IET Image Processing

Research Article

ISSN 1751-9659
Block-based discrete wavelet transform- Received on 4th April 2014
Revised on 2nd June 2015
singular value decomposition image Accepted on 8th July 2015
doi: 10.1049/iet-ipr.2014.0965
watermarking scheme using human visual www.ietdl.org

system characteristics
Nasrin M. Makbol 1, Bee Ee Khoo 1 ✉, Taha H. Rassem 2
1
School of Electrical and Electronic Engineering, Universiti Sains Malaysia, Malaysia
2
Faculty of Computer Systems and Software Engineering, Universiti Malaysia Pahang (UMP), Malaysia
✉ E-mail: [email protected]

Abstract: Digital watermarking has been suggested as a way to achieve digital protection. The aim of digital watermarking
is to insert the secret data into the image without significantly affecting the visual quality. This study presents a robust
block-based image watermarking scheme based on the singular value decomposition (SVD) and human visual system
in the discrete wavelet transform (DWT) domain. The proposed method is considered to be a block-based scheme that
utilises the entropy and edge entropy as HVS characteristics for the selection of significant blocks to embed the
watermark, which is a binary watermark logo. The blocks of the lowest entropy values and edge entropy values are
selected as the best regions to insert the watermark. After the first level of DWT decomposition, the SVD is performed
on the low-low sub-band to modify several elements in its U matrix according to predefined conditions. The
experimental results of the proposed scheme showed high imperceptibility and high robustness against all image
processing attacks and several geometrical attacks using examples of standard and real images. Furthermore, the
proposed scheme outperformed several previous schemes in terms of imperceptibility and robustness. The security
issue is improved by encrypting a portion of the important information using Advanced Standard Encryption a key size
of 192-bits (AES-192).

1 Introduction spatial domain, the watermark is directly embedded in the pixel

values, whereas in the transform domain, the data are embedded
The rapid spread of the internet, accompanied by the comprehensive by modulating the magnitude of the transformed coefficients.
development of digital technologies and easily reproduced digital Although spatial domain approaches are advantageous because of
media, has increased the popularity of such media. The current their low complexity and easy implementation, transform domain
challenge is how to protect the ownership of digital products while approaches are preferred because they allow more information to
allowing the full usage of internet resources. Cryptography be embedded and provide greater robustness against attacks. The
provides a solution for this problem. However, cryptography does discrete cosine transform (DCT) and discrete wavelet transform
not completely address the relevant concerns because control of (DWT) are the most commonly adopted transforms in the digital
data dissemination is lost when the encryption is removed. The watermarking world. Singular value decomposition (SVD) is
best alternative solution to preserve the rights of the authors and widely used as a transform in watermarking because of its
ensure easy and fast access to internet resources is digital powerful properties [7, 15, 16]. However, SVD alone is not
watermarking. Digital watermarking is the process of permanently preferred because of its extensive computation. Therefore, hybrid
embedding data (adding a watermark) into digital multimedia SVD-based watermarking schemes have been developed. Many of
content (host data) without degradation, such that this watermark these schemes embed the watermark into the singular values (S) of
can resist any extraneous operation. The watermark can be visible the host image. These watermarks show high robustness against
or invisible; invisible watermarks are the most commonly the most common attacks, but might be vulnerable to attacks
employed. The basic characteristics of any watermarking scheme based on singular value substitution [17]. As a result, several
are its capacity, robustness, security and imperceptibility [1–4]. A SVD-based watermarking techniques embed watermarks into
conflict arises among the capacity, robustness and imperceptibility singular vectors (U or V) or both vectors instead of (S) [17] as in
in which increasing the capacity, improves robustness, but may Chang et al. [18], Chung et al. [19], Fan et al. [20] and Lai [21].
reduce the imperceptibility. In the signal processing field, to This eliminates the probability of the false positive problem that
improve detection, more information should be included to may occurs due to embedding into (S) [17]. Besides, the
improve the signal energy. Similarly, a robust watermarking watermarking techniques are categorised as robust, fragile or
technique is established by embedding larger payloads into the semi-fragile based on the applications and into blind, semi-blind
host data to gain the required robustness [5]. Several schemes that and non-blind based on the information required for watermark
employ different properties have been developed to meet the extraction or detection.
requirements of the intended application, and not all approaches Currently, establishing block-based image watermarking techniques
are appropriate for all applications [5]. For instance, robustness is of interest because of their advantages, one of which is the ability to
and a good imperceptibility must accompany techniques used for process each block individually. The basis of this technique is to
copyright protection. embed the watermark into the selected blocks, which are the blocks
Existing digital watermarking techniques in which the watermark or regions bearing the basic character information of the image,
is embedded can be categorised into two main groups: spatial such as the texture and edges. A human visual system (HVS) is
domain techniques and transform domain techniques [6–14]. In the adopted as a good method to select the blocks and to improve the

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 1
robustness and imperceptibility by exploiting the characteristics of the entropy and edge entropy. Each of the selected blocks is then
block, such as its entropy and edge entropy [21]. Chang et al. [18] processed individually to embed the watermark. DWT is applied as
proposed an SVD block-based scheme. A host image is a wavelet transform on each selected block. Thus, wavelet
decomposed into non-overlapping blocks, and an SVD is applied to properties are employed to achieve higher performance by the
each block. The complexity of each block is then evaluated by proposed scheme. DWT has gained popularity because of its
calculating all non-zero coefficients of the S component. Chang wavelet properties. This method requires less computation time than
et al. [18] proposed that the greater non-zero coefficients are the DCT, and DWT is more accurate in the modelling aspects of HVS
greater complex blocks. Therefore, for block-based schemes, these because of the multi-resolution property of any wavelet transform.
complex blocks are preferred in embedding to preserve the visual The low-low (LL), low-high (LH), high-low (HL) and high-high
quality [18]. Finally, the embedding is performed by testing the (HH) sub-bands are developed; LL is the approximation sub-band,
relation between the coefficients in the first column of U. This and LH, HL and HH are the details sub-bands. Only the LL
technique is robust against general image processing attacks. Chung sub-band (which holds the most significant information) is selected
et al. [19] proposed two ways to increase the imperceptibility and to embed a more robust and imperceptible watermark. As
capacity of SVD-based watermarking schemes. The first proposal determined by previously performed experiments [18–20], only the
indicated that modifying coefficients in the column vector of U coefficients in the first column of the U component can be modified
would cause less visible degradation than modifying the row in a robust SVD-based watermarking scheme. Therefore, an SVD is
vectors. The second proposal highlighted that modifying coefficients applied to each LL, and the watermark bit is then embedded by
in the row vector of V T would cause less visible degradation than examining the entries in the U matrix of the LL sub-band for each
modifying the column vectors. Chung et al. [19] employed these selected block. Unlike Lai [21], the relationship between the U2,1
proposals and examined their effectiveness on imperceptibility and and U3,1 is altered to embed the watermark bits. Many experiments
capacity. Initially, they embedded the watermark only in U, and have been performed by applying the proposed scheme, leading to
then they embedded the watermark in both U and V. The select U2,1 and U3,1 instead of U3,1 and U4,1 by Lai. Selecting U2,1
experimental results showed that embedding the watermark in both and U3,1 achieves better performance than previous studies [18–21].
U and V increases the imperceptibility and capacity compared with This better performance has been noted in experiments that have
embedding only in U as proposed by Chang et al. [18]. The compared the proposed scheme with previously published schemes
experimental results demonstrated that the proposals have both the [18–21]. The proposed scheme outperforms the previously reported
capacity and peak signal-to-noise ratio (PSNR) advantages of an schemes and exhibits high imperceptibility and good robustness
SVD-based watermarking scheme. Fan et al. [20] conducted several against different attacks (both signal processing attacks and
experiments based on the proposals of Chung et al. [19]. A geometrical attacks).
theoretical analysis was performed, and experimental results were The rest of this paper is organised as follows. Section 2 presents
investigated to adopt the first proposal (only the coefficients of the the features and properties of the proposed scheme and an
first column of U and V may be altered) and establish a robust overview of the methods utilised in the proposed scheme. Section
SVD-based watermarking scheme. The second proposal was to use 3 illustrates the watermark embedding and extraction procedures.
U and V to embed the watermark and to adopt U and V to Section 4 presents the experimental setup and results. Finally, the
compensate for visible distortion that occurs when the watermark is conclusions are stated in Section 5.
embedded in U and V. These experiments verified the effectiveness
of the proposed observations to improve the invisibility and
robustness of the SVD-based watermarking scheme, and the results 2 Related outlines of the proposed scheme
showed the better performance of the scheme proposed by Fan
et al. [20] relative to that proposed by Chung et al. [19]. Lai [21] 2.1 Features and highlights of the proposed scheme
presented an improved SVD block-based scheme. He partitioned the
host image into 8 × 8 non-overlapping blocks and selected the A new hybrid block-based image watermarking scheme is developed
appropriate blocks based on HVS characteristics. First, the HVS in this work. The blocks are selected based on several HVS
characteristics of entropy and edge entropy were evaluated for each characteristics, and the embedding process is achieved by
block. Then, the obtained entropy and edge entropy values were modifying the orthogonal matrix U. These adopted criteria
summed and sorted into ascending order, and the blocks with the significantly contributed to establishing a scheme with high
lowest magnitude values were selected. The number of selected robustness against attacks without affecting the visual quality of
blocks matched the number of embedded watermark bits. DCT was the image. The following are the highlights and features of the
then applied to each selected block, and an SVD was applied to all proposed scheme:
DCT transformed blocks. Accordingly, U, S and V matrices were
obtained for each block. The watermark bits were embedded by (i) The proposed scheme is a block-based scheme, in which the
examining the entries in the first column of the U matrix for each watermark affects specific regions of the image. Such schemes
block. Lai [21] noted that the relationship between the first column have several features, such as the following:
entries of the U component could be preserved, whereas other † Segmenting the host image into small blocks satisfies the
column entries were altered during general image processing un-detectability and imperceptibility requirements [17].
operations [21]. U3,1 and U4,1 were selected for modification to
embed the watermark bits. Experimental results and theoretical
analyses demonstrated the efficiency of the scheme proposed by Lai
[21] under various types of image manipulations. In addition, this
method displayed better performance compared to previous schemes
in the literature.
Our research area focuses on establishing a hybrid SVD-based
image watermarking scheme. In this paper, the advantages of the
previously mentioned schemes are considered to develop a new and
improved robust SVD block-based scheme to achieve both
imperceptibility and robustness. First, blocking is utilised, in which
the host image is divided into non-overlapping 8 × 8 blocks. These
blocks are reduced to a 4 × 4 grid after the application of DWT to
each selected block. The smaller segmentation is used to satisfy the
un-detectability and imperceptibility requirements. Similar to Lai Fig. 1 DWT of the Lena image
[21], only some of these blocks that match the number of a Original Lena image
watermark bits are selected by utilising the HVS characteristics b One-level DWT (the image is 512 × 512 pixels and each sub-band is 256 × 256 pixels)

