Image Watermark using Visual Model Based

0 downloads 0 Views 518KB Size Report
Abstract. In this paper, a novel blind image watermark scheme is developed based on discrete wavelet transform (DWT). In order to make the watermark robust ...
IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.1, No.2, April 2007 CSES International ⓒ2007 ISSN 0973-4406

119

Image Watermark using Visual Model Based Discrete Wavelet Transform Cong JIN, Liang-Gang PAN, and Ting SU Department of Computer Science, Central China Normal University Wuhan 430079, P.R.China E-mail:[email protected]

Abstract In this paper, a novel blind image watermark scheme is developed based on discrete wavelet transform (DWT). In order to make the watermark robust and transparent, the watermark is embedded in the average of wavelet blocks using the visual model based on the human visual system. Low-pass wavelet coefficients n least significant bits (LSBs) are adjusted in concert with the average. Simulation results show that the proposed scheme is transparent and robust against many attacks such as JPEG compression, adding noise, rescaling, cropping, rotation, filtering, etc. Key words: Watermark, Visual model, Robustness, Transparent, Wavelet.

1. Introduction With the explosive growth of the Internet in recent years, protection of the copyright of multimedia data becomes one of the most important topics in the Internet world. As a potential and effective way to solve this problem, digital watermark becomes a very active research area of signal and information processing. A digital watermark is a signal that is embedded in a digital image or video sequence. It allows one to establish ownership, identify a buyer, or provide some additional information about the digital content. In general, it is required that the embedded watermark should be not only transparent to human observers, but also robust enough so that it cannot be destroyed or removed after some processing or attacks. Generally, the image watermark embedding process for digital images can be accomplished in either the spatial or frequency domain. It is shown that a better compromise between robustness and transparency can be obtained using the frequency domain scheme. Frequency-domain watermark techniques could be based on spatially local or global transforms. In the published literature, a famous frequency domain scheme for digital watermark of images is proposed by Cox et al.[1]. One of the significant contributions in this work is the realization that the watermark should be inserted in the perceptually significant portion of the image to be robust. A DCT is performed on the whole image. In recent years, with the development of JPEG2000 and MPEG-4, the application of DCT is limited due to the fact that the wavelet Manuscript received January 1, 2007 Manuscript revised March 1, 2007

transform is playing an important role in JPEG2000 and MPEG-4. Therefore, the wavelet transform approach is one of the promising techniques for image watermark [2, 3, 4]. In this paper, we propose a new blind watermark scheme of images based on DWT. The watermark signal is embedded in low-pass wavelet coefficients n LSBs. Unlike most watermark schemes, the watermark is not embedded by modulating individual wavelet coefficient but by modulating the average of coefficients in the wavelet blocks. Visual model is employed to achieve the best tradeoff between transparent and robustness to signal processing. Watermark detection is accomplished without the original. Simulation results show that the proposed scheme is transparent and robust against many common images processing and editing such as compression, adding noise, rescaling, cropping, rotation, filtering, etc. In the following of the paper, Section 2 describes the proposed scheme using DWT and the detailed steps of the proposed algorithm will be discussed in this section. Section 3 gives the detection theory of watermark. Experimental results are presented in Section 4 and the conclusion is given in Section 5.

2. Watermark Embedding Given a set of wavelet coefficients, it has been observed that the average has a smaller change than individual coefficient. Thus, the watermark embedded in the average of the wavelet blocks is more robust than in the individual coefficient [5]. In this paper, we choose the low-pass wavelet coefficient as the carrier of watermark signal. The low-pass wavelet coefficient matrix Ca can be obtained by 2D wavelet decomposition. M and N are the length and width of the wavelet block respectively and

I i ( k ) is the ith wavelet coefficient in the kth wavelet

[

]

block. Naturally, variable i ∈ 1, M × N . The n LSBs of

I i ( k ) is defined as

IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.1 No.2, April 2007

120

Iˆi ( k ) = mod ( I i ( k ) , 2n )

(1)

Then the average of the wavelet block is defined in the following way. M ×N

Average ( k ) =

∑ Iˆ ( k ) i =1

i

(2)

