
pixabay.com (public domain), image ID: 1636128. (c) RDWT-SVD watermarked image (PSNR = 54.44) (d) DWT-SVD
watermarked image (PSNR = 53.1313). The Bridge image, courtesy of pixabay.com (public domain), image ID: 1660417.

Each digital image watermarking scheme has to achieve the necessary requirements which are imperceptibility,
robustness, capacity and security. The imperceptibility requirement to indicate the similarity between the host image
and the watermark image. If the watermark image is looking same as the host image, which means the scheme
achieved the imperceptibility. The PSNR (Peak Signal-to-Noise Ratio) is used to estimate the similarity between the
host image and the watermarked image. High PSNR means good imperceptibility. It is calculated as follows:
 max( x(i, j )) 2 
PSNR  10 log10   (10)
 MSE 
Where the Mean Square Error (MSE) between the host image x and the watermarked image y is defined as:
2
1 m n
MSE   x(i, j )  y(i, j )
m * n i 1 j 1
(11)

050021-5
To evaluate the robustness of the proposed schemes, the watermarked images are exposed to various attacks. The
Normalized Cross-Correlation (NC) measures the similarity between the original watermark and the extracted
watermark from the attacked image. The NC is defined as:

w w i i
NC ( w, w)  i 1
(12)
N N

w w 2 2
i i
i 1 i 1

Table I shows the several attacks that were used in this paper. Different types of noises were added to the
watermarked image. They are speckle noise, pepper and salt noise and Gaussian noise. Median filter, wiener filter,
Gaussian filter and un-sharp filter were applied with different size windows. The JPEG compression attack is one of
the important image processing attacks. It applied to the watermarked image using multiple values of the quality
factor (Q=5, 10, 20 and 30). Two remaining image processing attacks were applied. They are gamma correction and
the histogram equalization. Some of the geometrical attacks also applied with different parameters such as rotate,
scale, share, cut, and shift and translate attacks. Table 1 presents the NC and PSNR results of DWT-SVD and
RDWT-SVD schemes. Although, both proposed schemes performed well against all attacks, RDWT-SVD is better
than DWT-SVD. Except the scaling attack, RDWT-SVD is more robust against the geometrical attacks. In addition,
it is more robust against sharpening attack, histogram equalization attack, and gamma correction attack. DWT-SVD
performed well against JPEG compression attack, filtering attacks and scaling attacks. The proposed schemes
performed mutually against noise attacks. Table 2 summarizes the RDWT-SVD and DWT-SVD performances. The
size of the watermark image in RDWT-SVD scheme is 512x512 pixels, while 256x256 pixels in DWT-SVD. So, the
RDWT-SVD achieved more capacity than DWT-SVD. The two proposed schemes are blind and that yields to
achieve the security issue. Fig 4 shows some attacked watermarked images and the extracted watermark from these
images.

