SlideShare a Scribd company logo
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
DOI : 10.5121/ijaia.2010.1304 45
PERCEPTUAL COPYRIGHT PROTECTION USING
MULTIRESOLUTION WAVELET-BASED
WATERMARKING AND FUZZY LOGIC
Ming-Shing Hsieh
Department of Computer Engineering and Information Science,
Aletheia University, Damshui, Taiwan 251
sms.sms@msa.hinet.net
ABSTRACT
In this paper, an efficiently DWT-based watermarking technique is proposed to embed signatures in images to
attest the owner identification and discourage the unauthorized copying. This paper deals with a fuzzy inference
filter to choose the larger entropy of coefficients to embed watermarks. Unlike most previous watermarking
frameworks which embedded watermarks in the larger coefficients of inner coarser subbands, the proposed
technique is based on utilizing a context model and fuzzy inference filter by embedding watermarks in the
larger-entropy coefficients of coarser DWT subbands. The proposed approaches allow us to embed adaptive
casting degree of watermarks for transparency and robustness to the general image-processing attacks such as
smoothing, sharpening, and JPEG compression. The approach has no need the original host image to extract
watermarks. Our schemes have been shown to provide very good results in both image transparency and
robustness.
KEYWORDS
digital watermarking, discrete wavelet transform, fuzzy inference filter, adaptive quantization, entropy.
1. INTRODUCTION
The increasingly easy access to digital media and increasingly powerful tools available for
manipulating digital media have made media security a very important issues. Due to the open
environment of Internet downloading, copyright protection introduces a new set of challenging
problems regarding security and illegal distribution of privately owned images. One potential solution
for declaring the ownership of the images is to use watermarks to embed an invisible signal into
multimedia data so as to prove the owner identification of the data and discourage the unauthorized
copying. In the study, we focus our research on digital image watermarking, but the method could be
modified for other applications as well.
In general, there are two common requirements of watermark. First requirement is the watermarks
must be perceptually transparency. They should not noticeable to the viewer. The second requirement
is the watermarks must be robust to intentional or unintentional attacks and common signal processing.
Particularly, the watermark should still be detectable even after attacks have been applied to the
watermarked image. These attacks include image compression, linear or nonlinear filtering, image
enhancements, etc.
The watermarking techniques can be divided into two different classifications. One is applied to
spatial domain, and the other is applied to the frequency domain. The spatial domain watermark
techniques are developed early [1, 2], simple but not robust is their obvious weakness. They can’t
against intentional or unintentional attacks and image processing. The embedded watermark signals
are easily disfigured, distorted, or removed. The frequency domain approach has some advantages
because most of the signal processing operations can be well characterized, and many good perceptual
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
46
models are developed in the frequency domain [3].
Cox et al. [4] proposed an image watermarking method based on spread spectrum theory, which
shows good performance among invisibility and robustness to signal processing operation and
common geometric transform. Hsu et al. [5] proposed DCT based image watermarking technique by
embedding a visually recognizable watermark. An image-content-based adaptive embedding scheme
is applied in discrete Fourier transform (DFT) domain of each perceptually high textured subimage to
ensure better visual quality and more robustness [6]. On the other hand, several methods used the
discrete wavelet transform (DWT) to hide data to the frequency domain to provide extra robustness
against attacks.
DWT is known that the wavelet image/video coding, such as embedded zero-tree wavelet (EZW)
coding [7], and play an important role in image/video compression standards, such as JPEG2000 and
MPEG4 due to its excellent performance in compression. In [8], they propose watermarking schemes
where visual models are used to determine image dependent upper bounds on watermark insertion.
This allows users to provide the maximum strength transparent watermark, which is extremely robust
to common image processing and editing such as image compression. Dugad et al. [9] picked all
coefficients in the subbands, which are larger than a given threshold, and watermark is added to these
coefficients only. The reason to set the threshold is to embed the watermarks in the edge regions of
the host image, so that the watermarked image is not distinguishable from the original image. Hsu et
al. [3] used DWT other than DCT and the same method proposed in [5] to embed signal into each
wavelet subbands. The empirical results showed that the watermarked image was more robustness and
imperceptibly then they early did in [3]. Inoue et al. [10] proposed a method based on zerotree [11],
which classified wavelet coefficients as insignificant or significant using zerotree, and select proper
position to insert watermarks. Embedding watermarks into host image using DWT are both in
transparency and robustness.