IET Image Process., pp. 1–19

2 & The Institution of Engineering and Technology 2015
Fig. 2 Embedding steps

† The ability to handle each block separately allows multiple secret watermarking, ROI blocks are preserved, and the watermark
keys to be used for the selected blocks and improves security. is embedded in the remaining portion of the image; this
† The watermark capacity will vary between blocks according to portion is called the region of non-interest [22].
their properties, improving the robustness and imperceptibility. (ii) The scheme employs HVS characteristics (entropy and edge
† Block-based schemes are widely used in numerous entropy) to determine in which parts (blocks) of the image will
applications, such as region of interest (ROI) functions in produce un-noticeable distortion and achieve maximum robustness
medical images. To secure medical images through during embedding.

Fig. 3 Extracting steps

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 3
Fig. 4 Modified U matrix in the embedding process

(iii) DWT is applied to each selected block individually. The

advantages of using DWT (wavelet transform) instead of DCT,
which was adopted in Lai [21], are as follows:
† DWT has multi-resolution properties, leading to a
hierarchical analysis of the signal and the ability to perform
an analysis similar to HVS [23]. Wavelet transformations
more accurately model aspects of HVS compared to DCT [24]. Fig. 6 Watermark logos in the experiments
† DWT requires less computational time compared to DCT a Suggested watermark logo
because it does not suffer from blocking artefacts [23]. b Lai’s watermark logo

Fig. 5 Pepper and Lena Fig. 7 Examples of real-host images

a Host Lena image a Engineering campus image
b Host pepper image b Penang Bridge image

IET Image Process., pp. 1–19

4 & The Institution of Engineering and Technology 2015
 † The basic function of the DCT is f (x) = exp (iωx). Therefore, (v) Several of the required secret keys (X and Y coordinates of the
the DCT has perfect localisation in frequency only. Wavelets selected blocks) are encrypted using Advanced Encryption Standard
offer a trade-off between time/space and frequency/scale. (AES) with a key size of 192-bits.
Therefore, incorporating DWT in a watermarking scheme
produces a watermark with a spatially global and spatially
local support [23]. Therefore, watermarking techniques based 2.2 Mathematical preliminaries
on wavelets are more robust against attacks than those based
on DCT [23]. This section explains the essential mathematical concepts and main
(iv) An SVD is performed on the transformed coefficients of the terminology that are adopted in the proposed scheme.
LL sub-band of each block. The embedding process is performed
by examining the second and third entries of the first column of 2.2.1 HVS characteristics: Incorporating HVS models into
the U matrix instead of examining the third and fourth entries [as watermarking schemes assists in achieving the robustness and
previously proposed by Lai [21]. Embedding in SVD vectors (U imperceptibility requirements. By utilising the characteristics of
matrix)] will improve the robustness, invisibility and security. HVS to select the most suitable components to embed the

Fig. 8 PSNR values and visual perception comparison for the Lena watermarked image

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 5
watermark, no significant distortions are produced [25, 26]. The pixel and does not consider the co-occurrence of the pixel values.
HVS characteristics of entropy and edge entropy have been Accordingly, the edge entropy together with the entropy of each
employed to select the significant embedding regions, because image block is considered to identify the suitable regions in the
they carry important information about the image. The entropy image for embedding the watermark. A minimum number of edge
(average information) is used to measure the spatial correlation of points of the host image should be modified during watermark
neighbouring pixels [27]. The following mathematical formula is embedding. This process can be defined as follows [27]
used to calculate the entropy as defined by Shannon for a set of
n-elements [27]

n

n
Eedge entropy = pi expui = pi exp1− pi (2)
i=1 i=1


n
Eentropy = − pi log ( pi ) (1)
i=1
where ui = 1 − pi indicates the ignorance or uncertainty of the pixel
value. Many watermarking techniques that employ HVS are
pi represents the occurrence probability of an event i with 0 ≤ p ≤ 1 available [14, 18, 19, 21, 27].

and ni=1 pi = 1.
This process does not capture the real pictorial information of the 2.2.2 Discrete wavelet transform: DWT is widely used in
image/sub-image; rather, it only measures the entropy. The value many signal processing applications. The wavelet energy is
depends only on the probability distribution of the intensity of the concentrated in time and is more suitable for the transient analysis

Fig. 9 PSNR values and visual perception comparison for the pepper watermarked image

IET Image Process., pp. 1–19

6 & The Institution of Engineering and Technology 2015
IET Image Process., pp. 1–19
Table 1 BCRs for the Lena and pepper images under image processing attacks with different threshold values for the proposed scheme and Lai’s scheme
Attack type T = 0.04 T = 0.02 T = 0.012 T = 0.002

Lena Pepper Lena Pepper Lena Pepper Lena Pepper

Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15]

‘No attack’ 1.0000 0.9863 1.0000 0.9893 1.0000 0.9736 0.9893 0.9795 1.0000 0.9619 0.9766 0.9795 0.8418 0.9023 0.7861 0.8945

