2009 IEEE International Advance Computing Conference (IACC 2009) Patiala, India, 6-7 March 2009

Robust Watermarking in Multiresolution Walsh-Hadamard Transform

Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee-247 667, India 1goravdma@gmail .com, [email protected] Abstract-In this paper, a newer version of Walsh-Hadamard Transform namely multiresolution Walsh-Hadamard Transform (MR-WHT) is proposed for images. Further, a robust watermarking scheme is proposed for copyright protection using MRWHT and singular value decomposition. The core idea of the proposed scheme is to decompose an image using MR-WHT and then middle singular values of high frequency sub-band at the coarsest and the finest level are modified with the singular values of the watermark. Finally, a reliable watermark extraction
the distorted image. The experimental results show better visual imperceptibility and resiliency of the proposed scheme against intentional or un-intentional variety of attacks.
scheme is developed for the extraction of the watermark from

Gaurav Bhatnagarl and Balasubramanian Raman2

quality and the smallest singular values are more sensitive to the noise. Here the middle singular values are selected
extraction scheme is introduced for finding the estimate of the watermark from both the coarsest and the finest level high frequency sub-bands. This paper is organized as follows: In section II and III introduction to the Multiresolution Walsh-Hadamard Transform and singular value decomposition. Our proposed watermarking embedding and extraction schemes are described in section IV. Section V presents experimental results using proposed watermarking scheme and finally the concluding remarks are given in section VI.
.

toaembed the watrmkurer aneint watermarkin

. .

I. INTRODUCTION

Now a days, some very crucial issues of digital media are duplication, distribution, editing, copyright protection etc. The main reason of these kind of issues is development of internet and multimedia technology. As a solution, Digital Hence, d Watermarking ius ve marking becomes very attractive research topic and many many taxonomies for digital watermarking have been proposed. Among these, the most common taxonomies are embedding in spatial[l] andtransf transform domain. Watermarking schemes of domain. Watermark transform domain, such as discrete Fourier transform (DFT)[2],

Watermarkang isltusediaerycf entloy. v

igitalw

ingsptiae[s]ea

embesdo

discrete cosine transform(DCT)[3], [4], Walsh-Hadamard Transform[9] and discrete wavelet transform(DWT)[6], [7], [8], Multi-Polarity Hadamard transform[9] are more popular, since they provide more advantages and better performances than those in the spatial domain. The wavelet-based methods become more prevalent in transform-domain watermarking algorithms due to their excellent spatial localization, frequency spread and multiresolution characteristics. In the recent years, singular value decomposition is used as a new tool for watermarking[IO], [11], [12]. The most common approach is to modify the singular values of the host image by singular values of the watermark. These modified singular values are combined with the known component to get the watermarked image.

(MR-WHT) Unlike Fourier and Cosine Transforms, in the case of WalshHadamard Transform the basic functions are not sinusoids. The fbasic functions are based on square or rectangular waves with peaks of +1. Here the term rectangular wave refers to any function of this form, where width of the pulse may vary. One primary advantage of the transform is that the computations aevr ipe hnw rjc niaeot h ai are very simple. When we project an image onto the basis functions, all we need to do is multiply each pixel by 1. Mahemat c

II. MULTIRESOLUTION WALSH-HADAMARD TRANSFORM

1 N-1 E bI(x) bi(u) (1) N f(x) (-1)i=o X=o where N = 2' and bk (z) is the kth bit in the binary representation of z. For example, if n = 3 and z = 6 ( 110 in binary) then bo (z) = 0, b1 (z) = 1 and b2 (z) = 1. As inverse Fourier and Cosine transform, inverse Walsh-Hadamard transform is defined as: ( nN-1 (1) i(O (2) fEz) z , 2

y,

WH(u)

troduced for images. The proposed watermarking scheme is based on multiresolution Walsh-Hadamard Transform(MRWHT) and SVD approaches. For embedding, the host image is transformed into frequency domain using MR-WHT and then watermark is embedded in the middle singular values of the high frequency sub-bands at the coarsest and the finest level, The largest singular values are more important to the image
978-1-4244-1 888-6/08/f$25 .00 Q 2008 IEEE 894

inage.

tipprarutlotmowtmrniin In this

paper, arobustalgorithmofwatermarkingi

along both the axis Hence the WHT of 2D functi f( ) is written as WH(u, v) =WHTYUv{WHTxW {f(x, y)}} (3) By definition, it is clear that WHT is not classified as a frequency transform, because the basis functions do not exhibit the frequency concept in the manner of sinusoidal functions.