M ×N

If a few of I i ( k ) are changed by

Ω due to some of

distortions, the average of the wavelet block will only have a small change. Supposing that I i′ ( k ) is the ith wavelet coefficient in the

kth

wavelet

Iˆi′ ( k )

embedding,

block

after

is

the

the

watermark n

LSBs

of I i′ ( k ) and Average′ ( k ) is the average of Iˆi′ ( k ) in the kth wavelet block accordingly. The watermark W, consisting of a binary pseudo random sequence, W ( k ) ∈ {−1,1} , is embedded by adjusting the average

of

wavelet

blocks

in

this

⎧ ⎡0, 2 ) , if W ( k ) = −1 ⎪⎣ Average′ ( k ) ∈ ⎨ n −1 n ⎪⎩ ⎣⎡ 2 , 2 ) , if W ( k ) = 1

coefficient n LSBs. These questions will be discussed in following text orderly. In early work on DCT based watermarks, each wavelet coefficient is adjusted to make the average equal to 2

n−2

or 2 − 2 n

n−2

.But in proposed scheme, as show in

Eq.3, the average trends to 2

n−2

or 2 − 2 n

n−2

;at the same

time, some, but not each Iˆi ( k ) are chosen to adjust. Hence,

a

flag

function

must

be

set

to

decide

Iˆi ( k ) whether need adjust or not. The flag rule is defined

as follows.

((

)

Fi ( k ) = sign 2 n −1 − Iˆi ( k ) × W ( k )

)

(4)

1, if x ≥ 0 Where, sign ( x ) = ⎧ . The detailed result of ⎨ ⎩0, if x < 0 flag function is showed in Table 1. Table 1. The detailed results of

Fi ( k ) and S ( k )

way

n −1

(3)

In order to further enhance the robustness of the watermark, we expect Average′ ( k ) could more trends to 2

n−2

or 2 − 2 n

n−2

. As show in Fig.1, there is maximal

distance between integers 2

n−2

and 2 − 2 n

n−2

In addition, it is possible that Average ( k ) has satisfied Eq.3 before doing any adjustment. So in this case, no adjust should be needed in these wavelet blocks. Considering the robustness of the watermark illuminated in above paragraphs, a very light adjustment is still indispensable. Contrarily, if Average ( k ) is not satisfied

Eq.3, a relative strong adjustment is required. So an estimation between 2

n −1

− Average ( k ) and W ( k ) to

determine adaptable strength is necessary when the watermark is embedding. The strength function S ( k ) is defined as Fig. 1 The distance of n bit binary number in circle of integer mod 2

n

Before adjusting Average ( k ) , there are three pending questions. First, whether or not each individual coefficient n LSBs need adjust, if it is not, how to mark them; Second, how to control embedded strength if Average ( k ) have

satisfied Eq.3 before doing any adjustment; Third, how to make use of visual model when adjust individual

S ( k ) = sign ( X ( k ) )

Where, X ( k ) = ( 2

n −1

(5)

− Average ( k ) ) × W ( k ) . The detailed

result of strength function is also showed in Table 1. To adapt the watermark sequence to the local properties of the wavelet block, we use the model based on HVS [5] in the watermark system. The model is similar to that of reference [5], but is developed independently. The visual model takes into account the brightness sensitivity and

Image Watermark using Visual Model Based Discrete Wavelet Transform

121

L

texture sensitivity of the wavelet block to noise. The visual model function Vm ( k ) is defined as

Vm ( k ) = brightness ( k ) × texture ( k )

ρ (W ′,W ) = β

(6)

M ×N

Where,

∑ ⎡⎣brightness ( k ) − I ( k )⎤⎦

texture ( k ) =

M ×N

,

brightness ( k ) =

i =1

i

M ×N

, and β to control the degree

of texture sensitivity. This visual model function indicates that the human eye is less sensitive to noise in the highly bright and the highly textured areas of the image. Hence, the wavelet blocks are divided into two parts depend on the value of Vm ( k ) : high activity wavelet block and low

activity wavelet block. For simplicity, the threshold Tc is

set to the average of Vm ( k ) .The following function can

be applied to distinguish high or low activity wavelet block.

