ADAPTIVE WATERMARKING OF COLOR IMAGES

IN THE DCT DOMAIN

Ibrahim Nasir, Jianmin Jiang and Stanley Ipson

School of Computing, Informatics and Media,University of Bradford, Bradford BD7 1DP, U.K.

Keywords: DCT domain, DC components, Classification, Adaptive embedding, Watermarking, Color images.

Abstract: This paper presents a new approach of embedding watermarks adaptively in DC components of color

images with respect to analysis of the image luminance in the YIQ model or the blue channel in the RGB

model. The embedding strength is determined by the classification of DCT blocks, which is simply

implemented in the DCT domain via just analyzing the values of two AC coefficients rather than using

method such as Sobel, Canny, or Prewitt. Furthermore, a new combination algorithm of both watermark

extraction and blind detection is designed, where the watermark is extracted directly in the spatial domain

without knowledge of the original image. Experimental results demonstrated that the embedded watermark

is robust to the attacks including JPEG-loss compression, JPEG2000, filtering, scaling, small cropping,

small rotation-crop and self-similarities attacks.

1 INTRODUCTION

Recently, digital watermarking techniques have

been utilized to maintain the copyright of digital

data by identifying the owner or distributor of digital

data. There are some contradictory requirements in

the watermarking process that, on one hand, the

embedded watermark should not affect the image

quality in a visible manner, but on the other, the

watermark should still be extractable from the

watermarked media after intentional or unintentional

attacks (Petitcolas et al

., 1999) and (Hartung et al.,

1999

A wide variety of image watermarking

algorithms has been proposed to provide copyright

protection of digital images. Most existing

watermarking algorithms use key-generated

pseudorandom patterns as watermarks and focus

mainly on embedding watermarks into the low or

the middle frequency coefficients of grey-scale

images. The extension to color images is

accomplished by marking the image luminance

(Kutter, 1998), or by processing each color channel

separately. Kutter et al. (1997) proposed a method

based on embedding a watermark by modifying a

selected set of pixel values in the blue channel, since

the human eye is less sensitive to changes in this

band. Fleet and Heeger (1997) proposed a method,

which takes into account the characteristics of the

human visual system (HVS) with respect to color

perception. They suggested embedding the

watermark into the yellow-blue channel of color

images and using S-CIELAB space to measure the

color reproduction error. However, their method can

only resist printing and rescanning attacks. Barni et

al. (2002) introduced another color image

watermarking method based on the cross-correlation

of RGB channels. However, it has relative high

computing costs and low processing speed since the

full-frame DCT is used for three color channels.

Tsai et al. (2004) provided a solution of embedding

the watermark on a quantized color image.

Instead of using AC coefficients to embed a

watermark, Huang et al. (2000) suggested that more

robustness can be achieved if the watermark is

embedded into the DC coefficients of the DCT and

demonstrated that the DC coefficients provide more

perceptual capacity than AC coefficients. Based on

Huang et al. suggestion, watermarks were embedded

in DC coefficients of images (Huang et al., 2005

)

and

(Nasir et al., 2008). In (Huang et al., 2005), the

watermark embedded into DC coefficients of a color

image directly in the spatial domain, followed by a

saturation adjustment technique performed in the

RGB color space. The main weaknesses of the

method presented in (Huang et al., 2005

) are as

follows: (i) The detection process requires the

original image, which may not necessarily available

in some applications; (ii) The image contents are not

171

Nasir I., Jiang J. and Ipson S. (2010).

ADAPTIVE WATERMARKING OF COLOR IMAGES IN THE DCT DOMAIN.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 171-177

DOI: 10.5220/0002895601710177

 SciTePress

taken into account when embedding the watermark

and, as a result, the maximum-possible

imperceptibility and robustness of the embedded

watermark cannot be guaranteed. In (Nasir et al.,

2008), the watermark was embedded into DC

coefficients of gray-scale images without taking into

account the content of the image.