Due to separability of the transform, two dimensional WHT can be obtained by successively taking one dimensional WHT y

yg

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.

However, it considers the number of zero crossings (or sign

changes).

III. SINGULAR VALUE DECOMPOSITION

quently

for an WHT. imageisthmodiicatinofheWashHpower transform (WHT) and is implemented in three steps:

of singular values i.e. singular values of an image are less affected if general image processing is performed. Step 1: WHT is performed on image F, which is denoted by Moreover, the size of the matrices from SVD transformation . . .is not fixed. It can be a square or a rectangle. Finally, singular Step 2: Row wise modification results *. values contain intrinsic algebraic image properties. i (:, X)
L F(:,2x-1)+F(:,2x) j

Walsh-Hadamardtrans WHT)foranimageisthe modificationloftthe

Recently,
in

Singular value decomposition is used very fredigital image processing due to excellent resist

(4)

Let A be a general real matrix or image of order m x n. The singular value decomposition (SVD) of A is the factorization A
=

+Width X ) = (:, 2x- 1) -(:, 2x) 2 Step 3: Column wise modification results L 7i(2y- 1,:)+1(2y,:) j (

1 ( :

(5)

U * S * VT
=

(12)

where U and V are orthogonal matrices of size m. x m. and n x n respectively while S diag(ui,,2, ...,or), where oJi, ( 2 i (I)r are the singular values of the matrix A with r minn(m, n) and satisfying (71 >. (2 > ... > Cr,. The first r ) 2 ( Height + (2y - 1,:) - (2y,:) (7) columns of V are called right singular vectors and the first r columns of U are called left singular vectors. Here each where F, , Tj and 2 refer to original, WHT transsingular value specifies the luminance of the image layer while formed, temporary and final MR-WHT images respectively. the corresponding pair of singular vectors specify the geometry Figure 1 shows the MR-WHT decomposition process of an of the image layer. image. Similarly, inverse MR-WHT is also implemented in IV. PROPOSED ALGORITHM three steps:

_2(y,:=

____l___
WHT

_ pera

________
Host Image
WHT Transformed Image

Opelruatmion
2-level MR-WHT

Transformed Image

Transformed Image

1-level MR-WHT

Without loss of generality, let us assume that F represents the host image of size MF x MF, W represents the initial watermark of size Mw x NW and the watermark image is smaller than the host image by a factor 2Qi and 2Q2 along both the axis, where Qi and Q2 are any integers greater than or equal to 1. Since Walsh-Hadamard transform is applied on the square matrix (images), we have taken square image as host image however for rectangular images make those images square by adding rows/columns having elements zero. Watermark is embedded in HH sub-band at the coarsest and the finest level. The block diagram of proposed embedding and extraction scheme are shown in figures 2 and 3 respectively.
MR-WHT image

Fig. 1.

Multiresolution Walsh-Hadamard Transform of an Image

Step 1: Undo Column wise modification, that results

ub-bands

SVD
Modified Frequency at

:(2y )
Step 2:
F(: 2x
/

~~~F2 (y,2+ 2(y ) + L ~~~~~F2

F2( Height
F2
(

y:
y

)+1
)

~~~~~~~~~~~~~~~~~~Watermnarked Waemre (8) Image _(8)

Neve

I nverse MR-WHT

i-level

Coarsest and Finest ~ level

High

Modified Inverse Singular Values of

_'VD

at Coarsest level

High Frequency

F2 (2y :=F2 (2y -1,:)

