Computer Standards & Interfaces 32 (2010) 54–60

Contents lists available at ScienceDirect

Computer Standards & Interfaces

j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / c s i

Improving the robustness of DCT-based image watermarking against

JPEG compression
Shinfeng D. Lin a, Shih-Chieh Shie b,⁎, J.Y. Guo a
a
Department of Computer Science and Information Engineering, National Dong Hwa University, Hualien County, Taiwan, ROC
b
Department of Computer Science and Information Engineering, National Formosa University, 64 Wen-Hua Road, Hu-Wei, Yunlin County 632, Taiwan, ROC

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

Article history: A DCT-based image watermarking technique is proposed in this article. To improve the robustness of
Received 1 February 2007 watermark against JPEG compression, the most recently proposed techniques embed watermark into the
Received in revised form 4 April 2009 low-frequency components of the image. However, these components hold significant information of the
Accepted 23 June 2009
image. Directly replacing the low-frequency components with watermark may introduce undesirable
Available online 3 July 2009
degradation to image quality. To preserve acceptable visual quality for watermarked images, we propose a
watermarking technique that adjusts the DCT low-frequency coefficients by the concept of mathematical
Keywords:
Image watermarking
remainder. Simulation results demonstrate that the embedded watermarks can be almost fully extracted
JPEG compression from the JPEG-compressed images with very high compression ratios.
Robustness © 2009 Elsevier B.V. All rights reserved.
Torus automorphism

1. Introduction literature. One is image authentication technique and the other is

image data hiding. The goal of image authentication is to verify the
Along with the progress relating to computer hardware and originality of an image by detecting malicious manipulations.
software, the Internet has become the most popular channel for However, most of the earlier methods for image authentication deal
transmitting various forms of digital media. Since the environment of with all types of manipulations equally and unacceptably. That is these
the Internet is open, the protection of digital images transmitted on methods treat some practical manipulations such as image compres-
the network has become an important research topic in recent years. sion and image enhancement as attacks. Therefore, Lin and Chang
Digital watermarking is a common technique to achieve copyright proposed a DCT-based image authentication method which can
protection. For digital images, watermarking is the process of prevent malicious manipulations but allow JPEG lossy compression
embedding significant data (watermark) into an image such that [4]. In Lin and Chang's method, the authentication signatures for
the embedded watermark can be detected or extracted later to make images are generated based on the invariance of the relationships
an assertion about the image [1,2]. between DCT coefficients at the same position in separate blocks of an
Generally, a watermarking scheme consists of three parts, the image. More detailed theoretical analysis and experimental results
watermark, the watermark embedding stage and the watermark can be found in [4]. The purpose of image data hiding is different from
verification stage. The watermark embedding algorithm incorporates traditional cryptography. Cryptography concentrates on encrypting
the watermark into the host image, whereas the verification algorithm meaningful messages into meaningless data while image data hiding
extracts and authenticates the watermark determining the ownership covers secret information with the host image as camouflage. Hiding
of the image. Usually, the watermark is a visually recognizable logo or data in images involves embedding a large amount of secret data into
a set of meaningless character strings that represents the copyright of a cover image with minimal perceptible degradation of image quality.
the owner or legal users. If a watermark can be extracted from an However, the hiding capacity for secret data and the distortion of the
image in the verification stage, it could prove the copyright of the cover image are a tradeoff since more hidden data always result in
owner [3]. more degradation on the cover image.
In the past few years, two kinds of techniques that are related to Digital watermarking has been shown as a valid solution to the
but different from image watermarking have been introduced in problem of image copyright protection [5,6]. There are two essential
requirements for image watermarking. One is invisibility, namely the
watermarked image should not be perceived the changes of the
⁎ Corresponding author. original image by human eyes. It means that the visual quality of
E-mail address: [email protected] (S.-C. Shie). watermarked images should not be destroyed by the embedding of