& The Institution of Engineering and Technology 2015

‘gaussian Noise 0.001’ 0.9092 0.5732 0.7910 0.5244 0.7559 0.5664 0.6738 0.4863 0.6533 0.5625 0.6113 0.4922 0.5381 0.5313 0.5186 0.4775
‘gaussian Noise 0.005’ 0.9033 0.5166 0.8135 0.5244 0.7490 0.5010 0.6865 0.5020 0.6533 0.4932 0.6279 0.4893 0.5420 0.4951 0.5508 0.4795
‘gaussian Noise 0.5’ 0.7354 0.5186 0.6758 0.5332 0.6211 0.4854 0.5781 0.5215 0.5752 0.4746 0.5488 0.5195 0.5000 0.4785 0.5049 0.5098
‘JPEG compression Q = 10’ 0.3555 0.3291 0.3350 0.3018 0.2705 0.2705 0.2832 0.2793 0.2705 0.2705 0.2822 0.2783 0.2705 0.2705 0.2803 0.2783
‘JPEG compression Q = 30’ 0.9951 0.6221 0.8496 0.6250 0.8467 0.7803 0.6855 0.6094 0.4502 0.3994 0.4814 0.4717 0.3145 0.2920 0.3691 0.3545
‘JPEG compression Q = 50’ 0.9980 0.6357 0.9238 0.6846 0.9883 0.6924 0.8525 0.7236 0.8711 0.8418 0.7578 0.7373 0.4043 0.3369 0.4668 0.4561
‘JPEG compression Q = 70’ 1.0000 0.7402 0.9746 0.8730 0.9971 0.8164 0.9414 0.8164 0.9932 0.6865 0.9121 0.8008 0.6182 0.6162 0.6504 0.6709
‘JPEG compression Q = 90’ 1.0000 0.9258 0.9971 0.9629 1.0000 0.9170 0.9834 0.9531 1.0000 0.9229 0.9697 0.9355 0.8350 0.8213 0.7676 0.7979
‘Sharpening 0.8’ 1.0000 0.4580 0.9824 0.6855 1.0000 0.6104 0.9658 0.7451 1.0000 0.7256 0.9492 0.7725 0.8154 0.7568 0.7588 0.6904
‘Median filter (3*3)’ 1.0000 0.9053 0.9385 0.8408 0.9980 0.9121 0.8936 0.8193 0.9941 0.9238 0.8418 0.7734 0.7500 0.7520 0.5752 0.6152
‘Median filter (5*5)’ 0.7842 0.8037 0.6875 0.7393 0.7617 0.8125 0.6279 0.7139 0.7129 0.7998 0.5723 0.6963 0.5801 0.6689 0.4414 0.5615
‘Median filter (7*7)’ 0.5205 0.6904 0.4238 0.6123 0.5303 0.7158 0.4238 0.5791 0.5186 0.6943 0.4170 0.5967 0.5117 0.6318 0.3711 0.5420
‘Gamma correction attack 0.8’ 1.0000 0.9805 1.0000 0.9902 1.0000 0.9697 0.9883 0.9736 1.0000 0.9551 0.9736 0.9580 0.8564 0.8916 0.7910 0.8740
‘Pepper and salt noise attack 0.3%’ 0.6348 0.4893 0.5684 0.5059 0.5439 0.5342 0.5518 0.4863 0.5430 0.5068 0.5283 0.5078 0.5273 0.5127 0.5254 0.5195
‘Pepper and salt noise attack 0.01%’ 0.9883 0.7969 0.9541 0.8115 0.9043 0.7842 0.8916 0.7900 0.8760 0.7539 0.8682 0.7725 0.7480 0.7393 0.7246 0.7412
‘Pepper and salt noise attack 0.001%’ 0.9990 0.9580 0.9980 0.9629 0.9873 0.9531 0.9795 0.9492 0.9795 0.9346 0.9619 0.9473 0.8242 0.8740 0.7773 0.8818
‘Pepper and salt noise attack 0.005%’ 0.9961 0.8818 0.9795 0.8838 0.9502 0.8564 0.9355 0.8594 0.9248 0.8662 0.9180 0.8662 0.8047 0.7900 0.7324 0.8018
‘Speckle noise attack 0.01’ 0.9922 0.6211 0.9893 0.7490 0.8779 0.5840 0.8730 0.6875 0.7373 0.5479 0.7852 0.6201 0.5781 0.5127 0.5957 0.5674
‘Speckle noise attack 0.04’ 0.8789 0.5361 0.8564 0.5859 0.7109 0.5264 0.7363 0.5420 0.6201 0.5225 0.6289 0.5225 0.5293 0.4814 0.5811 0.5303
‘Speckle noise attack 0.4’ 0.6328 0.5078 0.6484 0.5029 0.5508 0.5186 0.5537 0.4961 0.5146 0.4932 0.5566 0.5029 0.5254 0.4814 0.5156 0.4834
‘Adjust’ 1.0000 0.9834 0.9941 0.9785 1.0000 0.9697 0.9795 0.9717 1.0000 0.9561 0.9629 0.9678 0.8418 0.8984 0.7705 0.8799
‘Histogram equalisation attack’ 1.0000 0.9541 0.9902 0.9561 1.0000 0.9463 0.9756 0.8965 0.9990 0.9355 0.9639 0.8809 0.8193 0.8662 0.7959 0.7832

7
 8
Table 2 BCRs for the Lena and pepper images under geometrical attacks with different threshold values for the proposed scheme and Lai’s scheme
Attack type T = 0.04 T = 0.02 T = 0.012 T = 0.002

Lena Pepper Lena Pepper Lena Pepper Lena Pepper

Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15]

‘Centred cropping 20%’ 0.9980 0.9854 0.9873 0.9805 0.9980 0.9727 0.9766 0.9697 0.9980 0.9609 0.9639 0.9697 0.8408 0.9023 0.7773 0.8857
‘Centred cropping 50%’ 0.9648 0.9512 0.8643 0.8535 0.9648 0.9434 0.8535 0.8447 0.9648 0.9277 0.8408 0.8447 0.8105 0.8701 0.6875 0.7725
‘Cropping off 20%’ 0.9600 0.9434 0.9844 0.9746 0.9600 0.9307 0.9746 0.9648 0.9600 0.9248 0.9619 0.9648 0.8076 0.8682 0.7715 0.8809
‘Cropping off 50%’ 0.7979 0.7891 0.8145 0.8047 0.7979 0.7773 0.8047 0.7979 0.7979 0.7686 0.7939 0.7939 0.6855 0.7285 0.6553 0.7354
‘Cropping off (20, 30)’ 0.9209 0.9121 0.9561 0.9502 0.9209 0.8984 0.9463 0.9404 0.9209 0.8906 0.9316 0.9385 0.7813 0.8369 0.7510 0.8594
‘Cropping off (40, 40)’ 0.8457 0.8379 0.8887 0.8809 0.8457 0.8271 0.8789 0.8750 0.8457 0.8174 0.8682 0.8770 0.7227 0.7705 0.7021 0.8057
‘Cropping columns off 10%’ 0.9102 0.8926 0.9434 0.9316 0.9102 0.8809 0.9316 0.9248 0.9102 0.8760 0.9199 0.9229 0.7627 0.8281 0.7500 0.8408
‘Cropping columns off 50%’ 0.7305 0.7227 0.5547 0.5508 0.7305 0.7100 0.5469 0.5459 0.7305 0.7061 0.5420 0.5439 0.6328 0.6689 0.4648 0.5127
‘Cropping rows off 10’ 0.8154 0.8125 0.9346 0.9346 0.8154 0.7969 0.9238 0.9248 0.8154 0.7920 0.9111 0.9248 0.7012 0.7373 0.7334 0.8428
‘Cropping rows off 50’ 0.5703 0.5693 0.6582 0.6504 0.5703 0.5557 0.6484 0.6455 0.5703 0.5557 0.6406 0.6436 0.5156 0.5293 0.5400 0.6016
‘Shift 2’ 0.9990 0.9102 0.9717 0.8848 0.9980 0.9082 0.9424 0.8613 0.9844 0.8857 0.8945 0.8398 0.7520 0.7178 0.6816 0.6680
‘Shift 20’ 0.5674 0.5674 0.5811 0.5771 0.5664 0.5762 0.5791 0.5801 0.5693 0.5654 0.5693 0.5850 0.5371 0.5547 0.5615 0.5635
‘Cut attack 50’ 0.7734 0.7559 0.8574 0.8486 0.7734 0.7432 0.8496 0.8438 0.7734 0.7383 0.8379 0.8428 0.6631 0.6855 0.6748 0.7705
‘Cut attack 100 rows’ 0.6475 0.6455 0.6582 0.6523 0.6475 0.6299 0.6533 0.6514 0.6475 0.6270 0.6494 0.6504 0.5752 0.5898 0.5254 0.5996
‘Cut attack 100 columns’ 0.6338 0.6201 0.7373 0.7256 0.6338 0.6182 0.7324 0.7227 0.6338 0.6143 0.7266 0.7197 0.5352 0.5898 0.6104 0.6680
‘Rotation 45’ 0.4912 0.4883 0.4922 0.5264 0.4883 0.4814 0.4883 0.5176 0.4912 0.4941 0.4941 0.5215 0.4941 0.5029 0.4902 0.5264
‘Rotation 2’ 0.5303 0.5928 0.4707 0.5498 0.5303 0.5938 0.4854 0.5459 0.5313 0.5977 0.4951 0.5713 0.5381 0.5400 0.5107 0.5596
‘Rotation 70’ 0.4785 0.5420 0.5068 0.5264 0.4814 0.5391 0.5078 0.5264 0.4834 0.5420 0.5068 0.5283 0.4873 0.5371 0.5059 0.5361
‘Rotation 110’ 0.4307 0.5576 0.5068 0.5625 0.4248 0.5576 0.5146 0.5645 0.4238 0.5625 0.5146 0.5625 0.4268 0.5547 0.5156 0.5615
‘Rotation 50’ 0.5000 0.5234 0.4932 0.5322 0.4961 0.5303 0.4961 0.5313 0.4912 0.5352 0.4922 0.5273 0.5039 0.5371 0.4971 0.5264
‘Rotation −50’ 0.4805 0.5059 0.4512 0.5215 0.4854 0.5078 0.4453 0.5234 0.4883 0.5098 0.4434 0.5137 0.4902 0.5029 0.4482 0.5098
‘Translate attack (10, 10)’ 0.3984 0.6133 0.4922 0.5566 0.4111 0.5752 0.5029 0.5381 0.4199 0.5430 0.5020 0.5225 0.4473 0.5049 0.5117 0.5117
‘Translate attack (10, 20)’ 0.4395 0.5889 0.4697 0.5557 0.4453 0.5625 0.4727 0.5342 0.4580 0.5313 0.4834 0.5332 0.4648 0.5029 0.4990 0.5225
‘Translate attack (30, 40)’ 0.4434 0.4648 0.5010 0.4971 0.4443 0.4736 0.5039 0.5020 0.4521 0.4863 0.5127 0.5215 0.4619 0.5244 0.5186 0.5381
‘Shearing attack (1, 0.2)’ 0.4863 0.5439 0.4600 0.5381 0.4863 0.5557 0.4521 0.5303 0.4893 0.5605 0.4541 0.5547 0.5068 0.5459 0.4580 0.5645
‘Shearing attack (0.2, 1)’ 0.4863 0.5439 0.4600 0.5381 0.4863 0.5557 0.4521 0.5303 0.4893 0.5605 0.4541 0.5547 0.5068 0.5459 0.4580 0.5645
‘Scaling 0.5’ 1.0000 0.8701 0.9756 0.8965 0.9990 0.9043 0.9453 0.9033 0.9990 0.9209 0.9072 0.9092 0.7969 0.7979 0.6787 0.7520
‘Scaling 0.25’ 0.9951 0.3906 0.8584 0.4951 0.9219 0.4717 0.7754 0.5420 0.8193 0.5127 0.6885 0.5381 0.5674 0.5283 0.5439 0.5254