In [4] the watermark is a symbol or a random number, which comprises of a sequence of bits, and can
only be detected by employing detection theory. In [12], [13], the watermark is a visually
recognizable pattern. This kind of watermark is more intuitive for representing one’s identity than a
random sequence, and also can be measured the correlation between the detected visual pattern and
original watermark during the verification phase.
Originally, the determination of a wavelet coefficient to be an embedded target is done by comparing
subband’s coefficients uniformly with a fixed sorting order. If all coefficients of a wavelet subband
are less than the threshold value, they are set “insignificant” and had no chance to embed watermarks.
Actually, the determination is uncertain and mutual influence. For example, all coefficients of a
subband being slightly less than the threshold value should be not always set “insignificant”; on the
other hand, all coefficients of a subband being slightly greater than the threshold value should be not
always set “significant”. Therefore, a fuzzy filter is necessary to provide reasoning with the vague and
uncertain information. The use of fuzzy filters in image processing is based on the idea that pixels or
coefficients are not uniformly fired by every fuzzy rule. The fuzzy set theory has the potential
capability to efficiently represent input and output relationships of dynamic systems, thus it has
gained popularity. For example, the usage of fuzzy algorithms for vector quantization has been
proposed by Munteanu et al. [14]; fuzzy clustering was employed to preserve the textually important
image characteristics while a compression algorithm was proposed by Karras et al. [15]; Yang and
Toh [16] applied heuristic fuzzy rules to improve the performance of the traditional multilevel filter;
Farbiz et al. [17] applied a new fuzzy logic filter to improve the performance of image enhancement.
Hsieh et al. [18] applied a new fuzzy logic filter to improve the performance of image coding. In [19],
an adaptive watermarking algorithm is presented which exploits a biorthogonal wavelets-based human
visual system (HVS) and a Fuzzy Inference System (FIS) to protect the copyright of images in
learning object repositories. The FIS isutilized to compute the optimum watermark weighting function
thatwould enable the embedding of the maximum-energy and imperceptiblewatermark.
In this paper, we propose a wavelet-based watermarking approach by adding visually recognizable
image to the larger entropies of coefficients calculated by the fuzzy inference filter of selected wavelet
subband. The proposed approach has the following advantages: (i) the extracted watermark is visually
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
47
recognizable to claim one’s ownership, (ii) the approach is hierarchical and has multiresolution
characteristics, (iii) the embedded watermark is hard detected by human visual perceptivity. Our
experimental results show that the proposed watermarking approach is very robust to image
compression and image operations.
This paper is organized as follows. Wavelet transform of images, context-based method and fuzzy
filter is described in Section 2. Section 3 describes the watermark embedding approach. In Section 4,
the experimental results are shown. The conclusion of this paper is stated in Section 5.
2. PRELIMINARIES
2.1. Wavelet transform of images
The wavelet transform is identical to a hierarchical subband system, where the subbands are
logarithmically spaced in frequency. For an input sequence of length N, DWT will generate an output
sequence of length N. The 1-D DWT can be implemented by the Direct Pyramid Algorithm, which
was developed by Mallat [20] as follows.
Begin {Direct Pyramid Algorithm}
For k = 1 to K
For n = 1 to 2K-k
Y[k, n] = ∑
−
=
−−
1
0
21
N
m
ngmnkX )(],[
X[k, n] = ∑
−
=
−−
1
0
21
N
m
nhmnkX )(],[
End
where k is the current octave, K represents the total number of octaves, 1 ≤ k ≤ K, n is the current
input, and N is the total number of inputs, 1 ≤ n ≤ N. Mallat’s direct pyramid algorithm is executed by
down sampling.
For a 2-D image, a wavelet Ψ and a scaling function Φ are chosen such that the scaling function ΦLL(x,
y) of low-low subband in a 2-D wavelet transform can be written as ΦLL(x, y) = Φ(x) Φ(y). Three other
bi-dimensional wavelets can also be obtained using the wavelet associated function Ψ(x) as follows.
ΨLH(x, y) = Φ(x) Ψ(y) ; horizontal
ΨHL(x, y) = Ψ(x) Φ(y) ; vertical
ΨHH(x, y) = Ψ(x) Ψ(y) ; diagonal
where H is a high-pass filter and L is a low-pass filter.