050021-6
 Table 1. The NC of the extracted watermark under different attacks
RDWT-SVD DWT-SVD
Non-geometrical Attack
LL PSNR LL PSNR
Gamma Correction 0.8 0.995 22.75 0.992 22.74
Pepper & Salt Noise(density 0.3) 0.728 10.49 0.727 10.44
Pepper & Salt Noise(density 0.01) 0.893 25.24 0.896 25.32
Pepper & Salt Noise(density 0.001) 0.984 35.70 0.969 35.68
Pepper & Salt Noise(density 0.005) 0.927 28.46 0.908 28.45
Speckle Noise(var=0.01) 0.929 27.48 0.916 27.48
Speckle Noise(var=0.001) 0.989 37.36 0.980 37.33
Speckle Noise(var=0.004) 0.954 30.48 0.941 30.47
Gaussian Noise(M=0,var=0.001) 0.934 30.02 0.916 30.01
Gaussian Noise(M=0,var=0.005) 0.868 23.12 0.851 23.12
Gaussian Noise(M=0,var=0.5) 0.707 7.85 0.707 7.85
Histogram Equalization 0.980 15.57 0.975 15.57
JPEG Compression Q=5 0.933 29.24 0.951 29.24
JPEG Compression Q=10 0.967 32.18 0.976 32.17
JPEG Compression Q=20 0.979 34.54 0.985 34.53
JPEG Compression Q=30 0.985 35.81 0.990 35.79
Median Filter (3x3) 0.956 34.11 0.964 34.12
Median Filter (5x5) 0.910 31.55 0.911 31.56
Wiener Filtering(3x3) 0.987 44.92 0.990 44.81
Wiener Filtering(5x5) 0.980 42.92 0.981 42.88
Gaussian Filtering (3x3) 0.987 41.77 0.993 41.73
Gaussian Filtering (5x5) 0.987 41.75 0.993 41.70
Sharping 2 0.955 26.15 0.930 26.10
RDWT-SVD DWT-SVD
Geometrical Attack
LL PSNR LL PSNR
Scaling (zoomout=0.5,zoomin=2) 0.942 33.17 0.980 33.20
Scaling (zoomout=0.25,zoomin=4) 0.801 29.76 0.818 29.77
Scaling (zoomout=2,zoomin=0.5) 0.991 43.18 0.996 43.20
Shift Attack 2 0.996 28.77 0.990 28.70
Shearing (x=1,y=0.2) 0.923 9.83 0.902 9.831
Shearing (x=0.2,y=1) 0.922 9.55 0.888 9.54
Cut 10 0.995 24.00 0.994 24.00
Cut 20 0.995 20.87 0.994 20.70
Cut 30 0.990 18.98 0.990 18.98
Rotation (angle 45 ) 0.916 15.46 0.901 15.45
Rotation (angle 2 ) 0.937 17.15 0.923 17.12
Rotation (angle 70 ) 0.996 14.58 0.975 14.38
Rotation (angle 110 ) 0.937 13.94 0.923 13.93
Rotation (angle -50 ) 0.916 15.27 0.900 15.26
Rotation (angle 50 ) 0.980 17.84 0.977 17.82
Translate (10,10) 0.995 18.78 0.994 18.75
Translate (10,20) 0.993 18.47 0.991 18.45
Translate (20,35) 0.991 17.78 0.990 17.76
Translate (40,40) 0.992 16.87 0.991 16.83

Table 2.RDWT-SVD and DWT-SVD Comparison performance

Attack RDWT-SVD DWT-SVD
Filtering /
JPEG compression /
Scale /
Noise addition - -
Gamma correction /
Histogram Equalization /
sharpening /
Rotate /
shear /
cut /
Shift /
Translate /

050021-7
 (a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o) (p)

FIGURE 4. (a) DWT-SVD watermarked image (JPEG compression Q=5, PSNR=27.3230 db). (b)
DWT-SVD extracted watermark (JPEG compression Q=5, NC= 0.976). (c) RDWT-SVD watermarked
image (JPEG compression Q=5, PSNR=27.3243 db). (d) RDWT-SVD extracted watermark (JPEG
compression Q=5, NC= 0.951). (e) DWT-SVD watermarked image (Rotation 110, PSNR=10.1546
db). (f) DWT-SVD extracted watermark (Rotation 110, NC=0.971) (g) RDWT-SVD watermarked
image (Rotation 110, PSNR=10.1566 db) 4(h) RDWT-SVD extracted watermark (Rotation 110,
NC=0.980). (i) DWTSVD watermarked image (Pepper & Salt Noise (value=0.3), PSNR=10.6778 db).
(j) DWT-SVD extracted watermark (Pepper & Salt Noise(value=0.3), NC= 0.641) 4(k) RDWT-SVD
watermarked image (Pepper & Salt Noise(value=0.3) , PSNR=10.6778 db).(l) RDWT-SVD extracted
watermark ( Pepper &Salt Noise(value=0.3) , NC=0.668).(m) DWT-SVD watermarked image (
Translate (x=10, y=10), PSNR=14.2653 db).(n) DWT-SVD extracted watermark (Translate (x=10,
y=10), NC=0.991). (o) RDWT-SVD watermarked image (Translate (x=10, y=10), PSNR=14.2691
db). (p) RDWT-SVD extracted watermark (Translate (x=10, y=10), NC=0.994). The Bridge image,
courtesy of pixabay.com (public domain), image ID: 1660417. The Doll image, courtesy of
pixabay.com (public domain), image ID: 1636128.