Undo Row wise
-

1) =F2 (: X) + L

Modification, that results

~~~~~~~~'
'

Height +

(9)

Fig. 2.
)i
1

Watermark Embedding Algorithm

High Frequency SVD Extract Singular

K . Width + ( W 22 t +

(10)
(,

Original

/-level

2x)

(:, 2x

1)

2 (

W2h

(I l)

Wm

l SVD

Step 3: Inverse WHT is performed on Flto get the reconstructed image F'. where 2, p2', 5 and F' refer to MR-WHT, temporary, reconstructed WHT and the original images respectively.

Fns ee
Fig. 3.

mg

Watermark Extraction Algorithm

2009 IEEE Internxationxal Advanxce Computing Conference (IACC 2009)

895

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.

A. Embedding Algorithm The watermark embedding algorithm is formulated as follows: 1) Perform L-level MR-WHT on the host image which is denoted by fl v where 0 e { LL, LH, HL, HH } and 1 C [l, L].* 2) Select HH sub-band from the coarsest and the finest level which is denoted by fXIH, where X C {1, L}. Here l and L stand for the coarsest and the finest level respectively. 3) The process of embedding is described as follows: . Apply SVD on both HH sub-band and watermark image, fH UfH JHH

3) In this step, sub-bands fw$ H and HfjH are used for extraction because watermark is embedded in HH subband at the coarsest and the finest level. Extraction process is described as follows: . Perform SVD on both fI and fwXI, fXHH (17) HH Sf HH VT U S H fwHH = UfWHH SfWHH VTHH (18) . The estimate of singular values of watermark is given by et SfWHH (i +p - SfHH(i +p ) P)

fw+

ffH

(1)o~()amxUH)(19)
watermark,

W= Uw Sw vf

(14)

. Perform inverse SVD to construct the extracted

* Modify the middle singular values of the HH subband as follows:

u7fYrip
new

H(i+pf)a
)

UH tZP)

where is the watermark 1, i.. (MX , Nw) and p 2, mm 3~ for choosing middle singular values from HH subband at the coarsest and the finest level. For example, if the value of po is 63 then the watermark singular values are embedded among 64 to 63+mnin(Mw, Nw) singular values of HH subband. Since, the largest singular values are more important to the image quality, and the smallest singular values are more sensitive to the noise, middle singular values are selected to embed the watermark. * Perform inverse SVD to construct the watermarked f~JH
( fHH

a.

max(u7fHH) strength, Xc= are the factors

UW(i)

(15)

U w (20) Set 4 V. RESULTS AND DISCUSSIONS The robustness of proposed scheme is demonstrated using MATLAB. Different standard gray scale images of size 512 x 512 are used as the host images namely Payaso, Yacht and Zelda. For watermark, three gray scale images of size 64 x 64 namely Peacock, Cup and IEEE CS are used. Peacock, Duck and IEEE CS are embedded into Payaso, Yacht and Zelda images respectively. The watermarked image quality is measured using PSNR (Peak Signal to Noise Ratio). The factor for choosing middle singular values (p,) is taken as 63 and 32 for coarsest and finest level respectively. To verify the presence of watermark, the correlation coefficient between the original and the extracted singular values is given by

W,XL

(w(i) - Wmeam) (W(i) -

meam)
mean)

p(wJitl)=

Ir

)HTHW

UfHH

S7/et/ V/HH

(16)

(W(i)

-Wmea

Z)

4) Map modified sub-band to its original position and Llevel inverse MR-WHT is performed to get the watermarked image.

B.Etato*
The*extaction

loih

The objective of the watermark extraction is to obtain the estimate of the original watermark. For watermark extraction, host and watermarked images, Vw, Uw and p, are required.
rocess s formuated asfollows

(21) where w, wI, Wmeam and Timeam are the original, extracted singular values, mean of original and extracted singular values and r =mim(Mw, Nw). In figure 4, all the original and watermarked images are shown. Original and extracted watermarks are shown in figure 5. The corresponding PSNR and the correlation coefficient values are given in table I.
PEAK SIGNAL TO NOISE RATIO AND CORRELATION COEFFICIENTS OF ALL TEST IMAGES

TABLE I