The basic idea in the DWT of a 2-D image is as follows. An image is firstly decomposed into four
parts of high, middle, and low frequencies (i.e., LL1, HL1, LH1, HH1) subbands, by cascading
horizontal and vertical two-channel critically subsampled filter banks. The subbands labeled HL1, LH1,
and HH1 represent the finest scale wavelet coefficients. To obtain the next coarser scale of wavelet
coefficients, the subband LL1 is further decomposed and critically subsampled. This process is
continued an arbitrary number of times, which is determined by the application at hand. Fig. 1 shows
layout of the image subbands from three-level dyadic decomposition and an example of DWT
decomposition of the Lena image using a wavelet filter set. In the figure, Lena image is decomposed
into ten subbands for three scales. Each level has various band-information such as low-low, low-high,
high-low, and high-high frequency bands. Furthermore, from these DWT coefficients, the original
image can be reconstructed. The reconstruction process is called the inverse DWT (IDWT). Let I [m, n]
represent an image. The DWT and IDWT for I [m, n] can be similarly defined by implementing the
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
48
DWT and IDWT for each dimension m and n separately: DWTn [DWTm I [m, n]].
The wavelet transform is identical to a hierarchical subband system, where the subbands are
logarithmically spaced in frequency and represent an octaveband decomposition. An example of three
scales wavelet transform is show in Fig. 1. The image is first decomposed into for subband LL1, LH1,
HL1, HH1 by using horizontal and vertical two channel critically subsampled filter banks. Each
coefficient represents a spatial area corresponding to approximately a 2×2 area of the original image.
The LH1, HL1, HH1 represent the finest scale wavelet coefficient. After decomposed and critically
subsampled the subband LL1, we can obtain the next coarser scale of wavelet coefficients. Note that
for each coarser scale, the coefficients represent a larger spatial area of the image but a narrower band
of frequencies. The decomposed process continues until some final scale is reach.
LL3 HL3
LH3HH3
HL2
LH2 HH2
HH1LH1
HL1
(a) (b)
Fig. 1. (a) Layout of the image subbands from three-level dyadic decomposition. (b) An example
of DWT decomposition of the Lena image.
2.2. Context-based method
The main function of the context-based method is to find the relationship of one pixel/coefficient and
its adjacent pixels/coefficients in the image. The relationship among the context is to determine
characteristics of the target pixel/coefficient. Context-based method can be applied in image
compression, which is based on wavelet transform. Triantafyllidis et al. [21] proposed that current
wavelet coefficient could be estimated using the coefficients that lie in the current band, scale
bound(s), parent band, and the pyramid structure. Yoo et al. [21] introduce a context-based
classification technique, which classifies each subband coefficient based on the surrounding
coefficients, and different quantizer is then used for each class. In our approaches, we introduce
context-based method on watermark embedding. Unlike previous works that only use the larger
coefficients to insert watermarks, we use the current coefficient and its surrounding coefficients to
calculate the entropy of each coefficient in selected subband, and choose the larger entropy to embed
watermarks. If a coefficient has higher entropy denotes the violent variation of spatial domain in the
host image, to embed watermarks in the center coefficient of the context could improve the
transparency of the watermarked image and robustness to extracted watermarks.
2.3. Using fuzzy filter to calculate entropy of wavelet coefficient
In this section, the use of context and fuzzy filter to calculate the entropy corresponding to each
coefficient in one subband are described. General block diagram of a fuzzy inference system is shown
in Fig. 2.
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
49
Fig. 2. General block diagram of a fuzzy inference system (FIS).
In an arbitrary wavelet subband S, where we determine entropy of each coefficient in S is based on
current coefficient and its surrounding coefficients. Template made of nine coefficients form the
context is shown in Fig. 3.
Let x0 be the current target coefficient to estimate its entropy, xi, 1≤i≤8, is a x0’s surrounding
coefficient as shown in Fig. 3. Fuzzy rules shown in Table 1 are used as inference filter to calculate
the entropy corresponding to each coefficient in S:
x 1 x 2 x 3
x 4 x0 x 5
x 6 x 7 x 8
Fig. 3. Template made of nine
coefficients form the context.