050021-8
 CONCLUSION
In this paper, two new blind image watermarking methods based on DWT-SVD and RDWT-SVD were presented.
Each of the schemes showed resistance against attacks. While DWT-SVD scheme presents more resistance against
some attacks such as JPEG compression, scaling and filtering, the RDWT-SVD scheme performed well against the
remaining attacks that mentioned in this paper. From the result’s table, the performances of the RDWT-SVD and
DWT-SVD are convergent. However, because of the ability of RDWT-SVD scheme to resist the geometric attacks
(except the scaling attack) and the capacity produced by RDWT without loss of imperceptibility. The RDWTSVD is
shown more robustness than DWT-SVD.

ACKNOWLEDGEMENTS

The authors would like to thank the anonymous reviewers for their helpful comments and suggestions. This work is
supported by the Universiti Malaysia Pahang (UMP) via Research Grant UMP RDU150353.

REFERENCES
1. Makbol, Nasrin M., Bee Ee Khoo, and Taha H. Rassem. "Block-based discrete wavelet transform-singular value
decomposition image watermarking scheme using human visual system characteristics." Image Processing,
IET10, no. 1: 34-52, (2016).
2. Makbol, Nasrin M., and Bee Ee Khoo. "A Hybrid Robust Image Watermarking Scheme Using Integer Wavelet
Transform, Singular Value Decomposition and Arnold Transform." Lecturer Notes of Computer Science,
Springer, pp. 36-47, 2013.
3. S. Lin and C.-F. Chen. “A robust DCT-based watermarking for copyright right protection”. IEEE Transactions
on Consumer Electronics, 46(3), pp. 415 –421, (2000).
4. P. Premaratne and C. Ko. A novel watermark embedding and detection scheme for images in DFT domain.
Proceedings of Seventh International Conference on Image Processing and Its Applications, 780 –783, (1999).
5. E. Ganic and A. M. Eskicioglu. Robust embedding of visual watermarks using discrete wavelet transform and
singular value decomposition. Journal of Electronic Imaging. 14(4), 043004, (2005).
6. T. D. Hien, Z. Nakao, and Y.-W. Chen. “Robust multi-logo watermarking by RDWT and ICA”. Signal
Processing, 86, 2981–2993, (2006)
7. S. Lagzian, M. Soryani, and M. Fathy. A new robust watermarking scheme based on RDWT-SVD. IJIIP:
International Journal of Intelligent Information Processing, 2(1), 22– 29, (2011).
8. K. Loukhaoukha and J.-Y. Chouinard. Hybrid watermarking algorithm based on SVD and lifting wavelet
transform for ownership verification. Proceedings of 11th Canadian Workshop on Information Theory, 177 –
182, (2009).
9. K.-L. Chung, W.-N. Yang, Y.-H. Huang, S.-T. Wu, and Y.-C. Hsu. “On SVD-based watermarking algorithm”.
Applied Mathematics and Computation, 188(1), 54–57, (2007)
10. S. Cui and Y.Wang. Redundant wavelet transform in video signal processing. Proceedings of International
Conference on Image Processing, Computer Vision, Pattern Recognition, 191–196, (2006).
11. K. feng He, J. Gao, L. mei Hu, and H. yuan Gao. “Watermarking for images using the HVS and SVD in the
wavelet domain”. Proceedings of the 2006 IEEE International Conference on Mechatronics and Automation,
2352 –2356, (2006).

050021-9