0920-5489/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.csi.2009.06.004
 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60 55

watermark. The other is robustness, namely the watermark should be applying the YUV color transformation, as shown in Eq. (2), on the
able to resist attacks, even if these attacks are deliberately made [7]. To original RGB image. The two reasons for embedding watermark in Y
improve the robustness of watermark, most of the recently proposed (luminance) rather than U or V (chrominance) are: (i) human visual
techniques embed watermark information into the low-frequency system is more sensitive to luminance than to the other two
part of images [8–10]. Huang et al. proposed a novel DCT-based chrominance components, and (ii) the JPEG and MPEG standards
technique which embeds watermarks in the DC components of images typically use higher density for Y than for the other two components.
[8]. The main idea behind this technique is that more robustness can 0 1 0 1 0 1
be achieved if watermarks are embedded in DC components since DC Y 0:299 0:587 0:114 R
@ U A = @ −0:148 −0:289 0:437 A × @ G A ð2Þ
components have much larger perceptual capacity than ac compo-
nents. In addition, the authors took advantages of the feature of spatial V 0:615 −0:515 −0:100 B
masking (both luminance and texture masking) of the HVS to
adaptively embed watermarks into images. For digitized images to Step 3. DCT transformation and quantization: Discrete Cosine
be safely and efficiently transmitted on the Internet, watermarked Transform (DCT) is a kind of signal decomposition that converts
images should be particularly robust to JPEG compression. Recently, images from spatial domain to frequency domain. The equation of
Lin et al. proposed two kinds of DCT-based image watermarking forward DCT regarding an image block of size N × N pixels is given
techniques [9,10]. These schemes perform well under general JPEG in Eq. (3). Note that DCT(i, j) represents the coefficient at coordinate
compression. However, when the watermarked images have to be (i, j) in the DCT-transformed block and pixel(x, y) is the pixel value at
compressed with higher compression ratio, the embedded water- coordinate (x, y) in the original block.
marks may be destroyed seriously. To overcome this problem, an
NX
− 1 NX
−1
improved technique has been studied and proposed in this paper.
DCTði; jÞ = C ðiÞ × C ð jÞ × pixelðx; yÞ ð3Þ
The rest of this paper is organized as follows. The proposed DCT- x=0 y=0
based watermark embedding and extracting algorithms are described    
in Sections 2 and 3, respectively. The relative parameters for the ð2x + 1Þiπ ð 2y + 1Þjπ
× cos cos
proposed watermarking scheme are provided in Section 4. Section 5 2N 2N
addresses our experimental results and discussion, especially on JPEG
compression with very high compression ratios. Finally, the conclu- where
sions are given in Section 6. 8 rffiffiffiffi
>
> 1
>
< N for i; j = 0
2. Watermark embedding algorithm C ðiÞ; C ð jÞ = rffiffiffiffi
>
>
: 2
> for i; j = 1; 2; N ; N − 1
The detailed steps of the proposed watermark embedding N
algorithm are given as follows.
Step 1. Torus automorphism permutation: in order to increase the After color transformation, the luminance Y of image is divided
security and robustness of watermark, the watermark pattern should into non-overlapping blocks of size 8 × 8, and each block of Y is DCT
be disarranged before embedding. Torus automorphism (TA) is an transformed independently. In order to enhance the robustness of
effective method to disperse a watermark equally and randomly [11]. watermark against JPEG compression, each DCT-transformed block of
Applying the concept of TA for scrambling the binary watermark Y is quantized based on the quantization table provided by the JPEG
before it is embedded into the host image offers cryptographic compression standard as shown in Fig. 1.
protection against intentional reconstruction of watermark [12]. This Step 4. Embedded blocks selection: to improve the invisibility of
is because the key utilized in the TA permutation procedure (for watermark, the complex blocks of images are selected for watermark
scrambling the watermark) is also necessary in the inverse TA embedding. In the DCT domain, the number of non-zero coefficients in
permutation procedure (for reconstructing the watermark). The key each block is computed to estimate the complexity of the image block.
is held only by the distributor of the watermarked image. Without the Let N(n) represent the number of non-zero coefficients in block n. The
key, the attacker cannot determine the original permutation of image blocks can be sorted into descending order based on their
watermark bits. Consequently, it is difficult for the attacker to detect corresponding N(n)s.
and reconstruct the watermark pattern. The details of the Torus Step 5. Watermark insertion: in the proposed scheme, the binary
automorphism permutation can be found in [13]. information of watermark pattern W is sequentially embedded into
The watermark used in the proposed scheme is permuted based on the selected image blocks obtained in step 4. Two bits of watermark
the following equation before it is embedded into the host image. pattern are embedded into the low-frequency components of each



 selected block for preserving the visual quality of watermarked
i⁎ 1 1 i
= × mod m ð1Þ
j⁎ k k+1 j