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
Ratio
Degreeofmembership
Very Low Low Medium High Very High
Fig. 4. Membership functions depicting the degrees of
normalized fuzzy wavelet coefficients.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 5. Typical membership function for the fuzzy function “more”.
Fuzzifer
Fuzzy Inference
Engine
DeFuzziferFuzzy Rule Base
Crisp
input
x, y
Crisp
ouput
z
µ(x, y) µ(z)
Fuzzy Fuzzy
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
50
Table 1. Fuzzy Inference rules
Rule Rule Description
R1
R2
R3
R4
R5
R1:If more of xi are Very High Then En(xi) is Very High.
R2:If more of xi are High Then En(xi) is High.
R3:If more of xi are Medium Then En(xi) is Medium.
R4:If more of xi are Low Then En(xi) is Low.
R5:If more of xi are Very Low Then En(xi) is Very Low.
Note that in the above production rules in Table 1, xi’s, named normalized fuzzy coefficients (NFCs),
are the normalized fuzzy value (NFV) of wavelet coefficients. NFV of each coefficient is normalized
by dividing some of T to emphasis the importance of coefficients in its context. The predicate’s and
conclusion’s part, ‘Very High’, ‘High’, ‘Medium’, ‘Low’, and ‘Very Low’ are the degrees of fuzzy
membership as illustrated in Fig. 4. The output En(xi) denotes the entropy of xi, is given by the target
coefficient and its surrounding coefficients which yield the result by Equ (3). The term more denotes
an S-type fuzzy function whose typical shape is shown in Fig. 5. The curve of this function enables
the non-uniform firing of the basic fuzzy rules. This more function may be described by the following
formula:
)(
1
1
)( βα
µ −−
+
= zmore
e
z (1)
where α=7.80375, β=3.29596. The curve of the function enables the non-uniform firing of the basic
fuzzy rules.
NFV of each coefficient is defined as follows.
0
0
0
T
x
x = ,
1
7,5,4,2|
T
x
x n
nn == ,
2
8,6,3,1|
T
x
x m
mm == ,
where T0 = Avg(|S|), T1= T0+1/16×Std(|S|), T2= T0+1/8×Std(|S|), Avg, Std denote the average, stand
deviation, respectively.
The activity degree of rule Rk is computed by the following relationship:
×∈= )}sup(:)(min{ qualityxx iiqualityk µλ





 ∈
i
ii
more
xofnumbertotal
qualityxwhichxofnumber )sup(
µ (2)
where 1≤ k ≤ 5, and the corresponding quality ∈ {Very High | High | Medium | Low}.
After all λk’s are calculated, the entropy of coefficient xi is computed by equation:
∑=
=
5
1
)(
k
kki CxEn λ (3)
where Ck represents the center point of the membership function Rk.
3. WATERMARKING IN THE DWT DOMAIN
The proposed digital watermarking approach can hide visually recognizable patterns in images. The
main study task of digital watermarking is to make watermarks invisible to human eyes as well as
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
51
robust to various attacks. In the proposed approach, watermarks are embedded in the host image by
modifying the coefficients with larger weighted energy. The block diagram of embedding watermarks
in the host image is shown in Fig. 6.
Host
image DWT
Select
embedded
subband
Calculate
entropy
coef.'s
Select
embedded
coef.
Sorting/
Quantization
Embedded
position
Sorting
map
Calculate
weight
T0-T2
Embedding
Watermark Sorting
Sorting map
parameter
v
IDWT
Key
Watermarked
image
Calculate
casting
weight v
Fig. 6. The block diagram of embedding watermarks.
3.1. Watermark embedding method
The algorithm for embedding a watermark in a host image is described as follows:
Step 1: Sort the gray levels of watermark W in ascending order to generate the sorted watermark SW,
Assume that the size of watermark image W is n.
Step 2: Decompose a host image H into three levels with ten subbands of a wavelet pyramid structure.
Choose a subband S, for example HL3, to embed watermark W.
Step 3: Calculate the weighted entropy En of coefficients in subband HL3 using the proposed method
approach described in Section II.B.
