Anda di halaman 1dari 6

Digital Watermarking For Multimedia Protection

CHI-MAN PUN and CHI-PIO TOU

Department of Computer and Information Science


University of Macau
Macau S.A.R.

ABSTRACT: - A novel blind digital watermarking scheme for multimedia protection using wavelet packet
decomposition is proposed in this paper. The original image is decomposed into wavelet coefficients using a
full wavelet packet decomposition. For minimal loss in image fidelity, the watermark is embedded in those
wavelet coefficients with highest magnitudes except for those in the lowest frequency channel. The watermark
detection is achieved without use of the original image by computing a similarity measure between the input
watermark and the wavelet coefficients of the attack image. Experimental results show that the proposed
method is robust against the common image/signal processing operations.

Key Words: blind digital watermarking, wavelet packets, multimedia protection

1. INTRODUCTION like lossy compression. Besides, altering the


significant frequency coefficients can severely
With the rapid development of digital multimedia degrade the image quality.
technology, protecting the copyright of digital media
has become more and more important. Different In order to avoid the image quality degradation, Cox
digital watermarking schemes have been proposed to et al [12] employed the spread spectrum for
address the issue of multimedia copyright protection. embedding watermarks and suggested inserting the
In general, the watermarking schemes can be watermark, which is a set of independent and
classified into two main categories: spatial domain identical distributed sequences drawn from Gaussian
approach or frequency domain approach. Most of distribution, to only a small fraction of the number
the early research on digital watermarking employs of significant frequency coefficients of the image.
the spatial domain approach and concentrates on the However, this scheme requires the original image
imperceptibility [1-7]. However, these techniques for watermark extraction, that is not preferable for
have relative low-bit capacity and are not robust resolving rightful ownership [30, 31]. Piva et al [14]
enough to lossy compression and many other signal embedded the watermarks in the middle frequency
processing operations such as geometric distortion of DCT of the images. The original image is not
and cropping. Recently many robust watermarking required for watermark extraction. However, a larger
schemes using the frequency domain approach have number of coefficients are required for a significant
been proposed. Most of these approaches are based detector response. Hsu and Wu [22] employed
on discrete Fourier transform (DFT) [8-11], cosine visually recognizable patterns as watermarks which
transform (DCT) [12-18] or wavelet transform are embedded in middle frequency coefficients of
(DWT) [19-29]. Since most of the signal processing DWT of the images due to the fact that the human
operations can be well characterized in frequency visual system (HVS) inherently performs
domain, most of these approaches embed multiresolution structure. However, the original
watermarks in low or middle frequency coefficients image is also required for watermark extraction.
and provide the superior robustness to various
attacks like lossy compression, linear or non-linear In general, most of the existing methods require
filtering, scaling, cropping, and collusion, etc. either the original image or large number of
However, human visual system (HVS) is more coefficients for watermark extraction. In this paper,
sensitive to lower frequency noise while high an effective and blind wavelet-based scheme is
frequency coefficients are vulnerable to be proposed for image watermarking using wavelet
suppressed by common signal processing operations packet transform. The original image is decomposed
into wavelet coefficients using a full wavelet packet
decomposition. For minimal loss in image fidelity,
∑ h( n)
n
2
=1 (2.3)

the watermark is embedded in those wavelet


coefficients with highest magnitudes except for
those in the lowest frequency channel. The
∑ h(n)g (n + 2 j ) = 0 , for all j
n
(2.4)
watermark detection is achieved without use of the
original image by computing a similarity measure
between the input watermark and the wavelet The 2D wavelet packet decomposition (forward 2D-
coefficients of the attack image. The outline of this DWPT) of an N × M discrete image x up to level
paper is organized as follows. In next section, we
p+1 ( p ≤ min(log 2 ( N ), log 2 ( M )) , p ∈ ) is
briefly introduce and review the standard 2-D
wavelet packets decomposition techniques. In recursively defined in terms of the coefficients of
section III, we present our proposed scheme for level p as follows:
embedding and extracting watermarks. The
experiment results for the detection responses C4pk+,(1i , j ) = ∑∑ h(m)h(n)Ckp,( m+ 2i ,n + 2 j ) (2.5)
performance and robustness of our proposed method m n

to various attacks are presented in Section IV. C4pk++11,(i , j ) = ∑∑ h(m) g (n)Ckp,( m + 2i ,n + 2 j ) (2.6)
Finally, conclusions are drawn in Section V. m n

