[ieee conference record of the thirty-ninth asilomar conference onsignals, systems and computers,...
TRANSCRIPT
Abstract—Digital image watermark can be detected in
transform domain using maximum likelihood (ML) detection,
whereby the decision threshold is obtained using the Neyman-
Pearson criterion. A probability distribution function (PDF) is
required to correctly model the statistical behavior of the
transform coefficients. In the literature, this detection method
has been considered by modeling magnitude of a set of discrete
Fourier transform (DFT) coefficients using a Weibull PDF. In
this paper, we propose extending the Weibull model to a
generalized gamma model. For the work here, we also propose
new estimators for parameters of the generalized gamma PDF.
I. INTRODUCTION
Maximum likelihood (ML) detection scheme based on
Bayes’ decision theory has been considered for image
watermarking in transform domain [1], [5]-[7]. The Neyman-
Pearson criterion is used to derive a decision threshold to
minimize the probability of missed detection subject to a
given probability of false alarm. To achieve optimum behavior
of the ML detector, a probability distribution function (PDF)
that models correctly the distribution of the transform
coefficients is required. In [1], magnitude of a set of discrete
Fourier transform (DFT) coefficients is modeled using a
Weibull PDF. Discrete wavelet transform (DWT)
coefficients are modeled using a Gaussian PDF in [5], a zero
mean generalized Gaussian PDF in [6], and a Laplacian PDF
in [7]. Here, we consider extending the Weibull model of [1]
to a generalized gamma model.
The paper is organized as follows. The embedding rule
used with ML detection is given in Section II. The decision
rule and threshold of ML detection are briefly explained in
Section III. A more detailed description can be found in [1]
and [6]. Section IV shows the modeling of the magnitude of
the DFT coefficients by a generalized gamma PDF. We
propose a new method to estimate parameters of the
generalized gamma PDF in Section V. Experimental results
are described in Section VI and our work is concluded in
Section VII.
The notation used is as follows. Non-bold letters are used to
represent scalar quantities, sets and functions. Bold letters are
used for vectors. All vectors are real-valued and expressed in
column form.
II. WATERMARK INSERTION
LetT
Nxxx 21x andT
Nyyy 21y be
N-vectors representing magnitude of a set of DFT coefficients
of an original image and the associated watermarked image,
respectively. A watermark T
Nwww **2
*1
*w ,
chosen from a given set, is embedded into x giving y. The
elements *, ii wx and iy are viewed as realizations of the
random variables ,iX *iW and iY , with underlying PDFs
)( iX xfi
, )( ** iW
wfi
and )( iY yfi
, respectively, for
Ni ,,2,1 .
In ML detection scheme, *w is usually embedded into x
using the multiplicative rule, defined as
),1( *iiii wxy ,,,2,1 Ni (1)
where i is a positive scalar representing the embedding
strength. The embedding strength is tuned to provide a trade-
off between robustness and imperceptibility of the watermark.
III. ML DETECTION
Elements of the watermark are assumed to be statistically
independent and uniformly distributed in the interval ]1,1[ .
DFT coefficients are also assumed to be statistically
independent. The embedding strength is assumed to be much
lower than 1. Under these assumptions, the ML detection
decision rule [1] states that the watermark *w is detected if
,)(ln1
ln)(
1*
N
i
iX
ii
iX yf
w
yfz
iiy (2)
where is the decision threshold.
The Neyman-Pearson criterion can be used to obtain in
Maximum Likelihood Detection in Image
Watermarking Using Generalized Gamma
Model
T. M. Ng H. K. Garg
Department of Electrical and Computer Engineering,
National University of Singapore,
4 Engineering Drive 3, Singapore 117576
16801424401321/05/$20.00 ©2005 IEEE
such a way that the missed detection probability is minimized
subject to a fixed false alarm probability, say FAP . It is given
as [1]
)(2
)(1 2)2(erfc XX zzFAP , (3)
where
N
i
iX
ii
iX xf
w
xfzz
ii
1*
)(ln1
ln)()xy
y(x (4)
is the realization of the random variable )(Xz with mean
N
i
iX
ii
iXz xf
w
xfEzE
ii
1*)( )(ln
1ln)(XX
(5)
and variance
N
i
iX
ii
iXz xf
w
xfVzV
ii
1*
2)( )(ln
1ln)(XX .
(6)
Here, central limit theorem is used to assume that )(Xz has a
Gaussian PDF. As a rule of thumb, the value of N should be at
least 30 for the application of central limit theorem [2].
IV. GENERALIZED GAMMA MODEL
We consider modeling iX as a random variable having
generalized gamma PDF [8], defined as
i
p
i
i
i
pi
pi
iiX x
a
xx
a
pxf
iii
iii0,exp
)()(
1
, (7)
with positive shape ip , scale ia , and power i parameters,
and
0
1 0,)exp()( udtttu u is the gamma function.
Note that the exponential ( 1ii ), Weibull ( 1i ), and
gamma ( 1ip ) PDFs are included as special cases.