Step 4: Let the preset interval be τ, t be the number of referenced coefficients used as a key to extract
watermark W without the host image. p coefficients with larger entropy are chosen from
subband S, where p = n + t. The larger entropy of coefficients makes the watermarked image
more robustness and transparency. Obviously, if n/τ=0, the value of t is fix(n/τ) + 1, otherwise
t is fix(n/τ) + 2, where function fix is to get the integer part of its argument. Let {Cp} be the
set of referenced coefficients and the coefficients to be embedded watermarks; {Cp} is called
the alternative coefficients. Sorting {Cp} to generate {SCp} called the sorted alternative
coefficients.
Step 4: Quantize {SCp} using a preset interval could extract watermark W without the host image.
Step 6: Embed watermark SW into subband HL3 by the watermark embedding strategy, SCj = SCj +
vi×SWindex, where SCj is a sorted alternative coefficients except the referenced coefficients, vi =
Eni × (T0+T1+T2)/3 = Eni × T1. The value of scaling factor vi is the maximum variances to
modify the embedded coefficients for robustness while embedding watermarks to the target
coefficients and SWindex is a sorted waremark. A coefficient with larger weighed entropy could
embed watermark with larger scaling factor for robustness without obviously degrading the
host image. The embedding procedure can be implemented by the proposed Casting
watermarks Algorithm as follows.
Begin {Casting Watermarks Algorithm}
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
52
index=1
/* Embedded watermarks for each intervals with τ coefficients. */
For i = 1 to p-τ step τ+1
For j = i+1 to i+τ
SCj = SCj + vj×SWindex
index = index + 1
EndFor
EndFor
/* Embedded watermarks for the rest alternative coefficients. */
If (n mod τ) ≠ 0
For j = p-(n mod τ) to p-1
SCj = SCj + vj×SWindex
index = index + 1
EndFor
EndIf
End {Casting Watermarks Algorithm}
Step 7: Save scaling factor vi, the symbol of embedded subband, symbol table of SCi, corresponsive
map of Ci and SCi, and corresponsive map of Wi and SWi. To form watermarked image, one
can take the IDWT of the modified DWT coefficients and the unchanged DWT coefficients.
3.2. Watermark extracting method
The embedded watermark can be detected based on the stored parameters after the wavelet
decomposition of the watermarked image as follows:
Step 1: Decompose a watermarked image into three levels with ten subbands using DWT.
Step 2: Restore the scaling factor vi, the symbol of embedded subband, symbol map of SCi,
corresponsive map of Ci and SCi, and corresponsive map of Wi and SWi.
Step 3: Extract the sorted watermarks by the proposed Extracting Watermarks Algorithm as follows:
Begin {Extracting Watermarks Algorithm}
index=1
/* Extracting watermarks for each intervals with τ coefficients. */
For i = 1 to p-τ step (τ+1)
For j = i+1 to i+τ
value = (|SCi|+ |SCi+τ+1|) / 2
SWindex = (|SCj|– value) / vj
index = index + 1
EndFor
EndFor
/* Extracting watermarks for the rest alternative coefficients. */
If (n mod τ) ≠ 0
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
53
For j = p-(n mod τ) to p-1
value=(|SCp-(n mod τ)-1+ |SCx’p|) / 2
SWindex = (SCx’j – value) / vj
index = index + 1
EndFor
EndIf
End {Extracting Watermarks Algorithm}
Step 4: Rearranging watermarks from corresponsive map of Wi and SWi to get the extracted
watermarks W’.
In our scheme, the extracted watermark W’ is a visually recognizable image. A subjective
measurement based on the standard correlation coefficient,
∑∑
∑
−−
−−
=
22
)()''(
))(''(
WWWW
WWWW
ncorrelatio (4)
is used to evaluate the quality of the extracted watermarks by measuring the similarity of the original
watermark W and extracted watermark W’. The value of correlation is between zero and one. A larger
value of correlation represents more similarity of the original watermarks and extracted watermarks.
4. EXPERIMENTS
A 512×512 Lena image was taken as the host image to embed a 32×32 binary watermark image with
“NCU CSIE” characters following a sequence of attacks. The proposed approach emphasizes the local
characteristics in the experiments, we also examined the quality of watermarked images and
detectability of watermarks.