C p +1
4 k + 2,( i , j ) = ∑∑ g (m)h(n)Ckp,( m + 2i ,n + 2 j ) (2.7)
m n

2. DISCRETE WAVELET PACKET C p +1


4 k + 3,( i , j ) = ∑∑ g (m) g (n)Ckp,( m+ 2i ,n + 2 j ) (2.8)
TRANSFORM m n

i , j ) = x( i , j ) is given by the intensity levels


0
The 2-D discrete wavelet packet transform (DWPT) where C0,(
is a generalization of 2D discrete wavelet transform of the image x.
(DWT) that offers a richer range of possibilities for
image analysis. In 2D-DWT analysis, an image is
split into an approximation and three detail images. In practice, the image x has only a finite number of
The approximation image is then itself split into a pixels. Different methods such as symmetric,
second-level approximation and detail images, and periodic or zero padding could be used for the
the process is recursively repeated. So there are n+1 boundary handling. At each step, we decompose the
possible ways to decompose or encode the image for
image Ckp into four quarter-size images
an n-level decomposition. In 2D-DWPT analysis,
the three details images as well as the approximation C4pk+1 , C4pk++11 , C4pk++12 and C4pk++13 . This decomposition
image can also be split. So there are 4n different algorithm can also be illustrated by the block
ways to encode the image, which provide a better
diagram in Fig. 1. The rows of the image Ckp are
tool for image analysis. The standard 2D-DWPT
can be described by a pair of quadrature mirror first convolved with a one-dimensional filter and
filters (QMF) H and G [32]. The filter H is a low- every other row is retained. Then the columns of the
pass filter with a finite impulse response denoted by resulting images are convolved with another one-
h(n) . And the high-pass G with a finite impulse dimensional filter and every other column is
retained.
response is defined by:

g (n) = (−1) n h(1 − n) , for all n (2.1)

The low-pass filter is assumed to satisfy the

following conditions for orthonormal representation:

∑ h(n)h(n + 2 j ) = 0 , for all


n
j≠0 (2.2)
column
s
rows G 1↓ 2 C4pk+1
Test Image
G 2 ↓1
p +1
H 1↓ 2 C 4 k +1
Wavelet Packet
Ckp Transform

G 1↓ 2 C4pk++12
H 2 ↓1
Wavelet
H 1↓ 2 C4pk++13 Secret Key Coefficients

2 ↓1 keep one column out of two


Gauassian
Watermark
Random Vector
Extraction
1↓ 2 keep one row out of two Generation

X convolve with filter X Y = { y1 , y2 ,… , yK } T * = {t1* , t2* ,… , t K *}

sim(Y , T * ) > 0?
P
Fig. 1. Decomposition of a discrete image C k into four
P +1 P +1 P +1 P +1
quarter-size images C 4k ,C 4 k +1 , C 4k +2 and C4 k +3 by
using the conjugate filters H and G. Fig.3 Procedure of detecting a digital watermark in a
test image.

Secret Key Original Image


3. DIGITAL WATERMARKING
USING WAVELET PACKET
Gauassian
Wavelet Packet
Random Vector
Transform
Generation The procedure of embedding a digital watermark
into the original image is depicted in Fig. 2. The
X = {x1 , x2 ,… , xK } T = {t1 , t2 ,… , t K } MxN original image is decomposed adaptively into
wavelet coefficients by a 2D discrete wavelet packet
Embedding [32] up to level p = min(log 2 ( N ), log 2 ( M )) / 2 . In
Watermarks
order to have better perceptual invisibility and
Watermarked
robustness to various attacks, the first K highest
wavelet coefficients
wavelet coefficients are selected to generate a vector
Inverse Wavelet
T = {t1 , t2 ,… , t K } from all channels except the
Packet Transform
lowest frequency channel. Then, a watermark
X = {x1 , x2 ,… , xK } generated by an independent
Watermarked Image
and identical distributed pseudo-random sequence
drawn from Gaussian distribution with zero mean
Fig.2 Procedure of embedding a digital watermark into
and unit variance of length K is embedded into the
the original image. vector T to obtain a new vector
T ' = {t1 ', t2 ',… , t K '} according to the following
rule.

ti' = ti + α ti xi (3.1)