Based on the fact that the magnitude of DC

coefficients is much larger than any AC coefficients,

DC coefficients can provide more perceptual

capacity than AC coefficients. On the other hand,

DC coefficients are less affected than any AC

coefficients when the watermarked is attacked by

JPEG compression, lowpass filtering and

subsampling operations. Therefore, DC coefficients

are suitable for embedding watermark (Huang et al.,

2005

). Motivated by those observations, in this

paper, we propose a new adaptive and blind image

watermarking method, which is based on the

principle of embedding a watermark in the DC

coefficients of subimages in the DCT domain. These

subimages are obtained through subsampling the

original luminance component Y or the blue

component B of color images in the YIQ or the

RGB color models respectively. In comparison with

existing reported work, our proposed method

possesses significant advantages, which can be

highlighted as: (i) The watermark is embedded in

the DC coefficients to provide more robustness than

using AC coefficients; (ii) The watermark is

extracted without knowledge of the original non-

watermarked image; (iii) The watermark is extracted

directly in the spatial domain rather than applying

the DCT again to the watermarked image; (iv) The

strength of the watermark is determined adaptively

to the contents of the host image to guarantee the

best possible perceptibility and robustness of the

embedded watermark; (v) the DCT and its inverse

are applied only to the selected blocks, which are

used to embed the watermark.

The rest of this paper is structured as follows.

Section 2 describes the adaptive determination of

watermarking strength; Section 3 and 4 present the

embedding and the extraction processes and Section

5 presents some experimental results. Conclusions

are drawn in Section 6.

2 ADAPTIVE DETERMINATION

OF WATERMARKING

STRENTH

A major challenge in designing a watermarking

algorithm is to find a strategy that satisfies the

conflicting objectives that, on one hand, the added

watermark is imperceptible to the human eyes but,

on the other, it should be robust to removal attacks.

The best way to achieve better trade-offs between

imperceptibility and robustness requirements is to

take the characteristics of the non-watermarked

image into account when embedding the watermark.

Chang et al (2005) proposed a technique for

extracting 5 edge patterns directly in the DCT

domain, and proved that DCT blocks of size 8

can be classified as certain type of edge patterns.

Jiang et al. (2008) suggested that three edge patterns

rather than five are sufficient to describe and

characterize the visual content of the image in the

DCT domain. Therefore, the proposed method

exploits this block classification scheme to analyze

the visual content and hence determine the

watermark embedding strength.

Via exploitation of Jiang’s classification scheme,

all DCT blocks can be further analyzed as smooth or

non-smooth based on the specific values of the two

DCT coefficients. Non-smooth blocks are then

further classified to determine if they contain both

vertical and horizontal edges or contain one of the

edge patterns. To determine the embedding strength,

we proposed the following adaptive scheme

(

)

()

⎪

⎩

⎪

⎨

⎧

λ≥δδα

λ<δδα

=α

otherwise

,minifelse

,maxif

edge

22/0texture

12/0smooth

(1)

where α stands for the embedding strength and,

)0,1(

)1,0(

are the absolute value

of the DCT coefficients X(1,0) and X(0,1),

respectively,

and

are thresholds. The

derivation of

2/π

does not require any addition

or multiplication and only two DCT coefficients are

used to classify DCT blocks.

Figure 1 demonstrates the classification results

obtained by applying the proposed method to images

with different textures. As an example, Lena

includes large smooth areas with sharp edges;

Peppers includes large smooth areas without sharp

edges and Baboon includes textured areas.

The white areas shown in Figure 1 are classified

as smooth, and thus any small change incurred by

watermarking could be visible. As a result, the

corresponding watermark should have a low

embedding strength. Similarly, the black areas in

Figure 1 are classified as edge or textured blocks,

and hence changes incurred by watermarking would

be less visible. Therefore, the watermark embedding

SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications

172

(a)

(b)

Figure 1: (a) Host images, (b) Classified images.

strength should be arranged as follows:

textureedgesmooth

3 ADAPTIVE EMBEDDING

OF WATERMARKS

The proposed embedding algorithm can be applied

to color images in the YIQ (Luminance, Hue, and

Saturation) or in the RGB (Red, Green, and Blue)

color models. In the YIQ color model, the gray-scale

information is separated from color information;

therefore, the Luminance component Y is selected to

embed the watermark. In the RGB color model, the

blue component B is selected to embed the

watermark because the human eyes are less sensitive

to changes in this color component (Kutter, 1997).