Eq. (1) means that each pixel of the watermark pattern at co-
ordinate (i, j) is moved to (i⁎, j⁎), and k and m are the key parameters
given by the user. Let Wp be the permuted watermark after Torus
automorphism permutation, then divide Wp into non-overlapping
pairs.
Step 2. YUV color transformation: RGB color space is highly
correlated and is not suitable for watermarking applications, except
for the blue channel utilized by some researchers. However, the
potential of the RGB channels can be exploited for watermarking, by
decreasing the correlation among them. In the proposed scheme, we
adopt the YUV color space for watermark embedding. The watermark
information is embedded into the luminance components of images.
The luminance information Y of the color image is obtained by Fig. 1. The quantization table recommended in the JPEG specification.
56 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60

images and increasing the robustness of watermark. Here, the low-

frequency components of image block can be the DCT coefficients at
positions C(2, 0), C(1, 1), C(0, 2), C(0, 3), C(1, 2), C(2, 1), and C(3, 0).
The positions of DCT coefficients within an image block are shown as Fig. 3. Watermark pattern used in the experiments.
Fig. 2.
To embed watermark information, we have to modify the DCT
coefficients at the low-frequency positions. Let C be the value of DCT worst condition, the largest difference between C and C⁎ is M/2.
coefficient and M be the modulus, the other relative variables can be Cooperated with the watermark extracting algorithm, the proposed
definded as the following formulas watermark embedding rule provides a safe range for the value of C⁎.
This means that even the value of C⁎ is changed into a new value C⁎⁎
r = jC jmod M with a difference as large as M (due to image processing attacks) the
jC j embedded watermark bit can be still extracted successfully.
q= ð4Þ Step 6. Inverse DCT transformation: after all the watermark bits are
M(
1 ; if Cz0 inserted into the selected image blocks, each block of the host image is
sign = inverse DCT transformed independently. The equation of inverse DCT
−1 ; if Cb0
is given in Eq. (5).

In Eq. (4), r is the mathematical remainder of |C| and r ∈ {0,1,2,…, − 1 NX

NX −1
M − 1}, q is the mathematical quotient obtained by dividing |C| by pixelðx; yÞ = C ðiÞ × C ð jÞ × DCTði; jÞ ð5Þ
M, and sign represents that C is positive or negative. i=0 j=0
   
The binary information of watermark pattern is embedded into the ð2x + 1Þiπ ð 2y + 1Þjπ
selected block by changing the value of coefficient C. Let C⁎ be the × cos cos
2N 2N
modified coefficient corresponding to C. The value of C⁎ is obtained by
the following rule. where
If watermark bit = 0:
8 rffiffiffiffi
>
> 1
M >
< N for i; j = 0
rV= C ðiÞ; C ð jÞ = rffiffiffiffi
4 >
> 2
>
Clow = sign × ðq × M + r VÞ : for i; j = 1; 2; N ; N − 1
N
Chigh = sign × ððq + 1Þ × M + r VÞ
(
Clow ; if j Clow − C j Vj Chigh − C j Step 7. Weighted correction: to further improve the visual quality
⁎
C = of the watermarked image, the weighted correction procedure is
Chigh ; if j Clow − C j Nj Chigh − C j
applied and it makes a slight change on the watermarked (modified)
luminance Y⁎. Let D be the difference between the original luminance
If watermark bit = 1: Y and the watermarked luminance Y⁎, the weighted correction
procedure can be represented by Eq. (6).
3M
rV=
4 D = Y − Y⁎
ð6Þ
Clow = sign × ððq − 1Þ × M + r VÞ Y ⁎⁎ = Y ⁎ + f × D
Chigh = sign × ðq × M + r VÞ
( In Eq. (6), Y⁎⁎ is the final luminance of the watermarked image and
⁎ Clow ; if j Clow − C j Vj Chigh − C j f is a quality factor with a value between 0 and 1. When f is with a
C =
Chigh ; if j Clow − C j Nj Chigh − C j larger value, Y⁎⁎ is more close to Y and the visual quality of the final

Note that Clow and Chigh are the two candidate values adaptively
generated for replacing C. Based on the above rule, C⁎ can be easily Table 1
computed and the difference between C⁎ and C is very small. In the NC values under JPEG compression with different compression ratios (CR: compression
ratio; PSNR: in the unit of dB).