where i = 1,… , K . This new vector T’ is inserted


back into the corresponding positions of the wavelet
packet decomposition. Finally, the inverse discrete
wavelet packet transform is applied to obtain the h(0) 0.01885858 h(10) -0.02082962
watermarked image.
h(1) 0.13306109 h(11) 0.02348491
h(2) 0.37278754 h(12) 0.00255022
The procedure of detecting a digital watermark in a
h(3) 0.48681406 h(13) -0.00758950
test image under attack is depicted in Fig. 3. The
attacked MxN test image is first decomposed into h(4) 0.19881887 h(14) 0.00098666
wavelet coefficients by the 2D discrete wavelet h(5) -0.17666810 h(15) 0.00140884
packet transform up to
level p = min(log 2 ( N ), log 2 ( M )) / 2 . Then, the h(6) -0.13855494 h(16) -0.00048497
first K highest wavelet coefficients are selected to h(7) 0.09006372 h(17) -0.00008235
generate a vector T * = {t1* , t2* ,… , t K *} from all h(8) 0.06580149 h(18) 0.00006618
channels except the lowest frequency channel. The h(9) -0.05048329 h(19) -0.00000938
extracted wavelet coefficient vector is used to
compute the similarity measure to the given Table 1 20-tap Daubechies wavelet transform filter
coefficients
watermark Y = { y1 , y2 ,… , yK } generated by the
input secrete key. Without use of the original image,
the similarity measure sim(Y , T * ) is defined as
follows:

K
α2 K *
sim(Y , T * ) = ∑ ( yi ti* ) 2 − ∑ (ti + α ti* yi ) 2 (3.2)
i =1 β i =1

The existence of the given watermark


Y = { y1 , y2 ,… , yK } in the test image is determined
whether the similarity measure sim(Y , T * ) is
greater than zero or not.

Fig.4 (a) The original Lena image; (b) Watermarked


Lena image, PSNR=43.33 dB.

4. EXPERIMENTAL RESULTS 40

30
A number of experiments have been carried out to
evaluate the performance of our proposed method 20
Detector Response

using the common 512x512 Lena image. For these


experiments, we have chosen K=1000, α = 0.1 and 10

β = 6 , and employed the 20-tap orthonormal


0
Daubechies wavelet [33] (where the coefficients of
the low-pass filter h are listed in Table 1) for the
-10
wavelet packet decomposition.
-20
0 500 1000 1500 2000 2500 3000
Watermarks