& The Institution of Engineering and Technology 2015

IET Image Process., pp. 1–19
of time varying signals. In nature, most signals are time varying. process can be expressed as follows
Therefore, wavelet transforms are suitable for many applications
[28]. As with any wavelet transform, a DWT is used to describe an

yhigh [k] = x[k]g[2k − n] (3)
image as small waves (called wavelets) of varying frequencies and n
limited durations. A DWT is a mathematical tool that provides
multiple-resolution analysis of an image. This tool decomposes the
and
image into two components: high frequency and low frequency. The

decomposition process is achieved by passing the signal through a ylow [k] = x[k]h[2k − n] (4)
series of high-pass filters to analyse the high frequencies and passing n
the signal through a series of low-pass filters to analyse the low
frequencies. Moreover, signals are analysed at different resolutions where yhigh and ylow are the outputs of the high-pass and low-pass
by using filters with different cut-off frequencies. By definition x[n] filters after sub-sampling by 2, respectively. They indicate DWT
is the original signal that extends across the frequency range of 0 to coefficients, and the process is a DWT forward process that can be
p rad/s. First, x[n] passes through a half-band high-pass filter (g[n]) repeated for further decomposition. IDWT is the inverse DWT,
and a low-pass filter (h[n]). Due to Nyquist’s rule, half of the which uses the DWT coefficients to reconstruct the original image.
samples can be eliminated after the filtering process. Accordingly, The IDWT can be expressed as follows
the signal now has a maximum frequency of π/2 radians instead of π
radians. Therefore, the signal can be sub-sampled by 2. This process

y[n] = (yhigh [k]g[ −n + 2k] + ylow [k]h[ −n + 2k]) (5)
represents a one-level DWT decomposition, mathematically, this n

Fig. 10 Visual quality comparison of the extracted watermarks for the Lena and pepper images for the proposed scheme and Lai’s scheme under image
processing attacks when T = 0.04

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 9
In two-dimensional applications, the DWT decomposes an image or 2.2.3 Singular value decomposition: SVD is a numerical
frames of a video into sub-bands (LL, HL, LH and HH). LL is the low technique that is widely used because of its properties, such as
frequency and denotes the approximation of an image, whereas HL, good stability (e.g. applying an SVD to an image does not
LH and HH are the high frequencies and denote the image details. noticeably affect the appearance when a small interfering signal
Each sub-band has half the size of the original image. For higher is added to the image). SVD is widely applicable in a lot of
decomposition levels, the sub-band size decreases by half until they signal and image applications, such as noise reduction, image
reach the required decomposition level. Fig. 1 shows a one-level hiding, image compression and image watermarking. In image
DWT of the Lena image. Embedding the watermark in the high processing applications, image A is a matrix of n × n entries.
frequency sub-bands of an image provides more imperceptibility Applying an SVD to A helps to extract the information of the
because they include the edge components, which are less sensitive image feature. An SVD decomposes the matrix into U, S and V
to change. However, embedding the watermark in the low matrices, where U and V describe the geometric properties of the
frequency provides greater robustness. image and S describes the luminance. These matrices can be

Fig. 11 Visual quality comparison of the extracted watermarks for the Lena and pepper images for the proposed scheme and Lai’s scheme under geometrical
attacks when T = 0.04

IET Image Process., pp. 1–19

10 & The Institution of Engineering and Technology 2015
IET Image Process., pp. 1–19
Table 3 BCRs for the examples of real-world images (Engineering campus image and Penang Bridge image) under image processing attacks with different threshold values for the proposed scheme and Lai’s
scheme

Threshold Engineering campus image Penang Bridge image

Attack type 0.04 0.02 0.012 0.002 0.04 0.020 0.012 0.002

Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15]

1.0000 0.8740 1.0000 0.8750 1.0000 0.8389 0.7969 0.7246 1.0000 0.7422 1.0000 0.5273 1.0000 0.8896 0.6553 0.7480