M CR 0 10 20 30 40 43 47 53
14 PSNR 36.4 35.0 34.4 32.3 31.0 30.4 29.9 29.3
NC 1 1 0.97 0.96 0.32 0.44 0.46 0.44
18 PSNR 36.0 34.7 34.1 32.2 30.7 30.4 29.8 29.1
NC 1 1 1 0.94 0.95 0.72 0.37 0.43
26 PSNR 34.4 33.5 33.0 31.4 30.5 30.1 29.4 28.7
NC 0.97 0.97 0.97 0.97 0.93 0.96 0.97 0.97

Table 2
NC values under uniform noise attacks (NR: noise ratio in %).

M NR 0 5 10 15 20 25 30
14 PSNR 36.4 29.8 24.6 21.2 18.7 16.9 15.4
NC 1 0.99 0.72 0.48 0.31 0.23 0.12
18 PSNR 36.0 29.6 24.5 21.1 18.7 16.9 15.4
NC 1 1 0.90 0.68 0.45 0.32 0.24
26 PSNR 34.4 29.0 24.3 21.1 18.7 16.8 15.4
NC 0.97 0.97 0.96 0.87 0.71 0.56 0.42
Fig. 2. The positions of DCT coefficients within an image block.
 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60 57

Fig. 4. Watermarks retrieved from marked images Lena after JPEG compression with compression ratios 40 and 43, respectively, when the modulus M is 18.

watermarked image is better. However, the larger value of f will affect

the watermark extraction rate. Therefore, the value of f should not be
too large.
Step 8. Inverse YUV color transformation: The final step of the
proposed watermark embedding algorithm is to transform the YUV
color planes of the watermarked image back to the original RGB color
planes. Instead of the watermarked luminance Y⁎, the final water-
marked image is acquired by using the luminance Y⁎⁎ obtained in
step 7.

3. Watermark extracting algorithm

The steps of the proposed watermark extracting algorithm are very

similar to that of watermark embedding algorithm, except for the step
of watermark information extraction. The procedure for watermark
extraction is quite simple and it doesn't need any assistance of the
original host image. Note that two key parameters are needed in the
watermark extraction procedure: the record of selected image blocks
that were embedded with watermark bits and the modulus M. In
addition, the parameters of Torus permutation function are also
needed when reconstructing the watermark pattern. The steps for
watermark extraction are briefly listed as follows.

Step 1. Extract the luminance information of watermarked image

based on the YUV color transformation.
Step 2. Divide the luminance information of watermarked image into
Fig. 5. The watermarked color images. (a) Lena, PSNR = 31.47, (b) F16, PSNR = 33.9,
non-overlapping blocks of size 8 × 8. Transform each block into
(c) Baboon, PSNR = 26.94, and (d) Scene, PSNR = 30.53.
DCT frequency domain independently based on the DCT
transformation.

Fig. 6. Watermarks retrieved from marked images F16 after JPEG compression with compression ratios 93 and 100, respectively, when the modulus M is 18.
58 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60

Table 3 Table 4
NC values under JPEG compression with different quality factors (CR: compression Simulation results of common image processing attacks.
ratio).

Quality factor 90 60 30 20 15 10 9 8
Lena CR 15.6 26.2 46.8 63.4 78.2 102.5 111.7 124.0
PSNR 29.5 28.8 28.5 27.8 27.2 26.1 25.8 25.4
NC 1 0.99 0.98 0.98 0.90 0.96 0.73 0.49
F16 CR 15.5 25.0 43.3 57.7 70.4 93.3 100.7 110.3
PSNR 31.5 30.3 29.8 28.8 28.0 26.8 26.5 26.1
NC 1 1 1 1 0.84 0.91 0.71 0.40
Baboon CR 7.3 13.1 22.3 31.0 39.1 55.6 61.9 69.1
PSNR 24.7 24.2 23.6 22.9 22.4 21.5 21.3 21.0
NC 1 1 1 1 0.84 0.88 0.54 0.47
Scene CR 11.0 18.8 32.5 44.1 54.8 74.5 81.2 89.5
PSNR 27.7 26.8 26.4 25.7 25.2 24.1 23.9 23.6
NC 1 0.99 0.99 0.96 0.85 0.85 0.57 0.46

Step 3. Extract the binary information of watermark based on the

following rule. In Eq. (7), b is the extracted watermark bit and
C(i, j) is the DCT coefficients at the low-frequency positions
where the watermark bits are embedded.