T ( k ) = sign (Vm ( k ) − Tc )

L

∑ W ′(k )

k =1

2

(10)

L

Where, L is the length of the watermark signal. The threshold Tρ is set to minimize the sum of the probability of error detection and the probability of false alarm in this paper. If ρ ≥ Tρ , we considered the watermark is present,

M ×N

∑ I (k )

=

k =1

k =1

2

i

i =1

L

∑W ′ ( k ) ×W ( k ) ∑W ′ ( k ) ×W ( k )

(7)

Considering the tradeoff between robustness and transparence, the proposed watermark embedding algorithm can be formulated as follows. Iˆi′ ( k ) = Iˆi ( k ) + αW ( k ) Fi ( k ) ⎡⎣ 2n−2− S ( k ) + T ( k ) × 2n −3 ⎤⎦ (8) Where α is a scaling factor to control the strength of the inserted watermark. Based on the above discussion, the n LSBs of wavelet coefficients have been adjusted by using the Eq.8. Naturally, their average has been updated depending on the requirement of W ( k ) as show in Eq.3. In the other word, the watermark has been embedded.

3. Watermark Extraction and Detection The watermark sequence can be extracted without the original image. From the process of the watermark embedding, we can obtain the watermarked objects by applying the function of Eq.3. Thus, for a given watermarked object, the watermark can be extracted as

⎧−1, if Average ( k ) ∈ ⎡0, 2n −1 ) ⎪ ⎣ (9) W ′(k ) = ⎨ n −1 n ⎪⎩1, if Average ( k ) ∈ ⎡⎣ 2 , 2 ) In order to detect the watermark W ′ extracted from watermarked object, we firstly evaluate the detector response (or similarity of W ′ and W ) as

otherwise absent.

4. Simulation Experiments Results In this section, we present the experimental results to illustrate the robustness and transparency of the proposed watermark scheme. Simulations are carried out for several standard monochrome images but only report result in detail for 512 × 512 Lena and Baboon .In ours experiments, the parameters are presented as follows: the threshold Tρ = 0.15 , β = 0.318 , n = 5 , wavelet-

level = 2 , wavelettype = ' db1' and scaling factor α ∈ [1.25,1.73] . We set α as a constant and choose α = 1.45 in order to compare expediently. For simplicity, the length and width of the wavelet block are fit to 4 ( M = N = 4 ) .On this occasion, there are 1024

bits can be embedded for an 512 × 512 image, in other words, L = 1024 . In order to test the performance of the proposed watermark scheme, 200 watermarks were randomly generated. Lena image is the first object of we experiment. The PSNR result between the original object and the watermarked object is 42.41 dB. As shown in Fig.2, the watermark is transparent and the object with watermark appears visually identical to the object without watermark. Fig.2.(c) shows the response of the watermark detector to

200 randomly generated watermarks of which only one matches the watermark present in Fig.2.(b). The response to the correct watermark is much higher than the responses to incorrect watermarks. The result of experiment for Baboon image is showed in Fig.3. Added experiment results are listed in Table 2. To evaluate the robustness of our scheme against unintentional and intentional attacks, we test the watermarked object with JPEG compression, adding noise, filtering, rescaling, cropping and rotation.

IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.1 No.2, April 2007

122

Table 2. PSNR between original image and watermarked object

4.1 JPEG compression distortion

(a)

(b)

JPEG is a widely used compression format and the watermark should be resistant to it. As shown in Fig.5, with the decreasing of the quality of the JPEG compressed object, the response of the watermark detector also decreases. We have found that the proposed watermark can survive even when quality factor is as low as 20% (see Fig.4(d)), although the object is visibly distorted (see Fig.4(c))

4.2 Adding noise Noise is one of common distorts in the image processing and transmission. In the experiment, we add 15% Salt & Pepper noise and 0.1% Gaussian noise into the watermarked object as shown in Fig.6. The watermark can be retrieved successfully, and the responses of the watermark detector are 0.3591 and 0.1536 .

(c) Fig.1. (a) Original image Lena; (b) Watermarked object Lena (PSNR= 42.41 dB); (c) Detector response of the watermarked object Lena for 200 randomly generated watermark

4.3 Filtering Filtering is also one of the common image processing. The watermarked object was filtered with 3 × 3 blur filter and 5 × 5 median filter (see Fig.7). The responses of the watermark detector are 0.9017 and 0.6428 . These responses are well above the threshold Tc , even if the objects appear degraded.