& The Institution of Engineering and Technology 2015

‘No_attack’
‘gaussian_Noise_0001’ 0.8027 0.5635 0.6719 0.5342 0.6211 0.5176 0.5420 0.4941 0.9561 0.5303 0.8232 0.5381 0.7012 0.5469 0.5273 0.5156
‘gaussian_Noise_0005’ 0.8086 0.5703 0.6729 0.5518 0.6133 0.5498 0.5469 0.5400 0.9600 0.5068 0.8174 0.4990 0.7012 0.4951 0.5322 0.4961
‘gaussian_Noise_0.5’ 0.7588 0.5557 0.6465 0.5381 0.6045 0.5342 0.5225 0.5068 0.3721 0.3662 0.3369 0.3613 0.3232 0.3584 0.3115 0.3506
‘JPEG_compression_Q10’ 0.3584 0.3057 0.3115 0.2861 0.3047 0.2852 0.2979 0.2852 0.4727 0.4219 0.2705 0.2705 0.2705 0.2705 0.2705 0.2705
‘JPEG_compression_Q30’ 0.9521 0.7373 0.7764 0.6895 0.6113 0.5225 0.4951 0.4424 1.0000 0.6279 1.0000 0.5781 0.6201 0.4521 0.2715 0.2705
‘JPEG_compression_Q50’ 0.9961 0.7842 0.9541 0.8320 0.8721 0.7500 0.6475 0.5732 1.0000 0.3467 1.0000 0.6201 1.0000 0.6328 0.5186 0.2939
‘JPEG_compression_Q70’ 1.0000 0.8281 0.9902 0.8506 0.9658 0.8340 0.7559 0.6729 1.0000 0.6807 1.0000 0.5713 1.0000 0.7168 0.4648 0.5166
‘JPEG_compression_Q90’ 1.0000 0.8623 1.0000 0.8652 0.9971 0.8613 0.7979 0.7266 1.0000 0.7158 1.0000 0.5234 1.0000 0.8555 0.6494 0.5723
‘sharp_08’ 1.0000 0.7031 0.9971 0.7686 0.9814 0.7510 0.7891 0.6240 1.0000 0.4727 1.0000 0.4727 1.0000 0.5518 0.6680 0.6592
‘Median_filter33’ 0.9951 0.9092 0.9678 0.9072 0.9141 0.8418 0.6895 0.6670 1.0000 0.6875 1.0000 0.5957 1.0000 0.7207 0.5537 0.6797
‘Median_filter55’ 0.7041 0.7813 0.6738 0.7500 0.6631 0.7002 0.5967 0.6074 0.6875 0.7324 0.6865 0.7354 0.6865 0.7461 0.4248 0.6582
‘Median_filter_77’ 0.6016 0.6748 0.6016 0.6602 0.6035 0.6563 0.5938 0.6064 0.5225 0.4902 0.5225 0.5049 0.5225 0.5898 0.4404 0.5791
‘Gamma_Correction_Attack_08’ 1.0000 0.8652 1.0000 0.8779 1.0000 0.8428 0.7949 0.7207 1.0000 0.7295 1.0000 0.6084 1.0000 0.8018 0.6484 0.6582
‘Pepper_and_Salt_Noise_Attack_03’ 0.5566 0.4980 0.5303 0.5127 0.5557 0.4873 0.5107 0.4971 0.6387 0.4961 0.5840 0.5088 0.5557 0.5117 0.5049 0.4912
‘Pepper_and _Salt_Noise_Attack_001’ 0.9521 0.7314 0.8594 0.7236 0.8691 0.6963 0.7188 0.6387 0.9961 0.6777 0.9326 0.5176 0.9053 0.7383 0.6133 0.6533
‘Pepper_and_Salt_Noise_Attack_0001’ 0.9971 0.8545 0.9834 0.8555 0.9863 0.8262 0.7832 0.7207 1.0000 0.7295 0.9932 0.5303 0.9902 0.8770 0.6523 0.7295
‘Pepper_and_Salt_Noise_Attack_0005’ 0.9795 0.8184 0.9248 0.7793 0.9170 0.7295 0.7578 0.6855 0.9980 0.7021 0.9541 0.5303 0.9473 0.7920 0.6436 0.6885
‘Speckle_Noise_Attack_001’ 0.9932 0.7207 0.8623 0.6807 0.7871 0.6201 0.5947 0.5430 0.9873 0.5234 0.8789 0.5020 0.7666 0.4980 0.5313 0.5107
‘Speckle_Noise_Attack_004’ 0.8555 0.5977 0.7354 0.5488 0.6406 0.5283 0.5596 0.5146 0.8535 0.5234 0.6855 0.4795 0.6299 0.4980 0.5059 0.5195
‘Speckle_Noise_Attack_04’ 0.6484 0.5156 0.5498 0.4844 0.5635 0.5127 0.5186 0.5029 0.5762 0.5332 0.5439 0.4971 0.4795 0.4932 0.5244 0.4990
‘Adjust’ 0.9932 0.8701 0.9902 0.8701 0.9883 0.8438 0.7822 0.7295 1.0000 0.7285 1.0000 0.5840 0.9980 0.7520 0.6523 0.7412
‘Histogram_Equalization_Attack’ 0.9922 0.8682 0.9922 0.8750 0.9863 0.8525 0.7734 0.7461 1.0000 0.7207 0.9980 0.6064 0.9951 0.7627 0.6055 0.5938

11
 12
Table 4 BCRs for the examples of real-world images (Engineering campus image and Penang Bridge image) under geometrical attacks with different threshold values for the proposed scheme and Lai’s scheme
Threshold Engineering campus image Penang Bridge image

Attack type 0.04 0.02 0.012 0.002 0.04 0.020 0.012 0.002

Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15] Our Lai [15]

‘Centered_Cropping_50’ 0.9199 0.8096 0.9199 0.8174 0.9199 0.7871 0.7373 0.6807 0.8428 0.6543 0.8428 0.4727 0.8428 0.7412 0.5928 0.6543
‘Cropping_off_20’ 0.9521 0.8359 0.9512 0.8457 0.9512 0.8086 0.7549 0.6943 0.9658 0.7207 0.9658 0.5059 0.9658 0.8643 0.6299 0.7217
‘Cropping_off_50’ 0.8232 0.7285 0.8232 0.7324 0.8232 0.7100 0.6631 0.6113 0.7568 0.5527 0.7568 0.3916 0.7568 0.6650 0.5000 0.5361
‘Cropping_off_2030’ 0.9209 0.8086 0.9209 0.8174 0.9209 0.7813 0.7295 0.6729 0.9512 0.7119 0.9512 0.4980 0.9512 0.8525 0.6221 0.7100
‘Cropping_off_4040’ 0.8896 0.7754 0.8896 0.7852 0.8896 0.7490 0.7090 0.6436 0.8242 0.6123 0.8242 0.4248 0.8242 0.7295 0.5459 0.5918
‘Cropping_columns_off_10’ 0.9043 0.8008 0.9043 0.7949 0.9023 0.7744 0.7256 0.6660 0.9150 0.6670 0.9150 0.4824 0.9150 0.8086 0.5986 0.6787
‘Cropping_columns_off_50’ 0.6777 0.6113 0.6777 0.6064 0.6777 0.5947 0.5645 0.5225 0.6846 0.5049 0.6846 0.3828 0.6846 0.5938 0.4668 0.5020
‘Cropping_raws_off_10’ 0.7998 0.7217 0.7998 0.7246 0.7998 0.6963 0.6289 0.6035 0.9141 0.6914 0.9141 0.5039 0.9141 0.8271 0.5947 0.6816
‘Cropping_raws_off_50’ 0.6553 0.5938 0.6553 0.5928 0.6553 0.5791 0.5205 0.5020 0.3721 0.3262 0.3721 0.2852 0.3721 0.3496 0.3115 0.3096
‘shift_2’ 0.9932 0.8555 0.9561 0.8174 0.9063 0.7412 0.7021 0.6377 1.0000 0.7402 1.0000 0.5967 1.0000 0.8115 0.6943 0.6650
‘shift_20’ 0.5986 0.5947 0.6143 0.5703 0.5996 0.5596 0.5889 0.5625 0.6152 0.5303 0.6064 0.4131 0.6035 0.5791 0.4785 0.5234
‘cut_attack_50’ 0.6836 0.6055 0.6836 0.6133 0.6836 0.5762 0.5693 0.5303 0.9277 0.6914 0.9277 0.5039 0.9277 0.8271 0.6055 0.6816
‘cut_attack_100_raws’ 0.5215 0.4658 0.5215 0.4785 0.5215 0.4424 0.4561 0.4199 0.7900 0.5908 0.7900 0.4365 0.7900 0.6982 0.5186 0.5742
‘cut_attack_100_columns’ 0.6768 0.6182 0.6738 0.6104 0.6729 0.5967 0.5605 0.5293 0.6777 0.5303 0.6777 0.4043 0.6777 0.6396 0.4629 0.5371
‘rotation_45’ 0.4512 0.5332 0.4512 0.5195 0.4502 0.5117 0.4541 0.5176 0.4961 0.5684 0.5049 0.5654 0.5176 0.5752 0.5234 0.5742
‘rotation_2’ 0.5264 0.5732 0.5303 0.5811 0.5400 0.5713 0.5518 0.5645 0.4678 0.5508 0.4795 0.5361 0.5020 0.5547 0.5479 0.6143
‘rotation_70’ 0.4453 0.5244 0.4443 0.5205 0.4443 0.5156 0.4404 0.5195 0.4805 0.5420 0.4697 0.5410 0.4678 0.5439 0.4570 0.5459
‘rotation_110’ 0.4541 0.5400 0.4570 0.5361 0.4551 0.5342 0.4551 0.5449 0.4854 0.5781 0.4785 0.5820 0.4785 0.5674 0.4658 0.5674
‘rotation_50’ 0.4258 0.5000 0.4287 0.4941 0.4307 0.4980 0.4316 0.5000 0.4922 0.5703 0.4990 0.5645 0.4990 0.5752 0.5049 0.5742
‘rotation__50’ 0.4619 0.4961 0.4648 0.5039 0.4609 0.5059 0.4619 0.5107 0.5088 0.5742 0.5205 0.5732 0.5283 0.5762 0.5342 0.5586
‘Translate_attack_10_10’ 0.4121 0.5811 0.4570 0.5557 0.4883 0.5527 0.5234 0.5498 0.4746 0.6533 0.4834 0.6465 0.4805 0.6270 0.5518 0.5703
‘Translate_attack_10_20’ 0.4316 0.5811 0.4521 0.5674 0.4824 0.5635 0.5078 0.5635 0.4795 0.6563 0.4814 0.6387 0.4971 0.6055 0.5371 0.5713
‘Translate_attack_30_40’ 0.4922 0.5273 0.5029 0.5430 0.5273 0.5654 0.5518 0.5684 0.4893 0.4658 0.4883 0.4766 0.4902 0.4854 0.5410 0.5674
‘Shearing_attack_1_02’ 0.5098 0.5371 0.5254 0.5459 0.5293 0.5508 0.5361 0.5488 0.5342 0.5400 0.5371 0.5430 0.5420 0.5449 0.5439 0.6006
‘Shearing_attack_02_1’ 0.5098 0.5371 0.5254 0.5459 0.5293 0.5508 0.5361 0.5488 0.5342 0.5400 0.5371 0.5430 0.5420 0.5449 0.5439 0.6006
‘Scaling_05’ 0.9980 0.8975 0.9785 0.8936 0.9297 0.8301 0.6768 0.6436 1.0000 0.7227 1.0000 0.7461 1.0000 0.7227 0.6318 0.6846
‘Scaling_025’ 0.8076 0.5244 0.7266 0.5527 0.6777 0.5723 0.6162 0.5645 0.9990 0.5098 0.9697 0.5293 0.8682 0.5244 0.4834 0.4795