1) Perform L-level MR-WHT on the host and watermarked images which are denoted by fl, fwl, where 0 C { LL, LH, HL, HH } and 1 C [1, L]. 2) Select HH sub-band from the coarsest and finest level from both images, which are denoted by f,2H and to fw4afH respectively, where C {1,oL}. Here 1 and L stand for the coarsest and the finest level respectively.

Image Payaso Yacht Zelda

PSNR

P
From Finest level

48.7463 46.3570 47.8401

0.9999 1 1

From Coarsest level

0.9959 0.9997 0.9983

896

2009 IEEE Internlationdal Advance Computing Conference (IACC 2009)

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.

(a) Experimental Images

Finest

High Frequency at Extracted Watermark From Coarsest Finest

Coarsest

Finest

Coarsest

Fig. 7.

Results for Gaussian Noise Addition (100%)

(b) Watermarked Images

Fig. 4.
a) Original b) Watermarked images.

(b) Extrated Watermark Images fm

(a) Oginal Waternalrk Images +_g

6isest

leal

Extracted Watermark From High Frequency at Coarsest

Finest

(a) ~~~gina1 Watermark J m~~~~~~~ ~~~Finest

Coarsest

Finest

Coarsest

(3 Extrated Watermark Images fm ffiest leal

(b)

Fig. 5.

a) Original b) Extracted watermark images.

Fig. 8.

Results for Salt & Paper Noise Addition

(100%)

Extracted Watermark From High Frequency at Finest Coarsest Finest Finest Coarsest

(a)

Coarsest

Finest

Extracted Watermark From High Frequency at Coarsest Finest Coarsest Finest

(a-)=

Coarsest

Fig. 6.

Results for Gaussian Blurring

Fig. 9.

Results for JPEG Compression (CR=100)

To investigate the robustness of the proposed algorithm, the watermarked image is attacked by Gaussian blurring, Gaussian and Salt & Paper noise addition, JPEG compression, Row and Column deletion, Pixelation, Cropping, Vertical Flipping, Horizontal flipping, Sharpen and Wrapping attacks. After these attacks on the watermarked image, the extracted gray scale watermark is compared with the original one. In table II, correlation coefficients are given for all extracted watermark images. The most common manipulation in digital image is blurring. Watermark is extracted after applying 13 x 13 Gaussian blurring and the results are shown in the figure 6. To verify

the robustness of the watermarking scheme, another measure is noise addition. In our experiments, P% additive Gaussian and Salt & Paper noise are added in the watermarked image. In figure 7 and 8, extracted watermarks from 100% Gaussian and Salt & Paper noise attacked watermarked images are shown. For the storage and transmission of digital data, image compression is often performed to reduce the memory and increase efficiency. Hence, we have also tested our algorithm for JPEG compression(CR 100) and the results are shown in figure 9. The proposed algorithm has also been tested for rowcolumn deletion and resizing attacks. In row-column deletion, some rows and columns of the watermarked image are deleted
897

2009 IEEE Internlationlal Advance Computing Conference (IACC 2009)

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.

CORRELATION COEFFICIENTS OF ALL EXTRACTED WATERMARK IMAGES AFTER ATTACKS

TABLE II

Attacks Image

Payaso

Gaussian Blur(13 x 13) Gaussian Noise (100%) Salt & Paper Noise (100%) JPEG Compression (CR=100) Row & Column Deletion (20 row & column) Pixelation Cropping (2.5% area remaining) Vertical Flipping Horizontal Flipping Sharpen Wrapping

Finest -0.0596 0.0398 0.0215 0.5333 -0.8604 0.1771 -0.9957 -0.8441 -0.7861 -0.1998 0.7894

p Yatch Zelda Watermark Extracted from Coarsest Finest Coarsest Finest Coarsest -0.1698 -0.1763 -0.2369 -0.1393 -0.2704 0.0165 0.1461 0.2800 0.3477 0.1689 0.0403 -0.0119 0.2469 0.3621 0.1154 0.5635 0.5436 0.4265 0.4829 0.3106 -0.9354 -0.9529 -0.9500 -0.8269 -0.9695 -0.0719 -0.0400 -0.1820 0.0083 -0.1570 -0.9964 -0.9969 -0.9930 -0.9687 -0.9923 -0.1826 -0.8746 0.9534 0.7827 -0.0360 -0.9219 0.1804 0.8874 0.6571 0.8994 0.3092 0.2563 -0.2977 -0.0659 -0.3304 0.8039 -0.1158 0.7565 -0.4442 0.5685

Extracted Watermark From High Frequency at Finest Coarsest Finest Coarsest Finest

(a)

Coarsest

Extracted Watermark From High Frequency at Finest Coarsest Finest Coarsest Finest

(a)

Coarsest

(b)
Fig. 10. Results for Row-Column Deletion (Randomly deleted 20 Rows and Columns)
Fig. 12.

"1

(b)

Results for Cropping (2.5% of image area remaining)

a)
/