8
>
> M
<0 ; if ð jC ði; jÞjmod MÞb
2
b= ð7Þ
>
>
: 1 ; if ð jC ði; jÞjmod M Þz M
2

Step 4. Rearrange the extracted watermark bits into right order by

Torus automorphism permutation. Then the watermark
pattern is reconstructed.

4. Secret keys

Secret keys are necessary in the proposed watermark embedding

and extracting algorithms. Four kinds of secret keys are utilized in the
proposed scheme. These secret keys should be preserved well for
watermark extraction and verification.

1. The parameters for Torus automorphism permutation: Torus images is evaluated by the peak signal-to-noise ratio (PSNR) criterion
automorphism permutation has been introduced in Section 2. defined as Eq. (9).
The disarranging parameters k and m were defined in the first step
of watermark embedding algorithm. These parameters are neces- P
31 P
31
W ði; jÞ × W ⁎ði; jÞ
sary when reconstructing the extracted watermark pattern. i=0 j=0
NC = ð8Þ
2. The record of selected image blocks: the record of image blocks P
31 P
31
½W ði; jÞ2
selected for watermark bits embedding has to be preserved well for i=0 j=0
watermark information extraction.
3. The embedding positions of watermark bits: the number of DCT 2
coefficients and the positions of them that were designed for Emax × Iw × Ih
PSNR = 10 log10 P  2 ð9Þ
watermark bits embedding have to be protected well for water- ⁎
Ix;y −Ix;y
mark extraction.
4. The modulus used in the proposed algorithm: the modulus M used
In Eq. (9), Iw and Ih are the width and height of the watermarked
in the watermark information embedding step is necessary in the
image, respectively. Ix,y is the original image pixel value at coordinate
step of watermark bits extracting.

5. Simulation results and discussion Table 5

Performance comparison among [9,10], and proposed scheme in terms Of JPEG
The proposed scheme has been conducted on gray-level and color compression (CR: compression ratio).
images, respectively. All these test images are with size 512 × 512
CR 10 20 30 40 50 53
pixels. The watermark pattern used in the experiments is a binary
[9] PSNR 35.4 34.8 NA NA NA NA
image with size 32 × 32 pixels and it is given as Fig. 3. To provide NC 1 0.29 NA NA NA NA
objective judgment of the extracting fidelity, the similarity measure- [10] PSNR 35.4 34.8 32.6 31.1 NA NA
ment, the Normalized Correlation (NC) value between the original NC 0.99 0.50 0.12 0.07 NA NA
watermark W and the extracted watermark W⁎, is applied and it is Proposed PSNR 33.5 33.0 31.4 30.5 29.0 28.7
NC 0.97 0.97 0.97 0.93 0.96 0.97
defined as Eq. (8). In addition, the visual quality of watermarked
 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60 59