& The Institution of Engineering and Technology 2015

IET Image Process., pp. 1–19
defined as follows Recently, embedding the watermark in the U and V components
instead of S has been adopted by Chang et al. [18], Chung et al.
A = U × S × VT (6) [19], Fan et al. [20], Lai [21] and our proposed scheme.

⎡ ⎤⎡ ⎤⎡ ⎤
u1,1 , . . . , u1,n s1,1 , 0, . . . , 0 v1,1 , . . . , v1,n T
⎢ u2,1 , . . . , u2,n ⎥⎢ ⎥⎢ . . . , v2,n ⎥ 3 Proposed image watermarking scheme
⎢ ⎥⎢ 0, s2,2 , . . . , 0 ⎥⎢ v2,1 , ⎥
A=⎢
⎢ . ⎥⎢
⎥⎢ . ⎥⎢
⎥⎢ . ⎥
⎥ (7)
⎣ . ⎦⎣ . ⎦⎣ . ⎦ A robust hybrid digital image watermarking scheme based on DWT,
SVD and HVS characteristics is presented in this work. The
un,1 , . . . , un,n 0, 0, . . . , sn,n vn,1 , . . . , vn,n
watermark embedding and extracting procedures, depicted in
Figs. 2 and 3, are presented in the subsequent subsections. Before
U and V are n × n orthogonal matrices; U TU = I, V TV = I and S = listing the watermarking steps, we highlight two important points.
diag(σ1,1, σ2,2, …, σn,n) that includes non-negative singular First, the scheme is a block-based scheme embedding the
values arranged in a descending order. These matrices can be watermark in specified regions (blocks); therefore, the image is
rectangular or square. Many hybrid digital image watermarking decomposed into (8 × 8) non-overlapping blocks. This size is
schemes use SVD with other transforms watermark. A lot of selected to obtain 4 × 4 blocks after performing a DWT
them explore the singular values, S; to embed the watermark. decomposition; this size is considered the smallest possible size

Fig. 12 PSNR values and visual perception comparison for the Engineering campus watermarked image

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 13
Fig. 13 PSNR values and visual perception comparison for the Penang Bridge watermarked image

that can be obtained in the proposed scheme. HVS characteristics The changing process in each block is based on examination of
are employed to select the desired blocks by calculating the the binary watermark bit at the location of the block (as
entropy and edge entropy using (1) and (2), respectively, for each explained in the following algorithm). This algorithm can
block. These two obtained values of each block are summed, and preserve the image quality and provide greater robustness
the resulting magnitude values are sorted in ascending order. It is against attacks, notably against JPEG attacks. The watermark
desirable to modify the minimum number of edge points of the embedding and extracting procedures are shown in Algorithm 1
host image during the watermark embedding [27]. Therefore, the (see Fig. 4).
blocks with the lowest magnitude values (called low informative
blocks) are selected as the best regions to insert the watermark
[21, 29]. The number of selected blocks is 1024, equal to the
watermark size (32 × 32). 3.1 Embedding procedure
The second point is that the embedding process is achieved by
extracting the algebraic properties of the image from the first Before displaying the embedding steps, the coordinates (x, y) of the
column of the U vector, and the watermark bits are embedded selected blocks are stored in two matrices denoted X and Y. These
by changing the relationship between the U2,1 and U3,1. coordinates represents the secret keys and have been encrypted
Empirically, selecting U2,1 and U3,1 in the first column of U to using AES-192.
embed a watermark bit achieves better results than Lai [21]
scheme, who employed U3,1 and U4,1 to embed the watermark. † Divide the host image into (8 × 8) non-overlapping blocks.

IET Image Process., pp. 1–19

14 & The Institution of Engineering and Technology 2015
† Sum the entropy and edge entropy for each block. Arrange the † Decrypt the encoded X and Y matrices to obtain the coordinates of
blocks in ascending order, and select the first 1024 blocks that the blocks that contained the watermark bits.
have the lowest values as the appropriate blocks in our proposed † Apply a DWT to the all target blocks.
scheme for watermark embedding. † Apply an SVD to the LL sub-band of each of these blocks.
† Apply a DWT to the selected blocks; four sub-bands LL, LH, HL † Extract the watermark bits by examining U2,1 and U3,1 of matrix
and HH are obtained for each block. U. If U2,1 is larger than U3,1, then the watermark bit = 1, otherwise
† Apply an SVD on the LL sub-band of each block. the watermark bit = 0.
† Examine the U2,1 and U3,1 entries of matrix U to embed the
watermark by updating the relationship. Algorithm 1 (Fig. 4) states
the steps to achieve this updating process. 4 Experimental setup and results
† Perform an inverse SVD.
† Perform an inverse DWT to obtain the watermarked image. A Several experiments are conducted to evaluate the validity and the
threshold value is used to control the trade-off between the performance of the proposed scheme. Two 512 × 512 grey-level
imperceptibility and the robustness. images (pepper and Lena) are used as host images. They are
shown in Fig. 5, whereas Fig. 6 shows two different 32 × 32
3.2 Extracting procedure binary images that are used as watermark logos in the
experiments. Moreover, two 512 × 512 real grey-level images
† Decompose the watermarked image into non-overlapping blocks selected from USM dataset [30] (Engineering campus and Penang
of 8 × 8 pixels. Bridge) are also used as host images. They are shown in Fig. 7.

Fig. 14 Visual quality comparison of the extracted watermarks for the real-world images (Engineering campus image and Penang Bridge image) for the
proposed scheme and Lai’s scheme under image processing attacks when T = 0.04

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 15
The experiments are divided into three parts. In the first part, the scheme. The Lai [21] binary watermark is shown in Fig. 6b. In
proposed scheme is evaluated using the host images (standard both experiments, different threshold values are used for
Lena and pepper images) with the watermark that is shown in evaluation. In this paper, the performance of the proposed scheme
Fig. 6a. In this part, Lai [21] scheme is re-implemented for and the other schemes are compared and investigated in terms of
comparison with the proposed scheme. imperceptibility and robustness against various attacks. Many
In the second part, examples of real images (Engineering campus criteria are suggested to estimate the imperceptibility and
and Penang Bridge) are used as host images with the watermark logo robustness. The most widely used criteria are the PSNR and the
that shown in Fig. 6a. Similar to first experiments, the results in this normalised cross-correlation (NC). However, in previous studies
part are compared with [21] after re-implemented. [15, 18–20], the bit correction rate (BCR) has been used to
In the third experiments, to avoid re-implementation, the Lai measure the robustness of the schemes. This criterion is mostly
scheme [21] and several previous schemes [18–20] employ the used in binary watermark logos, whereas (NC) is used for grey
same binary watermark logo that is used in Lai [21] in the host watermark logos [29]. Therefore, the BCR is used in this paper to
images, pepper and Lena, for comparison with the proposed measure the robustness.