Substituting (7) in (2), we obtain the decision rule as
N
ip
iipi
piip
iii
i
i
wa
wyz
1*
*
)1(
1)1()(y . (8)
In view of (3), the mean and variance of
N
ip
iipi
piip
iii
i
i
wa
wXz
1*
*
)1(
1)1()(X (9)
are required to obtain the decision threshold . By
considering the rth moment of iX [8],
,)(
)( 1
i
iiri
ri
rpaXE (10)
it is straightforward to show that
N
ip
iii
piii
zi
i
w
w
1*
*
()1)((
]1)1)[(1(X) (11)
and
N
ip
iii
piiiii
zi
i
w
w
12*2
2*22
)()1)((
]1)1)][(1()()2([X .
(12)
Taking 1i , we obtain X)(z and 2)(Xz under the Weibull
model of [1]. The present of the power parameter i in (11)
and (12) provides more degree of freedom in the adjustment
of decision threshold to improve watermark detection.
V. PARAMETERS ESTIMATION
In applying the decision rule and threshold, the parameters
,, iia and ip need to be pre-determined. For blind detection
[1] of the watermark, these parameters can be estimated from
magnitude of the DFT coefficients of the watermarked image
instead of the original image. In other words, the original
image is not required in the detection process.
Estimating all three parameters of the generalized gamma
PDF is rather complex [8]. Here, we consider fixing either i
or ip , and then estimate the other two parameters. The
estimators are derived from the mean and variance of iX
given as
)(
)(][
1
i
iiii
paXE (13)
and
)(
)(
)(
)2(][
2
122
12
i
iii
i
iiii
pa
paXV , (14)
respectively.
First, if i is fixed, say 0i , by use of (13) and (14) we
define
1681
][][
][
)()2(
)()(
2
2
01
10
2
ii
i
io
ii
XEXV
XE
p
pp . (15)
As in [1], magnitude of the DFT spectrum is divided into
different regions for watermark embedding. Let R be the
region containing ix and having RN coefficients. All
coefficients in R are assumed to be identically distributed, i.e.,
they have identical PDF. The unbiased estimators of ][ iXE
and ][ iXV are
RyRi y
N
1ˆ and 22 )ˆ(
1
1ˆ
i
RyRi y
N,
respectively, where y is the corresponding magnitude of the
DFT coefficient of the watermarked image (possibly distorted)
in R. With these, we propose estimating ip as
22
21
ˆˆ
ˆˆ
ii
iip , (16)
and then ia as
)(
)(ˆˆ
10
0
i
iip
a . (17)
Similarly, if ip is fixed, say 0ppi , then we define
another function
][][
][
)()2(
)()(
2
2
10
10
2
ii
i
ii
ii
XEXV
XE
p
p (18)
and propose estimating i and ia as
22
21
ˆˆ
ˆˆ
ii
ii (19)
and
)(
)(ˆˆ
10p
a
i
iii , (20)
respectively.
One way to solve (16) and (19) is to approximate the
inverse functions 1 and 1 using any of the well-known
function interpolation methods [3]. Note that for the
interpolation process to be practical, the range of the functions
and must be finite. In Appendix A, we show that both
functions are indeed of finite range.
VI. EXPERIMENTAL RESULTS
Using 512512 test images as shown in Fig. 1, three
special cases of the generalized gamma model are explored,
(i) 1i (ii) 1.1i (iii) 1ip . In (i), we have the Weibull
model. This is similar to the model of [1] except that in [1] ia
and ip are estimated using maximum likelihood method
which requires the use of iterative procedure like the Newton-
Raphson method. In this regard, the proposed estimators are
considered simpler. We refer to (ii) as GG model
with 1.1i . In (iii), we have the gamma model. All these
models are selected because the plot of their PDFs closely
resembles the shape of the histograms derived from subsets of
Nxxx ,,, 21 .
In [1], the DFT magnitude spectrum is divided into 16
regions. Here, for simplicity of testing, we consider only two
regions, shown in Fig. 2 as two squares at the upper half of the
DFT matrix. Each region has 2500 identically distributed
coefficients. The watermark is embedded into these
coefficients and then duplicated to the corresponding
coefficients in the two regions at the lower half of the DFT
matrix. This is done to preserve the symmetry property of the
DFT magnitude spectrum. A constant embedding strength
3.0 is used for all the coefficients.
With 610FAP , robustness of the watermark is tested
under different standard image processing operations. Table I
shows the results for watermarked images compressed by
JPEG with 70% quality factor. In Table II, each watermarked
image is cropped to retain only 400400 pixels at the center.
The missing portion is replaced by zero pixels so that the size
of each image remains at 512512 . In Table III, they are
corrupted by Gaussian noise of zero mean and variance equals
to 0.5. These results are based on 10,000 trials. In each trial,
we embed into each image a watermark *w chosen from a set
W of 100 randomly generated watermarks. A detection is said
to be successful only if the decision threshold is exceeded for
the watermark *w but not for any other watermarks in W. In
general, our simulation results reveal that watermark in JPEG
compressed images can be detected more effectively under
GG model with 1.1i . However, the Weibull model
performs better when the images are cropped. All models
perform equally well under image scaling.