(x, y) and ⁎Ix,yis the altered image pixel value at coordinate (x, y). Emax is To show the improvement of the proposed scheme in terms of JPEG
the largest energy of the image pixels (i.e., Emax = 255 for 256 gray-level compression, we compare our scheme with the earlier works, [9] and
images). Moreover, the compression ratio of the JPEG-compressed image [10], and the performance comparison is given in Table 5. Note that
is obtained by dividing the file size of the original uncompressed image by “NA” means the corresponding experimental datum is unavailable. As
that of the JPEG-compressed image. shown in Table 5, the watermark embedded by the proposed scheme
The simulation results of the gray-level image, Lena, are given in can be almost fully extracted from the marked image when the com-
Tables 1 and 2. Table 1 lists the NC values of extracted watermarks and pression ratio is 53. However, under the same circumstance, the
the image quality (i.e., the PSNR values) of watermarked images after watermarks embedded by [9] and [10] cannot survive after attacked
attacked by JPEG compression with different compression ratios. It by JPEG compression with compression ratios 20 and 25, respect-
shows the embedded watermark can be almost fully extracted from ively. It demonstrates that the proposed scheme achieves a good
the JPEG-compressed image with compression ratio 53 when the improvement on the robustness against JPEG compression. Note
modulus M is 26. Table 2 shows the NC values of extracted watermarks that in Tables 1, 3, and 5, some of the NC values increase when the
after attacked by adding different percentages of uniform noise on compression ratio increases. This is because both the proposed water-
watermarked image. Note that in Tables 1 and 2, the NC values are not marking system and the JPEG compression system are not linear
equal to 1 when the modulus M equals 26 in the circumstance that no systems. For the robustness at high compression ratio and the visual
attack has been performed on the watermarked image. This is because quality of marked image, the applied steps or parameters in the
the steps of inverse DCT transformation, the weighted correction, and proposed algorithm may introduce a little distortion on the embedded
the inverse YUV color transformation in the proposed watermark watermark information. In addition, during JPEG compression attack,
embedding algorithm introduce a little distortion on the embedded the marked image has to undergo several nonlinear transformation or
watermark information. Fig. 4 shows the watermarks retrieved from quantization procedures. Therefore, a small number of the embedded
the watermarked image Lena after attacked by JPEG compression with watermark bits can be distorted at higher compression ratio.
compression ratios 40 and 43, respectively. Note that although there are several alternative DCT coefficients
The simulation results of the color images, Lena, F16, Baboon, and at the low-frequency positions, the watermark bits are embedded into
Scene, are given in Figs. 5 and 6 and Tables 3 and 4. The modulus M the coefficients at positions C(0, 2) and C(2, 0) in our experiments. The
for watermark embedding is 18 in this experiment. Table 3 shows the reasons why we choose the two components to embed watermark
NC values of extracted watermarks after attacked by JPEG compres- are summarized as follows. The DCT coefficients at the positions C(0, 1),
sion with different quality factors. Note that the quality factor for C(1, 0), and C(1, 1) possess more significant information of image.
images is an integer value ranging from 1 to 99, which denotes the Embedding watermark into these positions based on the proposed
predetermined image quality. The larger the quality factor is assigned, scheme introduces too much quality degradation on the marked image.
the lower compression ratio the compressed image obtains and the In addition, it also makes the watermark information more perceptible.
better visual quality the compressed image retains. Fig. 5 illustrates The DCT coefficients at the positions C(0, 3), C(1, 2), C(2, 1), and C(3, 0)
the watermarked color images, together with their PNSR values. It are the higher-frequency components of image. Embedding watermark
reveals that good visual quality of watermarked images can be ob- into these components by the proposed scheme makes it easy to detect
tained by the proposed scheme. Fig. 6 shows the watermarks retrieved the existence of watermark by statistic methods. In addition, to survive
from the marked image F16 after JPEG compression with high compres- high-ratio JPEG lossy compression, it is improper to embed watermark
sion rations 93 and 100, respectively. As shown in Fig. 6, the extracted in the higher-frequency components of images. Therefore, for the
watermarks can be easily recognized by the human eyes. To show the purpose of preserving good visual quality for marked images and
robustness of the proposed watermarking scheme against common increasing the robustness of watermark, the watermark information are
image processing attacks other than JPEG compression, we conduct the embedded in the AC components at positions C(0, 2) and C(2, 0).
options provided in the software of PhotoShop on the marked image Table 6 shows the quantitative analysis of modulus M with respect
Lena. The marked images after being attacked and the retrieved to the visual quality of marked images and the robustness of em-
watermarks are listed in Table 4. As shown in Table 4, the proposed bedded watermarks. We use sixteen gray-level and color images in
scheme is robust to common image processing attacks such as cropping, this experiment. The averaged PSNR and NC values with respect to the
rotation, sharpening, invert, and so on. These experimental results modulus M are listed in this table. The experimental data demonstrate
illustrate that a noticeable improvement, both on the robustness and on that the larger the modulus M is, the more robustness the watermark
the imperceptibility, is achieved by the proposed scheme. obtains. However, a larger modulus introduces more distortion on the

Table 6
Modulus value M with respect to visual quality of marked image and robustness of watermark (CR: compression ratio).

M CR 0 10 20 30 40 50 60 70 80 90 100
4 PSNR 39.4 35.3 32.8 30.9 29.7 28.4 27.1 26.7 26.4 26.1 25.3
NC 1 1 0.98 0.96 0.93 0.84 0.76 0.61 0.49 0.43 0.36
8 PSNR 37.7 34.9 32.6 30.4 29.1 27.9 27.0 26.5 26.3 26.1 25.1
NC 1 1 0.98 0.96 0.94 0.87 0.81 0.71 0.65 0.57 0.42
12 PSNR 36.9 34.2 31.8 30.1 29.0 27.7 27.0 26.3 26.1 25.8 25.1
NC 1 1 1 0.99 0.93 0.90 0.83 0.74 0.65 0.56 0.43
16 PSNR 35.3 33.3 31.5 30.1 28.7 28.0 27.4 26.3 26.0 25.6 25.1
NC 1 1 0.99 0.99 0.98 0.93 0.91 0.84 0.79 0.66 0.57
20 PSNR 34.2 32.4 31.0 29.8 28.6 28.0 27.3 26.1 25.5 25.3 24.9
NC 1 1 0.98 0.97 0.97 0.97 0.94 0.89 0.83 0.76 0.67
24 PSNR 32.4 32.0 31.0 29.6 28.5 27.4 26.9 25.8 25.2 24.9 23.4
NC 1 0.99 0.98 0.98 0.98 0.96 0.97 0.89 0.85 0.79 0.70
28 PSNR 31.3 30.5 29.7 29.0 28.1 27.3 26.1 25.3 24.8 24.1 22.9
NC 1 0.99 0.99 0.97 0.98 0.95 0.92 0.90 0.85 0.80 0.71
60 S.D. Lin et al. / Computer Standards & Interfaces 32 (2010) 54–60