Fig.5 Detector response of 3000 randomly generated


watermarks. Only watermark number 1000 is correct.
Attack PSNR Sim(Y,T*) Further work can be focused to investigate the
proposed method for geometric attack and image
3x3 median filter 35.53 8.9 normalization.
5x5 media filter 33.21 8.3
Gaussian Noise (0,0.01) 37.68 9.5
REFERENCES
JPEG compression 30.33 9.3
1. Schyndel, R.G.v., A.Z. Tirkel, and C.F.
(QF = 30)
Osborne. A digital watermark. in IEEE Int.
JPEG compression 31.26 9.7 Conf. Image Processing. 1994.
2. Caronni, G. Assuring ownership rights for
(QF = 40)
digital images. in Reliable IT Systems (VIS'95).
JPEG compression 32.13 10.8 1995.
3. Macq, B.M. and J.-J. Quisquater, Cryptology
(QF = 50) for digital TV broadcasting. Proc. IEEE, 1995.
83: p. 944-957.
Table 2 Detector response of various signal processing
attacks.
4. Bender, W.R., D. Gruhl, and N. Morimoto.
Techniques for data hiding. in SPIE: Storage
and Retrieval of Image and Video Database.
First, we evaluate the perceptual quality of the 1995.
watermarked image. Fig. 4 shows the original and 5. Nikolaidis, N. and I. Putas. Copyright
the watermarked Lena image, which are visually protection of images using robust digital
indistinguishable. The PSNR is 43.33 dB. In signatures. in IEEE Int. Conf. Acoustics,
principle, it is very difficult to recognize the Speech and Signal Processing. 1996.
difference visually between the original image and 6. Wolfgang, R. and E. Delp. A watermark for
the watermarked one if the PSNR is greater than 30 digital image. in IEEE Int. Conf. Image
dB. Processing. 1996.
7. Pitas, I. A method for signature casting on
Fig. 5 shows the watermark detector response of digital image. in IEEE Int. Conf. Image
3000 randomly generated watermarks where only Processing. 1996.
the correct one (number 1000) matches the 8. A. Herrigel, J.J.K.Ó.R., H. Petersen, S. Pereira,
watermark as embebbed in Fig. 5. The much and T. Pun, Secure copyright protection
stronger response of the correct watermark suggests techniques for digital images, in Information
the very low false positive response rates of the Hiding (Lecture Notes in Computer Science),
proposed method. D. Aucsmith, Editor. 1998, Springer. p. 169-
190.
In Table 2, the robustness of the proposed method 9. Pereira, S., et al. Template based recovery of
for watermarking to most common signal processing Fourier-based watermarks using log-polar and
attacks of median filtering, additive noise and JPEG log-log maps. in IEEE Int. Conf. Multimedia
lossy compression. In general, the detector response Computing and Systems. 1999.
is much greater than zero, even when the 10. Ruanaidh, J.J.K.Ó. and T. Pun, Rotation, scale
watermarked image is highly compressed up to and translation invariant spread spectrum
quality factor as low as 50. digital image watermarking. Signal Processing,
1998. 66(3): p. 303-317.
5. CONCLUSION 11. Pereira, S. and T. Pun, Robust template
matching for affine resistant image
watermarks. IEEE Trans. Image Processing,
In this paper, we proposed a novel blind
watermarking scheme based on wavelet packet 2000. 9(6): p. 1123 -1129.
transform. Experiments show that the watermark 12. Cox, I.J., et al., Secure spread spectrum
with the size of 1000 can be correctly detected watermarking for multimedia. IEEE Trans.
without use of the original image. This proposed Image Processing, 1997. 6: p. 1673-1687.
scheme is robust to most common signal processing 13. Hsu, C.-T. and J.-L. Wu, Hidden digital
attacks of median filtering, additive noise and JPEG watermarks in images. IEEE Trans. Image
lossy compression with quality factor as low as 50. Processing, 1999. 8: p. 58-68.
14. Piva, A., et al. DCT-based watermark Industrial Electronics, 2001. 48 (5): p. 875 -
recovering without resorting to the uncorrupted 882.
original image. in IEEE Int. Conf. Image 29. Wang, S.-H. and Y.-P. Lin, Wavelet tree
Processing. 1997. quantization for copyright protection
15. Podilchuk, C. and W. Zeng. Perceptual watermarking. IEEE Trans. Image Processing,
watermarking of still images. in IEEE 1st 2004. 13: p. 154 -165.
Workshop Multimedia Signal Processing. 1997. 30. Craver, S., et al., Resolving rightful ownerships
16. Tao, B. and B. Dickinson. Adaptive with invisible watermarking techniques:
watermarking in the DCT domain. in IEEE Int. Limitation, attacks, and applicaitons. IEEE J.
Conf. Acoustics, Speech and Signal Processing. Select. Areas Commun., 1998. 16: p. 573-586.
1997. 31. Zeng, W. and B. Liu, A statistical watermark
17. Wolfgang, R.B., C.I. Podilchuk, and E.J. Delp, detection technique without using original
Perceptual watermarks for digital images and images for resolving rightful ownerships of
video. Proc. IEEE, 1999. 87: p. 1108-1126. digital images. IEEE Trans. Image Processing,
18. Hernandez, J.R., M. Amado, and F. Perez- 1999. 8: p. 1534-1548.
Gonzalez, DCT-domain watermarking 32. Mallat, S., A theory for multiresolution signal
techniques for still image: Detector decomposition: The wavelet decomposition.
performance analysis and a new structure. IEEE Trans. PAMI, 1989. 11(7): p. 674-693.
IEEE Trans. Image Processing, 2000. 9: p. 55- 33. Daubechies, I., Ten Lectures on Wavelets, in
68. CBMS-NSF Regional Conference Series in
19. Podilchuk, C.I. and W. Zeng, Image-adaptive Applied Mathematics. 1992, SIAM Press,
watermarking using visual models. IEEE J. Philadelphia, Pennsylvania.
Select. Areas Commun., 1998. 16: p. 525-539.
20. Wei, Z.H., P. Qin, and Y.Q. Fu, Percept digital
watermark of images using wavelet transform.
IEEE Trans. Consumer Electron., 1998. 44: p.
1267-1272.
21. Dugad, R., K. Ratakonda, and N. Ahuja. A new
wavelet based scheme for watermarking
images. in IEEE ICIP. 1997.
22. Hsu, C.-T. and J.-L. Wu, Multiresolution
watermarking for digital images. IEEE Trans.
Circuits Syst. II, 1997. 45: p. 1097-1101.
23. Wang, H.-J. and C.-C.J. Kuo. Image protection
via watermarking on perceptually significant
wavelet coefficient. in IEEE 2nd Workshop
Multimedia Signal Processing. 1998.
24. Lumini, A. and D. Maio. A wavelet-based
image watermarking scheme. in Int. Conf.
Information Technology. 2000.
25. Tsai, M.-J., K.-Y. Yu, and Y.-Z. Chen, Joint
wavelet and spatial transformation for digital
watermarking. IEEE Trans. Consumer
Electron., 2000. 46: p. 137-144.
26. Xie, L. and G.R. Arce. Joint wavelet
compression and authentication watermarking.
in IEEE Int. Conf. Image Processing. 1998.
27. Zhu, W., Z. Xiong, and Y.-Q. Zhang,
Multiresolution watermarking for images and
video. IEEE Trans. Circuits Syst. Video
Technol., 1999. 9: p. 545-550.
28. Hsieh, M.-S., D.-C. Tseng, and Y.-H. Huang,
Hiding digital watermarks using
multiresolution wavelet transform. IEEE Trans.

Anda mungkin juga menyukai