Fig. 15 Visual quality comparison of the extracted watermarks for the real-world images (Engineering campus image and Penang Bridge image) for the
proposed scheme and Lai’s scheme under geometrical attacks when T = 0.04

IET Image Process., pp. 1–19

16 & The Institution of Engineering and Technology 2015
 The PSNR is used to estimate the imperceptibility. 4.1 Experiment part 1
Imperceptibility is a term used to evaluate the similarity between
the host image and the watermarked image and can be defined as In this part of the experiment, the proposed scheme and Lai’s scheme
follows [21] are tested using the watermark logo that is shown in Fig. 6a. Four
different threshold values are used: 0.04, 0.02, 0.012 and 0.002. Figs. 8

 and 9 show and compare the PSNR values and visual perceptions of
max(x(i, j))2
PSNR = 10log10 (8) the Lena and pepper watermarked images, respectively. The results
MSE in Figs. 8 and 9 show that the proposed scheme achieves a higher
imperceptibility and outperforms the imperceptibility of Lai’s
where i, j are the coordinates of each pixel of the host image x and the scheme. In addition to the imperceptibility, the proposed scheme is
mean-square error (MSE) between the host image x and the exposed to various types of image processing and geometrical
watermarked image y is defined as follows attacks to evaluate the robustness. The image processing attacks are
noise addition, (Gaussian noise attack, salt and pepper noise attack
1
m
n and speckle noise attack) and de-noising attacks (median filter
MSE = [x(i, j) − y(i, j)]2 (9) attack, Gaussian filter attack, gamma correction attack, JPEG
m∗n i=1 j=1 compression, sharpening attack, adjust attack and histogram
equalisation attack). The geometrical attacks include a cropping
When good imperceptibility is achieved, the watermarked image attack, shift attack, cut attack, rotation attack, translate attack,
appears nearly identical to the host image; in other words, we can shearing attack and scaling attack. Tables 1 and 2 show the BCR
say that the host image is not affected by the embedding process. values of the similarity between the original watermark and the
Generally, imperceptibility within 38 dB is the minimum accepted watermark that is extracted after the attacks. From the results, the
value of PSNR [31]. proposed scheme shows a higher robustness with high threshold
The BCR is a criterion used to measure the robustness by values, and it simultaneously preserves high imperceptibility. For
evaluating the similarity between the original watermark and the image processing attacks, the proposed scheme outperforms Lai’s
extracted watermark after an attack. When the BCR value is close scheme in all attacks when the threshold values are 0.04, 0.02 and
to 1 under applicable attacks, the scheme is robust against those 0.012, except in median filter 5 × 5 and 7 × 7 attacks and adjust attack
attacks. The BCR can be estimated as follows when the threshold value is 0.012. For geometrical attacks, the
cropping, scaling, cut and shifting attacks outperforms Lai’s scheme
M N when T = 0.04, 0.02 and 0.012 (Lena image only). However, the
i=1  ij
j=1 wij ⊗ w scheme proposed by Lai [21] performs best in rotation, shearing and
) =
BCR(w, w (10)
M ∗N translate attacks for all T values and approximately all attacks with T
= 0.002 and T = 0.012 (pepper image only). Furthermore, the
where N and M represent the number of pixels in the watermark and extracted watermarks for the two schemes are shown in Figs. 10 and
 indicate the original and the extracted watermarks, respectively.
w, w 11 for visual quality comparison.

Table 5 Comparing the PSNR values for the host images, Lena and pepper, under different threshold values
Image Scheme T = 0.002 T = 0.012 T = 0.02 T = 0.04 Average

Lena Chang et al. [18] 48.80 48.02 46.90 43.74 46.86

Chung et al. [19] 50.17 47.83 45.94 42.04 46.49
Fan et al. [20] 48.91 48.12 46.98 43.81 46.95
Lai [21] 61.69 49.37 44.75 38.51 48.58
our scheme 61.31 51.54 47.32 41.39 50.39
Pepper Chang et al. [18] 45.71 45.26 44.61 42.44 44.51
Chung et al. [19] 50.86 48.37 46.34 42.24 46.95
Fan et al. [20] 45.87 45.41 44.75 42.54 44.64
Lai [21] 56.20 50.20 45.73 39.61 47.94
our scheme 57.94 52.06 48.55 43.01 50.39

Table 6 Comparing the BCR values of the extracted watermark for the host images, Lena and pepper, under different attacks when 0.012 and 0.04
thresholds

Threshold Image Scheme Cropping off 25% Gauss. noise 3% Median filter [3 × 3] JPEG comp. 70% Sharpening Average

0.012 Lena Chang et al. [18] 0.8115 0.6230 0.5283 0.6806 0.8847 0.7056
Chung et al. [19] 0.8115 0.7080 0.4980 0.6855 0.9589 0.7324
Fan et al. [20] 0.8115 0.6250 0.5292 0.6806 0.8867 0.7066
Lai [21] 0.8476 0.5928 0.9688 0.9941 0.9990 0.8805
our scheme 0.9307 0.6504 0.9961 0.9932 0.9999 0.9141
Pepper Chang et al. [18] 0.8310 0.6435 0.4794 0.6943 0.8837 0.7064
Chung et al. [19] 0.8330 0.6611 0.4628 0.6562 0.9365 0.7099
Fan et al. [20] 0.8311 0.6464 0.4833 0.6933 0.8828 0.7074
Lai [21] 0.7969 0.5908 0.8369 0.9160 0.9316 0.8145
our scheme 0.9385 0.6152 0.8340 0.9063 0.9434 0.8475
0.04 Lena Chang et al. [18] 0.8125 0.8750 0.5507 0.9687 0.9873 0.8388
Chung et al. [19] 0.8125 0.9179 0.5107 0.9794 0.9892 0.8419
Fan et al. [20] 0.8125 0.8740 0.5468 0.9658 0.9873 0.8373
Lai [21] 0.8477 0.9521 1.0000 1.0000 1.0000 0.9599
our scheme 0.9307 0.9053 1.0000 1.0000 1.0000 0.9672
Pepper Chang et al. [18] 0.8359 0.8720 0.4921 0.9511 0.9667 0.8236
Chung et al. [19] 0.8359 0.8896 0.4726 0.9443 0.9628 0.8211
Fan et al. [20] 0.8359 0.8720 0.4941 0.9472 0.9658 0.8230
Lai [21] 0.8271 0.8544 0.9541 0.9609 0.9717 0.9136
our scheme 0.9629 0.7988 0.9375 0.9717 0.9775 0.9297

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 17
4.2 Experiment part 2 good resiliency was observed against all types of image processing
attacks and several types of geometrical attacks. Unlike most
In this part, the proposed scheme and the Lai’s scheme are tested SVD-based watermarking schemes that embed the watermark into
with real-world host images. These images are collected from the singular values S, in this proposed scheme, the watermark was
USM dataset [30]. Two 512 × 512 images which are Engineering embedded by examining elements in U. This process avoided the
campus and Penang Bridge images are used with the watermark false positive problem and provided a reliable SVD-based scheme.
logo that is shown in Fig. 6a. Four different threshold values are The proposed scheme achieved the security requirement.
used: 0.04, 0.02, 0.012 and 0.002. Tables 3 and 4 show the BCR Moreover, the scheme was blind, and the important keys for
values of the similarity between the original watermark and the the extraction process were encrypted using AES-192 to maintain
watermark that is extracted after the attacks using different the information strictly confidential. This process improved the
threshold values. These results proved that the proposed scheme security of the scheme. The experiments are performed using
outperforms Lai’s scheme in many attacks. The visual perception examples of standard images and real-world images.
of both real images is shown in Figs. 12 and 13. Fig. 12 shows Our further research efforts will focus on improving the efficiency
the Engineering campus image visual perception with different of the robustness against some geometrical attacks; which is the
values of the threshold using the proposed scheme and the Lai’s limitation of the proposed scheme, by using more HVS
scheme. While Fig. 13 shows the Penang Bridge image visual characteristics and selecting more appropriate embedding regions.
perception. In both figures, the proposed scheme outperforms Lai’s
scheme in imperceptibility. Similar to experiments in part 1, the
proposed scheme and the Lai’s scheme are tested with the real 6 Acknowledgments
images with different types of attacks to evaluate the schemes
robustness. The extracted watermarks for the two schemes are The authors thank the anonymous reviewers for their helpful and
shown in Figs. 14 and 15 for visual quality comparison. constructive comments that greatly contributed to improve the
quality of the paper.
4.3 Experiment part 3