4.1. Image quality
The proposed perceptual watermarking framework was implemented for evaluating both properties of
transparency and robustness. The two conflicting characteristics are referred to the embedded
positions and casting degree of watermarks. For the embedded positions, as we chose the larger
entropy of coefficients to embedded watermarks, which stands for the violent variance of spatial
domain in a host image, the transparency and robustness of host image could be hewn. For the casting
strength, the larger casting degree makes it more robust to the watermarked image. The side effect is
that the lower transparence makes the watermarks the perceptual break. In the experiments, an
adaptive casting strength was proposed to adjust casting degree of the coefficients, which makes the
balance of transparency and robustness. Fig. 7 shows an example of embedding results, where Lena is
used as the test image, a binary image with “NCU CSIE” is used as the watermark. The performance
of the proposed image watermarking approach was evaluated. The common-used 512×512 gray-scale
Lena image was used as original image and the peak signal-to-noise ratio (PSNR) was used for
objective comparison,
MSE
PSNR
2
10
255
10log= , (5)
where MSE is the mean square error between a watermarked image and its original image.
Fig. 7 also shows the watermarked image, in the figure, the PSNR of the watermarked image is 45.68,
correlations of the original watermarks and extracted watermarks is 1.
Table 2 shows an example of extracting results from Fig. 7(b) without any attacks using the proposed
method.
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
54
(a) (b)
(c)
Fig. 7. Example of digital watermarking using visual watermarks, where (a) is the host
Lena image, (b) is the watermarked image with PSNR= 45.68 after embedding, (c) the
watermarks.
Table 2. The extracted watermarks from Fig. 7 are visually recognizable
Embedded watermarks
PSNR (dB) after embedding watermarks 45.68
Extracted watermarks
Correlation 1
Error bits 0
4.2. On the robustness against JPEG lossy compression
Table 3 shows the extracted results from JPEG compressed version of the watermarked images with
different compression quality. The quality of watermarked images is still in good situation even under
the high compressed ratio. The extracted watermarks and the original watermarks are with high
correlation. In the proposed method, error rate of the extracted watermarks is rarely small (<0.5%)
even under the situation of compression ratio, 16.59. From the experimental results, we can find that
the proposed watermarking techniques yield satisfactory results in terms of transparency for
watermarked images.
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
55
Table 3. Changes of correlation and error rate values of the proposed method under JPEG lossy
compression.
Attack
Extracted
Watermark
Correlation Error rate
Quality=100, CR=1.6810 1 0
Quality=90, CR=4.6681 1 0
Quality=80, CR=7.1466 1 0
Quality=70, CR=9.1314 1 0
Quality=60, CR=10.9222 1 0
Quality=50, CR=12.4298 0.9972 0.09%
Quality=40, CR=14.1922 0.9972 0.09%
Quality=30, CR=16.5903 0.9864 0.49%
Quality=20, CR=20.5619 0.8022 7.8%
*CR=compression ratio
4.2. On the robustness against JPEG lossy compression
Sharpen operations are used to enhance the subjective quality. Table 4 shows the extracted results of
applying enhanced operation to a watermarked image. The extracted results are highly similar to the
original watermark.
Smoothing operations such as median filter are used to decrease spurious effects that may be present
in images from a poor transmission channel. Table 4 shows the extracted results of applying median
filter to a watermarked image. The extracted watermark is still visually recognizable.
Table 4. The extracted of watermarks, correlation and error rate under sharpen and smoothing
operation.
Attack Extracted Watermark Correlation Error rate
Sharpen 0.9945 0.19%
Median filter 0.9488 1.8%
5. CONCLUSION
We have introduced a watermarking framework for embedding visually recognizable watermarks in
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
56
images, which can resist image-processing attacks, such as JPEG compression, smoothing,
sharpening, blur, etc. The proposed new techniques are based on the context ideas and fuzzy filter. In
the experiments, the fuzzy filter is employed to conclude the entropy of each coefficient in the center
of the context. The watermark embedding strategies of the proposed methods consider the local
characteristics by choosing the larger-entropy coefficients of DWT subbands to embed watermarks.
The experimental results show that the proposed methods provide extra robustness against
JPEG-compression and image processing compared to the traditional embedding methods. Moreover,
the embedded coefficients in the proposed approaches have their own embedding degrees, which are
calculated automatically in accordance with the entropy contributed from the context. Finally, the
proposed approaches have no need the original host image to extract watermarks.