(a)

(b)

(a)

(b)

(b)

(d)

(c) Fig.2. (a) Original image Baboon: (b) Watermarked object Baboon (PSNR= 41.67 dB); (c) Detector response of the watermarked object Baboon

Image Watermark using Visual Model Based Discrete Wavelet Transform

123

Fig. 3. JPEG compressed of the watermarked object Lena: (a) 60% quality object; (b) Detector response of 60% quality; (c) 20% quality object; (d) Detector response of 20% quality

(a)

(b)

Fig.6. Filtering: (a) 3 × 3 blur filtering; (b) 5 × 5 median filtering

Fig.4. Watermark detector response on the decreasing of the quality of the JPEG compressed object Lena

4.4 Rescaling Scaling is also very easy to perform during the editing of digital images. So the watermark technique must be robust to the scaling attack. We test our scheme in the case of scaling the watermarked object by 0.5 × 0.5 .The experiment results show the watermark can still be retrieved as shown in Fig.8 with the detector response 0.4802 .

(a)

(b)

Fig.7. Rescaling: (a) 0.5 × 0.5 rescaling; (b) Detector response of rescaling

4.5 Cropping Cropping is very common distort during the editing operation of digital images. In the experiment, Fig.9(a) illustrates the undisturbed watermarked image is cropped to 1/ 4 of the original size and 1/ 8 of the original size in different way [Fig.9(c) and Fig.9(e)]. The detector response for these cropping are 0.2949 , 0.1954 and 0.1736 respectively as show in Fig.9(b), Fig.9(d) and Fig.9(f).

(a)

(b)

Fig.5. Noise adding: (a) 15% Salt & Pepper noise; (b) 0.1% Gaussian noise

(a)

(c)

(b)

(d)

IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.1 No.2, April 2007

124

(e)

(f)

Fig.8. Cropping: (a) 1/4 of the original size; (c) and (e) 1/8 of the original size; (b), (d) and (f) various detector response of cropping

4.6 Rotation Rotation is a widely used editing operation of digital images and the watermark should be resistant to this distortion. We test the scheme in the case of rotation the watermarked object −3 degree. The experiment results show the watermark can still be retrieved as shown in Fig.10 with the detector response 0.2637 .

(1) It has been observed that the average has a smaller change than that of individual coefficient. Thus, unlike most watermarking schemes, the watermark is not embedded by adjust individual wavelet coefficient but by modulating the average of the wavelet blocks. (2) Visual model ployed to achieve the best tradeoff between transparent and robustness. (3) Watermark detection is accomplished without the original. (4) Mary parameters can be used as private key to that they are unknown to public. The proposed scheme was compared to the Podilchuk’s scheme [6], the results shown in Table 3. The proposed scheme was found to be superior to Podilchuk’s scheme in PNSR. Table 3. The PSNR of the proposed algorithm and Podilchulk’s algorithm (dB)

The insufficiencies of our proposed approach are that robustness is relatively inferior to the Gaussian noise and visual mode is applied roughly. Both of them need farther exploration in our work in future.

References (a)

(b)

Fig.9. Rotation: (a) Rotate -3 degree; (b) Detector response of rotation

5. Conclusions We described a new blind digital watermark scheme using DWT for images in this paper. To make the watermark robust and transparent, we embed it in the average of the low-pass wavelet blocks using visual mode. Robustness of the watermark scheme is tested against image compression and common geometrical attacks. We give a comparison of the proposed scheme to other existing schemes. Some of the following information is already mentioned in this paper, but we have it summarized at this place. There are several advantages in our proposed method:

[1] Cox.I.J, Kilian.J, Leighton.T, Shamoon.T, “Secure spread spectrum watermark for multimedia,” Proceedings of NEC, Princeton, USA, pp.95-100, Oct.1995. [2] H. Kamran. M. Adeel., Gilani.S.A.M, “Digital image watermarking in the wavelet transform domain,” Transactions on Engineering, Computing and Technology, Vol.13, No.5, pp.86-89, May. 2006. [3] Peining.T, Ahmet.M.E, “A robust multiple watermarking scheme in the discrete wavelet transform domain,” Proceedings of SPIE, Philadelphia, USA, pp.133-144, Oct. 2004. [4] Yu.G, Lu.C, Liao.H, “Average quantization blind watermark for image authentication,” Proceedings of ICIP, Vancouver, Canada, pp.706-709, Mar. 2000. [5] Xiangwei.K, Yu.L, Huajian.L, Deli.Y, “Object watermarks for digital images and video,” Image and Vision Computing, Vol.22, No.8, pp.583-595, Aug.2004. [6] Podilchuk.C.I, Zeng.W, “Image-adaptive watermarking using visual models,” IEEE Journal on Selected Areas in Communications, Vol.16, No.4, pp.525-539, May.1998.