In this experiment, the proposed scheme is tested with the same 7 References
watermark logo that is used previously [21]. The same host 1 Gerhard, C., Setyawan, I., Lagendijk, R.: ‘Watermarking digital image and video
images, Lena and pepper, are used with identical threshold values. data’, IEEE Signal Process. Mag., 2000, 17, (5), pp. 20–46
Table 5 shows the PSNR of the imperceptibility of all schemes 2 Chang, C., Lin, P., Yeh, J.: ‘Preserving robustness and removability for digital
under different threshold values. The imperceptibility of the watermarks using subsampling and difference correlation’, Inf. Sci., 2009, 179,
(13), pp. 2283–2293
proposed scheme is best, and our scheme outperforms all others 3 Huang, C., Wu, J.: ‘Fidelity-guaranteed robustness enhancement of blind-detection
under all threshold values with both host images, Lena and watermarking schemes’, Inf. Sci., 2009, 179, (6), pp. 791–808
pepper. In the robustness comparison, only threshold values of 4 Liu, K.C., Chou, C.H.: ‘Robust and transparent watermarking scheme for colour
0.04 and 0.012 are tested. Table 6 shows the similarity between images’, IET Image Process., 2009, 3, (4), pp. 228–242
5 de Vugt, F.: ‘Quantization-watermarking methods for digital audio’, 2001
the extracted watermark and the original watermark in terms of 6 Ruanaidh, J., Pun, T.: ‘Rotation, scale and translation invariant digital image
BCR. Only five different attacks that are exposed to our proposed watermarking’. Proc. IEEE Int. Conf. Image Processing, 1997, pp. 536–539
scheme are similar to other schemes. These are the cropping attack 7 Makbol, N.M., Khoo, B.E.: ‘Robust blind image watermarking scheme based on
(cropping of 25%), Gaussian noise 3%, median filter [3 × 3], JPEG redundant discrete wavelet transform and singular value decomposition’,
AEU-Int. J. Electron. C, 2013, 67, (2), pp. 102–112
compression attack (Q = 70%) and sharpening attack. At a 8 Chang, C., Lin, C., Tseng, C., et al.: ‘Reversible hiding in DCT-based compressed
threshold value = 0.04, the proposed scheme performs better than images’, Inf. Sci., 2007, 177, (13), pp. 2768–2786
other schemes, except in the median filter attack for the Lena 9 Cox, I., Kilian, J., Leighton, F., et al.: ‘Secure spread spectrum watermarking for
image and in the median filter and Gaussian noise attack for the multimedia’, IEEE Trans. Image Process., 1997, 6, (12), pp. 1673–1687
10 Li, Q., Cox, I.: ‘Using perceptual models to improve fidelity and provide resistance
pepper image. At a threshold value = 0.04, the proposed scheme to valumetric scaling for quantization index modulation watermarking’, IEEE
performs better than other schemes, except in the median filter Trans. Inf. Forensics Sec., 2007, 2, (2), pp. 127–139
attack for the Lena image and in the median filter and Gaussian 11 Aslantas, V., Ozer, S., Ozturk, S.: ‘Improving the performance of DCT-based
noise attack for the pepper image. Under the other threshold fragile watermarking using intelligent optimization algorithms’, Opt. Commun.,
2009, 282, (14), pp. 2806–2817
values (T = 0.012), the proposed scheme is the best for the Lena 12 Li, L., Yuan, X., Lu, Z., et al.: ‘Rotation invariant watermark embedding based on
image in the sharpening and cropping attacks and in the scale-adapted characteristic regions’, Inf. Sci., 2010, 180, (15), pp. 2875–2888
sharpening, cropping and median filter attacks for the pepper host 13 Lin, T., Lin, C.: ‘Wavelet-based copyright-protection scheme for digital images
image. Based on the average of the BCR values of all attacks, our based on local features’, Inf. Sci., 2009, 179, (19), pp. 3349–3358
14 Maity, S., Kundu, M., Das, T.: ‘Robust SS watermarking with improved capacity’,
proposed scheme is the best when T = 0.04 and T = 0.012. Pattern Recognit. Lett., 2007, 28, (3), pp. 350–356
15 Lai, C.C., Tsai, C.C.: ‘Digital image watermarking using discrete wavelet
transform and singular value decomposition’, IEEE Trans. Instrum. Meas., 2010,
59, (11), pp. 3060–3063
5 Conclusions 16 Aslantas, V.: ‘An optimal robust digital image watermarking based on SVD using
differential evolution algorithm’, Opt. Commun., 2009, 282, (5), pp. 769–777
In this paper, a DWT–SVD block-based image watermarking 17 Basso, A., Bergadano, F., Cavagnino, D., et al.: ‘A novel block-based
watermarking scheme using the svd transform’, Algorithms, 2009, 2, (1), pp. 46–75
scheme is presented. Several characteristics were employed to 18 Chang, C.C., Tsai, P., Lin, C.C.: ‘SVD-based digital image watermarking scheme’,
achieve high-level grades for the watermarking requirements and Pattern Recognit. Lett, 2005, 26, (10), pp. 1577–1586
maintain the trade-off between them. Initially, blocking was used 19 Chung, K.L., Yang, W.N., Huang, Y.H., et al.: ‘On SVD-based watermarking
to divide the image into blocks. Then, only a portion of these algorithm’, Appl. Math. Comput., 2007, 188, (1), pp. 54–57
20 Fan, M.Q., Wang, H.X., Li, S.K.: ‘Restudy on SVD-based watermarking scheme’,
blocks were selected to include the watermark; this selection Appl. Math. Comput., 2008, 203, (2), pp. 926–930
ensured that the embedding process would affect specific regions 21 Lai, C.C.: ‘An improved SVD-based watermarking scheme using human visual
of the image. The HVS characteristics of entropy and edge entropy characteristics’, Opt. Commun., 2011, 284, (4), pp. 938–944
were used to select the low informative blocks as the best 22 Rathi, S., Inamdar, V.: ‘Medical images authentication through watermarking
preserving ROI’, Healthcare Inf., Int. J. (HIIJ), 2012, 1, (1), pp. 27–42
embedding regions. The scheme employed the properties of a 23 Serdean, C.V., Tomlinson, M., Wade, J.G, Ambroze, M.: ‘Protecting intellectual
DWT and SVD. These methods aimed to provide high robustness rights: Digital watermaking in the wavelet domain’. Proc. IEEE Int. Workshop
by selecting the most robust regions with an emphasis on Trends and Recent Achievements in Information Technology, 2002
maintaining non-noticeable distortions, in other words, to maintain 24 Liu, L.: ‘A survey of digital watermarking technologies’. Tech. Rep., Stoney Brook
University, USA, 2005
imperceptibility. 25 Padmanabha Reddy, Varadarajan, S.: ‘Human visual system sentient imperceptible
The proposed scheme outperformed all similar previous schemes and efficient wavelet-based watermarking scheme for copyright protection of
and achieved the highest imperceptibility. In terms of robustness, digital images’, IJCSNS Int. J. Comput. Sci. Netw. Sec., 2009, 9, (4), pp. 255–264

IET Image Process., pp. 1–19

18 & The Institution of Engineering and Technology 2015
26 Lin, Y.T., Huang, C.Y., Lee, G.: ‘Rotation, scaling, and translation resilient 29 Maity, S., Kundu, M.: ‘Perceptually adaptive spread transform image watermarking
watermarking for images’, IET Image Process., 2011, 5, (4), pp. 328–340 scheme using hadamard transform’, Inf. Sci., 2011, 181, (3), pp. 450–465
27 Maity, S.P., Kundu, M.K.: ‘Dht domain digital watermarking with low loss in 30 Rassem, T.H.: ‘New histogram-based interest point detectors, and new appearance
image informations’, AEU-Int. J. Electron. C, 2010, 64, (3), pp. 243–257 and texture image descriptors for object class recognition’. Ph.D. thesis, School of
28 Kashyap, N., Sinha, G.: ‘Image watermarking using 3-level discrete wavelet Electrical and Electronic Engineering, Universiti Sains Malaysia (USM), 2014
transform (dwt)’, Int. J. Modern Educ. Comput. Sci. (IJMECS), 2012, 4, (3), 31 Lee, Y.P., Lee, J.C., Chen, W.K., et al.: ‘High-payload image hiding with quality
p. 50 recovery using tri-way pixel-value differencing’, Inf. Sci., 2012, 191, pp. 214–225

IET Image Process., pp. 1–19

& The Institution of Engineering and Technology 2015 19