ACKNOWLEDGEMENTS
The authors would like to thank Mr. Y.-H. Huang, Prof Tseng D-C, and the anonymous reviewers of this
paper.
REFERENCES
[1] R. G. van Schyndel, A. Z. Tirkel, and C. F. Osborne, (1994) “A digital watermark,” Proc. of IEEE Int.
Conf. Image Processing, Vol. 2, pp. 86-90.
[2] R. Wolfgang and E. Delp, (1996) “A watermark for digital image,” Proc. Int. Conf. on Image Processing,
(3), pp. 211-214.
[3] C.-T. Hsu and J.-L. Wu, (1998) “Multiresolution watermarking for digital images,” IEEE Trans.
Consumer Electronics, Vol. 45, No. 8, pp. 1097-1101.
[4] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, (1997) “Secure spread spectrum watermarking for
multimedia,” IEEE Trans. Image Processing, Vol. 6, No. 12, pp. 1673-1687.
[5] C.-T. Hsu and J.-L. Wu, (1999) “Hidden digital watermarks in images, ” IEEE Trans. Image Processing,
Vol. 8, No. 1, pp. 58-68.
[6] Xiaojun Qi, and Ji Qi, (2007), “A robust content-based digital image watermarking scheme”, Signal
Processing, Vol. 87, pp. 1264–1280.
[7] J. M. Shapiro, (1993) “Embedded image coding using zerotrees of wavelet coefficients,” IEEE Trans.
Signal Processing, Vol. 41, No. 12, pp. 3445-3462.
[8] Manabu Shinohara and Fumiaki Motoyoshi, (2007), “Wavelet-Based Robust Digital Watermarking
Considering Human Visual System”, Proc. 2007 WSEAS International Conference on Computer
Engineering and Applications, Gold Coast, Australia, pp. 177-180.
[9] R. Dugad, K. Ratakonda and N. Ahuja, (1998) “A new wavelet-based scheme for watermarking image,”
Proc. Int. Conf. on Image Processing, pp. 419-423.
[10] H. Inoue, A. Miyazaki, A. Yamamoto and T. Katsura, (1998), “A digital watermark base on the wavelet
transform and its robustness in image compression,” Proc. IEEE Int. Conf. Image Processing, 2, , pp.
391-395.
[11] R. Dugad, K. Ratakonda and N. Ahuja, (1998), “A new wavelet-based scheme for watermarking image,”
Proc. Int. Conf. on Image Processing, 2, pp. 419-423.
[12] X.-M. Niu, Z.-M. Lu and S.-H. Sun, (2000) “Digital watermark of still image with gray-level digital
watermarks,” IEEE Trans. Consumer Electronics, Vol. 46, no.1, pp. 137-144.
[13] M.-S. Hsieh, D.-C. Tseng, and Y.-H. Huang, (2001) “Hidden digital watermarks using multiresolution
wavelet transform,” IEEE Trans. Industrial Electronics, Vol. 48, No. 5, pp.875-882.
[14] A. Munteanu, J. Cornelis, G. V. d. Auwera, and P. Cristea, (1999) “Wavelet image compression-the
quadree coding approach,” IEEE Trans. Information Technology in Biomedicine, Vol. 3, No. 3, pp.
176-185.
[15] D. A. Karras, S. A. Karkanis, and B. G. Mertzios, (1998) “Image compression using the wavelet transform
on textural regions of interest,” in Proc. IEEE Int. Conf. Euromicro, Vol. 2, pp. 633-939.
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010
57
[16] X. Yang, and P. S. Toh, (1995) “Adaptive fuzzy multilevel median filter,” IEEE Trans. Image Processing,
Vol. 4, No. 5, pp. 680-682.
[17] F. Farbiz, M. B. Menhaj, S. A. Motamedi, and M. T. Hagan, (2000) “A new fuzzy logic filter for image
enhancement,” IEEE Trans. Systems, Man, and Cybernetics, Vol. 30, No. 1, pp. 110-119.
[18] M.-S. Hsieh and D.-C. Tseng, (2003), “Image subband coding using fuzzy inference and adaptive
quantization,” IEEE Trans. Systems, Man, and Cybernetics B: Cybernetics, Vol. 33, No. 3, pp. 509-513.