marked images. In our opinion, the modulus M should be set a value [3] K. Gopalakrishnan, N. Memon, P.L. Vora, Protocols for watermark verification, IEEE
Transactions on Multimedia 8 (4) (2001) 66–70.
between 16 and 26. [4] C.Y. Lin, S.F. Chang, A robust image authentication method distinguishing JPEG
compression from malicious manipulation, IEEE Transactions on Circuits and
6. Conclusions Systems for Video Technology 11 (2) (2001) 153–168.
[5] W. Lu, H. Lu, F.L. Chung, Novel robust image watermarking using difference
correlation detector, Computer Standards & Interfaces 29 (2007) 132–137.
An improved DCT-based image watermarking technique has been [6] M.D. Swanson, M. Kobayashi, A.H. Tewfik, Multimedia data-embedding and
introduced in this paper. The watermark is embedded into a digital watermarking technologies, Proceedings of the IEEE 86 (6) (1998) 1064–1087.
[7] C.D. Vleeschouwer, J.F. Delaigle, B. Macq, Invisibility and application functionalities
image, based on the concept of mathematical remainder, by modifying in perceptual watermarking an overview, Proceedings of the IEEE 90 (1) (2002)
the low-frequency coefficients in DCT frequency domain. With the 64–77.
proposed scheme, the embedded watermark can successfully survive [8] J. Huang, Y.Q. Shi, Y. Shi, Embedding image watermarks in DC components, IEEE
Transactions on Circuits and Systems for Video Technology 10 (6) (2000) 974–979.
after attacked by image processing operations, especially for the JPEG
[9] S.D. Lin, C.F. Chen, A robust DCT-based watermarking for copyright protection, IEEE
compression with high compression ratio. Moreover, the watermark Transactions on Consumer Electronics 46 (3) (2000) 415–421.
embedding and extracting processes are very simple and the water- [10] S.D. Lin, S.C. Shie, C.F. Chen, A DCT based image watermarking with threshold
mark is self-extractable. Simulation results show that the proposed embedding, International Journal of Computers and Applications 25 (2) (2003)
130–135.
scheme outperforms the earlier works. Therefore, we conclude that [11] C.C. Chang, J.Y. Hsiao, C.L. Chiang, An image copyright protection scheme based on
the new proposed technique is more suitable for images that will be torus automorphism, Proceedings of the First International Symposium on Cyber
highly JPEG-compressed and transmitted on the Internet. Worlds (2002) 217–224.
[12] M. Engedy, V.N.K. Munaga, A. Saxena, A robust wavelet based digital watermarking
scheme using chaotic mixing, Proceedings of the First International Conference on
Acknowledgements Digital Information Management (2006) 36–40.
[13] G. Voyatzis, I. Pitas, Chaotic mixing of digital images and applications to
watermarking, Proceedings of the European Conference on Multimedia Applica-
The authors are supported by National Science Council (Taipei, tions, Services and Techniques 2 (1996) 687–695.
Taiwan, ROC), National Dong Hwa University (Hualien, Taiwan, ROC),
and National Formosa University (Hu-Wei, Yunlin, Taiwan, ROC).

References
[1] J.L. Liu, D.C. Lou, M.C. Chang, H.K. Tso, A robust watermarking scheme using self-
reference image, Computer Standards & Interfaces 28 (2006) 356–367.
[2] D.C. Lou, H.K. Tso, J.L. Liu, A copyright protection scheme for digital images using
visual cryptography technique, Computer Standards & Interfaces 29 (2007) 125–131.