VII. CONCLUSION
An ML watermark detection scheme based on using a
generalized gamma PDF to model magnitude of a set of DFT
coefficients is formulated. By use of the Neyman-Pearson
criterion, a decision threshold is explicitly derived. We have
also proposed a new method to estimate parameters of the
generalized gamma PDF. These estimators are also useful in
area like reliability analysis where the generalized gamma
PDF is widely used.
1682
Further research will be focused on giving a more
comprehensive performance evaluation of the generalized
gamma model that includes robustness under other standard
image processing operations. In particular, it is desirable to
look for an optimum value of i that can yield good detection
over a wide range of images.
APPENDIX A
The Weierstrass formula [4] for gamma function is given as
1
1
/1)(
k
kuu ek
uueu , (21)
where }ln)/13/12/11{(lim mmm
is a positive
constant called the Euler-Mascheroni constant. Using (21), we
can write the function )( ip in (15) as
1
1
2
0
)2(
0
2
1
1
0
)1(
0
00
00
21
)2(
11
)1(
)(
k
kp
p
i
ip
p
i
i
k
kp
p
i
ip
p
i
i
i
i
i
i
i
i
i
i
i
ekp
pe
p
p
ekp
pe
p
p
p
1
1
/00
00 1
1
k
ke
ke
1
20
002
0
00
)1(
))(2(
)1(
)2(
k ii
iii
i
ii
pkp
kpkpp
p
pp.
(22)
The first derivatives of 2000 )1/()2( iii ppp and
2000 )1/())(2( iiiii pkpkpkpp can be shown
to be positive. This implies that is a strictly increasing
function. Furthermore, it can be verified from (15) and (22)
that 1)(lim ip
pi
and 0)(lim0
ip
pi
, respectively. Hence,
we conclude that the range of is finite. Using the same
approach, we can also show that the range of is also finite.
REFERENCES
[1] M. Barni, F. Bartolini, A. De Rosa and A. Piva, “A new decoder for the
optimum recovery of nonadditive watermarks,” IEEE Trans. Image
Processing, vol. 10, no. 5, pp. 755-766, 2001.
[2] J. L. Devore, Probability and Statistics for Engineering and the
Sciences, Brooks/Cole, USA, 2004.
[3] D. Kincaid and W. Cheney, Numerical Analysis, 2nd ed., Brooks/Cole
Pub. Co., 1996.
[4] S. G. Krantz, Handbook of Complex Variables, Birkhauser, Boston,
1999.
[5] S.-G. Kwon, S.-H. Lee, K.-K. Kwon, K.-R. Kwon and K.-I. Lee,
“Watermark detection algorithm using statistical decision theory,” Proc.
IEEE Int. Conf. on Multimedia and Expo, Vol. 1, pp. 561-564, 2002.
[6] T. M. Ng and H. K. Garg, “Wavelet domain watermarking using
maximum-likelihood detection,” Proc. SPIE Conf. on Security,
Steganography, and Watermarking of Multimedia Contents VI, vol.
5306, San Jose, pp. 816-826, 2004.
[7] T. M. Ng and H. K. Garg, “Maximum-likelihood detection in DWT
domain image watermarking using Laplacian modeling”, IEEE Signal
Processing Letters, pp. 285-288, 2005.
[8] E. W. Stacy and G. A. Mihram, “Parameter estimation for a generalized
gamma distribution,” Technometrics, vol. 7, no. 3, pp. 349-358, 1965.
Baboon Barbara Bridge F16
Lena Fishing Boat Peppers Pentagon
Fig 1. Test Images
Watermarked region
Symmetric coefficients
Fig 2. Watermark region in DFT (magnitude) matrix
TABLE I. PERCENTAGE OF SUCCESSFUL DETECTIONS UNDER JPEG
COMPRESSION
Image Weibull GG ( 1.1i ) Gamma
Baboon 100.00 100.00 100.00
Barbara 97.88 99.07 94.09
Bridge 100.00 100.00 100.00
F16 98.15 98.87 97.33
Lena 93.61 99.53 92.04
Fishingboat 94.59 99.02 90.17
Peppers 94.58 98.82 91.64
Pentagon 100.00 100.00 100.00
TABLE II. PERCENTAGE OF SUCCESSFUL DETECTIONS UNDER IMAGE CROPPING
Image Weibull GG ( 1.1i ) Gamma
Baboon 100.00 100.00 100.00
Barbara 99.81 99.07 99.19
Bridge 100.00 100.00 100.00
F16 100.00 100.00 100.00
Lena 100.00 100.00 100.00
Fishingboat 99.93 99.12 99.89
1683
Peppers 100.00 100.00 100.00
Pentagon 90.32 89.57 89.49
TABLE III. PERCENTAGE OF SUCCESSFUL DETECTIONS UNDER IMAGE SCALING
Image Weibull GG ( 1.1i ) Gamma
Baboon 100.00 100.00 100.00
Barbara 100.00 100.00 100.00
Bridge 100.00 100.00 100.00
F16 100.00 100.00 100.00
Lena 100.00 100.00 100.00
Fishingboat 100.00 100.00 100.00
Peppers 100.00 100.00 100.00
Pentagon 100.00 100.00 100.00
1684