~~~~~Finest

Extracted Watermark From High Frequency at Finest Coarsest Coarsest Finest

~~~~~~~~~~~~~~~~~~~~~~~~~(a)
Coarsest

(b)
Fig. 11.
Results for Pixelation
(b)

randomly and then extract watermark. Figure 10 shows the result of randomly deleted 20 rows and 20 columns attack. Figure 11 shows the results of pixelation. Pixelation is the display of a digitized image where the individual pixels are apparent to a viewer. This can happen unintentionally when a low-resolution image designed for an ordinary computer display is projected on a large screen and each pixel becomes separately viewable. Another frequently used action on images is cropping. In figure 12, results for cropping are shown, when
898

Fig. 13.

Results for Vertical Flipping

only 2.5% of area is remaining. Results for Vertical flipping, Horizontal flipping, sharpen and wrapping are shown in figures 13, 14, 15 and 16 respectively. Wrapping is the process of giving 3D effect to an object by wrap a selection around a shape. Figure 16 shows the extracted watermarks when object is wrapped around spherical shape.

2009 IEEE Internlationlal Advance Computing Conference (IACC 2009)

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.

is constructed. Further, the proposed algorithm stands with various attacks, which shows the robustness of the proposed

algorithm.

ACKNOWLEDGEMENTS (a)

One of the authors, Gaurav Bhatnagar, gratefully acknowledges the financial support of the Council of Scientific
Coarsest
and

Extracted Watermark From High Frequency at Finest Coarsest Finest Coarsest Finest

nior Research Fellowship (JRF) scheme (CSIR Award no.:

Industrial

Research, New

Delhi, India

through

his

Ju-

09/143(0559)/2006-EMR-I) for his research work.

Wa1 Fig. 14.

(b)
Results for Horizontal Flipping

(a)__ Extracted Watermark From High Frequency at Finest Coarsest Finest Coarsest Finest

Coarsest

Fig. 15.

Results for

Sharpening (100%)

(b)

Electronics, 45(2), 1999, pp. 286-294. [2] I.J. Cox, J. Killian, F.T. Leighton and T. Shamoon, "Secure Spread Spectrum Watermarking for Multimedia", IEEE Transcation on Image Processing, 6(12), 1997, pp. 1673-1687. [3] M. Barni, F. Bartiloni, V. Cappellini and A. Piva, "A DCT Domain System for Robust Image watermarking", Signal Processing, 66(3), 1998, pp. 357372. [4] A. G. Bors, I. Pitas, "Image Watermarking using DCT domain constraints", Proceedings of IEEE International Conference on Image Processing, 3, 1996, pp. 231-234. [5] B.J. Falkowski, "Phase Watermarking Algorithm using Hybrid MultiPolarity Hadamard Transform," Journal of Mathematical Imaging and Vision, Springer, 30(1), 2008, pp. 13-21. [6] X. Xia, C.G. Boncelet, and G.R. Arce, "A multiresolution watermark for digital images", Proceedings 4th IEEE International Conference on Image Processing, Santa Barbara, CA, 3, 1997, pp. 548-551. 7]D. Kundur, D. Hatzinakos, "Towards robust logo watermarking using ~~~~~~~~~~meltiresolution image fusion", IEEE Transcations on Multimedia, 6, 2004,

[1] M. S. Hwang, C. C. Chang, K. F. Hwang, "A watermarking technique based on one-way hash functions", IEEE Transcations on Consumer

)REFERENCES

[8] J. Ohnishi, K. Matsui, "Embedding a seal in to a picture under orthogonal wavelet transform", Proceedings of IEEE International Conference on Multimedia and Computing system, Hiroshima, Japan, 1996, pp. 5 14-521. [9] B.J. Falkowski, "Phase Watermarking Algorithm using Hybrid MultiPolarity Hadamard Transform," Journal of Mathematical Imaging and Vision, Springer, 30(1), 2008, pp. 13-21. [10] R. Liu, T. Tan, "An SVD-Based Watermarking Scheme for Protecting Rightful Ownership", IEEE Transactions on Multimedia, 4(1), 2002, pp. [11] D. S. Chandra, "Digital image watermarking using singular value deProceedings (a)ecomposition",(MWSCAS'02), of the 45th Midwest Symposium on Circuits and Systems 3, 2002, pp. 264-267. [12] E. Ganic, AM. Eskicioglu, "Robust Embedding of Visual Watermarks Using DWT-SVD", Journal of Electronic Imaging, 14(4) (043004), 2005.
121-128.

pp. 185-197.

Finest

Extracted Watermark From High Frequency at Finest Coarsest Coarsest Finest

Coarsest

(b)

Fig. 16.

Results for Wrapping (Wrap around a circle)

VI. CONCLUSIONS

(for images) and its application in digital watermarking is proposed. Proposed algorithm is executed in the multiresolution Walsh-Hadamard domain using singular value decomposition (SVD). In which the watermark is a visually meaningful gray scale image/logo instead of a noise type Gaussian se-

In this paper, multiresolution Walsh-Hadamard transform

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 8, 2009 at 07:40 from IEEE Xplore. Restrictions apply.