The proposed embedding algorithm consists of five

steps. In the first step, four subimages are produced

by subsampling the luminance component (Y) or the

blue component (B) of a color image as given in

(Chu, 2003). In the second step, the embedding

positions of the watermark are specified by a secret

key, which is used to generate a random sequence

}ji),4,3,2,1(j,i),j,i{(S

≠

∈=

, with the length equal

to the length of the watermark, which serves as

blocks selector. where L=1, 2,…, the length of the

watermark. In the third step, the DCT is applied to

non-overlapping blocks of size 8×8 of the two

selected subimages, which are selected according to

random sequence

. In the fourth step, the

watermark embedding strength (α) is determined

according to the block pattern classification as given

in (1). In the final step, watermark bits are

embedded in pairs of DC coefficients of the selected

blocks of size 8×8

)}Sj,i(),DC,DC{(

Lji

∈

and

leaving the AC coefficients unchanged. The

procedure of embedding watermark bits in DC

coefficients of blocks of size 8×8 is achieved as

follows: If

and

DCDC >

, then swapping

j,i in

, and if

and

DCDC <

, then

swapping

j,i in

. Using the modified

, the

watermark embedding algorithm is defined as

follows:

DCDC

Avg

(2)

DCDC

Diff

−

(3)

Avg

Diff

Threshold

DC =

(4)

⎪

⎩

⎪

⎨

⎧

α−

=β≤α+

β>

otherwiseDC.DC

)1W(&)DC(elseifDC.DC

)DC(ifDC

Avgi

LThresholdAvgi

Thresholdi

(5)

()

.()&(0)

jThreshold

jj Avg Threshold L

jAvg

DC if DC

DC DC DC elseif DC W

DC DC otherwise

αβ

∗

⎧

⎪

+≤=

⎨

⎪

−

⎪

⎩

(6)

where α is the watermark embedding strength and β

is the threshold,

and

are the DC

coefficients of the blocks inside the two selected

subimages,

and

are the watermarked DC

coefficients,

stands for watermark bit and L=1,

2,…, the length of the watermark. The values of the

selected pairs of DC coefficients (

) are

altered if they do not satisfy the embedding

threshold β. The embedding depends on the

watermark bits as given in Equations (5) and (6). In

the case of embedding a watermark bit ‘1’, the value

of the

will be increased and the value of

the

will be decreased. In the case of embedding a

watermark bit ‘0’, the value of the

will be

decreased and the value of the

will be increased.

The values of the increment and decrement are made

large enough depending on the watermark

embedding strength α. The proposed method applies

the DCT and its inverse only to the selected 8×8

blocks of subimages and uses the DC coefficients of

those selected blocks to embed the watermark.

ADAPTIVE WATERMARKING OF COLOR IMAGES IN THE DCT DOMAIN

173

4 WATERMARK ETRACTION

PROCESS

The proposed extraction algorithm can be divided

into four stages, which include: (i) producing four

subimages by subsampling the luminance

component (Y) or the blue component (B) of a color

image, which is subjected to watermark extraction;

(ii) specifying the extraction positions of the

watermark using the secret key; (iii) DC coefficients

are computed directly in the pixel domain by simple

averaging the pixels inside each block as given in

(7); and (iv) The watermark extracted bits are

determined by comparing the DC coefficients values

of selected subimages as given in equation (8).

∑∑

)y,x(f

(7)

where

)y,x(f stands for the intensity pixel value at

location )y,x( .

⎪

⎩

⎪

⎨

⎧

≥

otherwise0

DCDCif1

(8)

where

is the extracted bit,

}N,....,2,1{L ∈

, N

represents the length of the watermark,

and

are DC coefficient values of 8×8 blocks of the

selected subimages.

The Correlation coefficient (CC) given in (Huang et

al., 2005

) is used to measure the similarity between

the original watermark and the extracted watermark.

An appropriate threshold T is used to make the

binary decision as to whether a given watermark is

present or not. Figure 2 shows the watermark

detector response with 1000 watermark seeds, only

one of which is the correct watermark. The

threshold T was set as 0.4 and when CC value

exceeded this, the existence of the watermark was

declared.

0 100 200 300 400 500 600 700 800 900 1000

-0.2

0.2

0.4

0.6

0.8

1.2

Figure 2: Detector responses for 1000 watermark seeds.

The proposed extraction algorithm possesses the

following advantages; (i) The watermark is

extracted directly from the watermarked images and

yet such extraction is conducted directly in pixel

domain rather than applying DCT again to the

watermarked image; and (ii) The extraction process

is not required the un-watermarked image; which

makes the proposed method applicable in a wider

range of application than those which require the un-

watermarked image for detection.

To achieve the security requirement of the

proposed method, the embedding positions of the

watermark are determined by a secret key.

Therefore, the watermark can not be extracted

without knowing the secret key, whose length is

equal to the watermark length, long enough to make

it impractical for hacking. To achieve the robustness

requirement, DC coefficients are selected to embed

the watermark because they can provide more

perceptual capacity than the AC coefficients and

they are less affected than any AC coefficient when

the watermarked image is attacked by JPEG

compression, low pass filtering or subsampling

operations. To achieve better tradeoffs between the

requirements for imperceptibility and robustness, the

watermark embedding strength is determined

adaptively to the image features. The embedding

capacity of the proposed scheme depends on the size

of the image. For the image with size

M×N, the

maximum embedding capacity is given as below

⎥

⎦

⎤

⎢

⎣

⎡

⎥

⎦

⎤

⎢

⎣

⎡

(9)

5 EXPERIMENTAL RESULTS

In the experiments, the logo used for watermarking

was a binary image of size 32×32 and the values of

and β are defined as given in (Jianmin et al.,

2008) and (Lu et al., 2006). The values of

watermark embedding strength in the YIQ color

model are

01.0

smooth

=α

04.0

texture

=α

025.0

edge

=α

and in the RGB color model are

07.0

smooth

2.0

texture

=α

1.0

edge

. These

empirical values are appropriate for various tested

images. In the embedding process, the distortion of

an image depends on the threshold

β, the watermark

strength

α, and the length of the watermark. The

threshold

β controls the difference between the

watermarked

DC coefficients. The higher β, the

more embedding distortion is introduced on the

SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications

174

watermarked image as given in equations (5) and

(6). Meanwhile, the embedding of the watermark is

controlled by the watermark strength

α. The

classification of the image determines where the

watermark should be strong and where the

watermark should be weak. The watermark strength

α is lower in smooth regions and higher in texture

regions. The higher

α, the more the distortion is

introduced on the watermarked image. Hence, there

is a tradeoff between robustness and

imperceptibility.

To evaluate the watermark imperceptibility,

fifteen different images are tested including Lena,

Peppers, Baboon, F16-plane, Barbara, Tiffany, Lake

and Couple, etc. The Peak signal to noise ratio

(PSNR) is adopted to evaluate the perceptual

distortion of the proposed scheme. The

PSNR values

of the fifteen watermarked images are between 42

and 50 dB. These values are all greater than 30.00

dB, which is the empirically tested threshold value

for the image without any perceivable degradation

(Qi et al., 2007). Taking Peppers image as examples,

the un-watermarked and the watermarked images are

shown in Figure 3.

(a) (b)

Figure 3: (a) Original Peppers image; (b) Watermarked

image; (c) original watermark; (d) extracted watermark.

It can be seen that the differences between the

corresponding watermarked and un-watermarked

images are imperceptibly. The PSNR values

between the original and the watermarked images

are shown in Table 1.

Various common signal processing and geometric

attacks were applied to the watermarked images.

Most of these were performed using the benchmark

software StirMark 4.0 (Petitcolas et al., 1998) and

(Petitcolas et al., 2000). As an example, figure 4

shows the extracted watermarks from watermarked

Table 1: PSNR between original and watermarked images.

Lena Peppers Baboon

Proposed YIQ scheme

Proposed RGB scheme

Scheme in (Huang et al.,

2005)

45.67

45.75

44.06

43.44

43.98

43.23

44.37

43.50

42.42

Lena image after filtering attacks include: low-pass,

median, Gaussian and mean filtering attacks with

3×3 windows size. As seen, all the extracted

watermarks can be visually identified and the

detector’s response correctly declares the existence

of the watermark.

(a) (b)

Figure 4:

Results of attack by (a) low-pass filtering; (b)

median filtering(c) Gaussian filtering (d) mean filtering

Figure 5 (a) and (b) show experimental results from

JPEG attacks to the watermarked images ‘Lena’ and

‘Peppers’. As shown, the robustness achieved by

embedding the watermark in the luminance

component Y is better than embedding the

watermark in the blue component. This is because

the blue component has characteristic of the highest

frequency range than the red and the green

components of the RGB and these high frequency

components may be discarded by lossy compression

quantization. As shown, the proposed method

performs better than Huang’s method under JPEG

compression attack.

Self-similarities (SS) attacks in different color space

were applied to the watermarked images. As an

example, the results for the Lena image is shown in

figure 6.

As seen, the proposed method performs

well under self similarities attacks.

ADAPTIVE WATERMARKING OF COLOR IMAGES IN THE DCT DOMAIN

175

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 102030405060708090100

JPEG-loss compression quality %

correlation coefficient (CC)

RGB

YIQ

Huang Method

(a)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 102030405060708090100

JPEG-loss compression quality %

correlation coefficient (CC)

RGB

YIQ

Huang Method

(b)

Figure 5:

Results of attack by JPEG-loss compression (a),

(b) and (c) detector response CC versus JPEG

compression quality for Lena and Peppers images,

respectively.

(a) (b) (c)

(d) (e) (f)

Figure 6: Extracted watermarks after self-similarities

attacks where (a), (b) and (c) from RGB color space and

(d), (e) and (f) from YIQ color space .(a) and (d)

extracted after ss1 attack, (b) and (e) extracted after ss2

attack, (c) and (f) extracted after ss3 attack.

Table 2 summarizes some experimental results by

applying several common image processing and

some geometric attacks on Peppers and Baboon

images watermarked using the proposed method and

Huang’s method (Huang et al., 2005

). It can be seen

that, the proposed method performs better than

Huang’s method under common image processes

and some geometric attacks, such as JPEG

compression, JPEG2000, scaling by 200%, self-

similarities. The extracted watermark can also be

correctly identified by the proposed method after

filtering attacks. However, Huang’s method

performs better than our method for filtering attacks.

This is because the un-watermarked image was used

in the extraction process in Huang’s method. By

contrast, the proposed method does not require the

un-watermarked image, which is not available for

most application. The experimental results of the

proposed methods show that more robustness can be

achieved when the watermark is embedded

adaptively to features of the host image.

Table 2: Comparison between the proposed method and

Huang’s method (Huang et al., 2005) under common

image processing and geometric attacks.

Attack operation on Peppers and Baboon images

Huang’s

method

Proposed method

Peppers

Baboon

Peppers

RGB YIQ

Baboon

RGB YIQ

Attack operation CC CC CC CC CC CC

Jpeg 100

Jpeg 50

Jpeg 20

JPEG2000 100

JPEG2000 50

JPEG2000 20

Scaling 2.0

Scaling 0.5

Median filter 7×7

Gaussian filter

3×3

SS1(self

similarity)

SS2

SS3

Crop 25%

0.90

0.68

0.52

0.98

0.68

0.56

0.90

0.88

0.81

0.90

0.94

0.93

0.50

0.57

0.59

0.53

0.45

0.98

0.44

0.41

0.74

0.75

0.53

0.90

0.60

0.64

0.42

0.59

0.99

0.68

0.58

1.0

0.67

0.63

0.99

0.64

0.61

0.62

1.0

0.86

0.97

0.73

1.0

0.77

0.70

1.0

0.71

0.68

0.97

0.70

0.56

0.73

0.96

1.0

0.71

0.75

1.0

0.93

0.78

1.0

0.72

0.61

1.0

0.52

0.46

0.69

1.0

0.85

1.0

0.76

1.0

0.95

0.78

1.0

0.74

0.56

0.94

0.54

0.45

0.65

0.78

1.0

0.54

0.75

6 CONCLUSIONS

We have presented a new adaptive color image

watermarking scheme based on embedding the

watermark adaptively to the features of the host

color image in perceptually significant DC

coefficients in the DCT domain.

The proposed watermark embedding process takes

into account the content of the image to achieve the

maximum-possible imperceptibility and robustness

of the embedded watermark. The experimental

results show that the proposed technique succeeds in

making the watermark perceptually invisible and

also robust against various signal processing

operation and some geometric attacks. The results

demonstrate that more robustness is achieved when

the watermark embedded in the Luminance

component rather than using the blue component.

The performance of the proposed scheme could be

improved by taken into account the correlation

SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications

176

between RGB bands and color gradients. Moreover,

robustness against geometric attacks could be

improved by using reference points, which act as

synchronization marks between the watermark

embedding and detection.

REFERENCES

F. Petitcolas, R. Anderson, and M. Kuhn, 1999.

Information hiding-a survey,

Proc. IEEE, 87(7), pp.

1062-1078.

F. Hartung and M. Kutter, 1999. Multimedia

watermarking techniques,

Proc. IEEE, 87(7), pp.

1079-1107.

M. Kutter, 1998. Watermarking resisting to translation,

rotation, and scaling, in:

Proc. Of the SPIE Conf. on

Multimedia Systems and Applications

, pp. 423-431.

M. Kutter, F. Jordan, F. Bossen, Digital signature of color

image using amplitude modulation, 1997. in:

Proc. Of

Storage and Retrieval for Image and Video Database

SPIE, vol. 3022, pp. 516-526.

D. Fleet, D. Heeger, 1997. Embedding invisible

information in color images, in

Proc. IEEE Int. Conf.

on Image Processing

’97, vol. 1, pp. 532-535.

M. Barni, F. Bartlini, A. Piva, 2002. Multichannel

watermarking of color image,

Proc. IEEE Trans. On

Circuit Systems for Video Technology

, vol. 12(3), pp.

142-156.

P. Tsai, Y.C. Hu, C.C. Chang, 2004. A color image

watermarking scheme based on color quantization,

Signal Process, pp. 95-105.

J. Huang, Y. Shi, and S. Yi, 2000. Embedding image

watermarks in dc components,

Proc. IEEE

Transactions on Circuits and Systems for Video

Technology

, vol. 10, no. 6, pp. 974-979.

P. S. Huang, C. S. Chiang, C. P. Chang, and T. M. Tu,

2005. Robust spatial watermarking technique for

colour images via direct saturation adjustment,

IEEE

Proceedings -Vision, Image and Signal Processing

vol. 152 (5,7), pp. 561-574.

I. Nasir, Y. Weng, J. Jiang, S. Ipson, 2008. Subsampling-

based image watermarking in compressed DCT

domain,

Proc. 10 the IASTED Int. Conf. on signal and

image processing

, pp. 339-344.

K. K. H. S. Chang, A compressed domain scheme for

classification block edge patterns, 2005.

Proc. IEEE

Transactions on Image Processing

, vol. 14(2), pp.145-

151.

J. Jianmin, Q. Kaijin, X. Guogiang, 2008. A block-edge-

pattern based content descriptor in DCT domain,

Proc.

IEEE Transactions on Circuits and Systems for Video

Technology

, vol. 18(7), pp. 994-998.

W. C. Chu, 2003. DCT-based image watermarking using

subsampling,

Proc. IEEE Transactions on Multimedia,

vol. 5

(1), pp. 34-38.

W. Lu, H. Lu, and F. Chung, 2006. Robust digital image

watermarking based on subsampling,

Applied

Mathematics and Computation

, vol. 181, pp. 886-893.

X. Qi and J. Qi, 2007. A robust content-based digital

image watermarking scheme,

Signal Processing, 87

(6), 1264-1280.

F. A. P. Petitcolas, 2000. Watermarking schemes

evaluation,

Proc. IEEE Signal Processing Magazine,

vol. 17

(5, pp. 58-64.

F. Petitcolas, R. Anderson, and M. Kuhn, 1998. Attacks

on copyright marking systems,

Proc. 3

Int. Int.

workshop on Information Hiding

, vol. 1525, Lecture

Notes in Computer Science, pp. 218-238.

ADAPTIVE WATERMARKING OF COLOR IMAGES IN THE DCT DOMAIN

177