[19] Nizar Sakr, Nicolas Georganas, and Jiying Zhao, (2006) “Copyright Protection of Image Learning Objects
using Wavelet-based Watermarking and Fuzzy Logic,” Proc. I2LOR 2006 - 3rd annual e-learning
conference on Intelligent Interactive Learning Object Repositories Montreal, Quebec, Canada, 9–11
November.
[20] S. Mallat (1989), , “Multifrequency channel decomposition of images and wavelets models,” IEEE Trans.
Acoustics Speech and Signal Processing, vol. 37, no. 12, 1989, pp. 2091-2110.
[21] G. A. Triantafyllidis and M. G. Strintzis(1999), “A context base adaptive arithmetic coding technique for
lossless image compression,” IEEE Signal Processing Letters, vol. 6, no. 7, pp. 168-170.
[22] Y. Yoo, A. Ortega, and Bin Yu (1999), “Image subband coding using context-base classification and
adaptive quantization,” IEEE Trans. Image Processing, vol. 8, no. 12, pp. 1402-1714.

More Related Content

PDF
Compression technique using dct fractal compression
PDF
11.compression technique using dct fractal compression
PDF
DCT and Simulink Based Realtime Robust Image Watermarking
PDF
A Novel Technique for Image Steganography Based on DWT and Huffman Encoding
PDF
A novel steganographic technique based on lsb dct approach by Mohit Goel
PDF
Maximizing Strength of Digital Watermarks Using Fuzzy Logic
DOCX
Implementation of digital image watermarking techniques using dwt and dwt svd...
PDF
Ijetcas14 527
Compression technique using dct fractal compression
11.compression technique using dct fractal compression
DCT and Simulink Based Realtime Robust Image Watermarking
A Novel Technique for Image Steganography Based on DWT and Huffman Encoding
A novel steganographic technique based on lsb dct approach by Mohit Goel
Maximizing Strength of Digital Watermarks Using Fuzzy Logic
Implementation of digital image watermarking techniques using dwt and dwt svd...
Ijetcas14 527

What's hot (18)

PDF
1674 1677
PDF
call for papers, research paper publishing, where to publish research paper, ...
PDF
SIGNIFICANCE OF RATIONAL 6TH ORDER DISTORTION MODEL IN THE FIELD OF MOBILE’S ...
PDF
A Review of Comparison Techniques of Image Steganography
PDF
IRJET- An Improved Technique for Hiding Secret Image on Colour Images usi...
PDF
A NOVEL APPROACH FOR IMAGE WATERMARKING USING DCT AND JND TECHNIQUES
PDF
Digital Image Watermarking Basics
PDF
I017535359
PDF
A new approach on noise estimation of images
PDF
1820 1824
PDF
A systematic image compression in the combination of linear vector quantisati...
PDF
40120140505005
PDF
40120140505005 2
PDF
Cecimg an ste cryptographic approach for data security in image
PDF
Robust Digital Image-Adaptive Watermarking Using BSS Based
PDF
A NOVEL IMAGE STEGANOGRAPHY APPROACH USING MULTI-LAYERS DCT FEATURES BASED ON...
PDF
A robust combination of dwt and chaotic function for image watermarking
PDF
A novel secure image steganography method based on chaos theory in spatial do...
1674 1677
call for papers, research paper publishing, where to publish research paper, ...
SIGNIFICANCE OF RATIONAL 6TH ORDER DISTORTION MODEL IN THE FIELD OF MOBILE’S ...
A Review of Comparison Techniques of Image Steganography
IRJET- An Improved Technique for Hiding Secret Image on Colour Images usi...
A NOVEL APPROACH FOR IMAGE WATERMARKING USING DCT AND JND TECHNIQUES
Digital Image Watermarking Basics
I017535359
A new approach on noise estimation of images
1820 1824
A systematic image compression in the combination of linear vector quantisati...
40120140505005
40120140505005 2
Cecimg an ste cryptographic approach for data security in image
Robust Digital Image-Adaptive Watermarking Using BSS Based
A NOVEL IMAGE STEGANOGRAPHY APPROACH USING MULTI-LAYERS DCT FEATURES BASED ON...
A robust combination of dwt and chaotic function for image watermarking
A novel secure image steganography method based on chaos theory in spatial do...
Ad

Similar to PERCEPTUAL COPYRIGHT PROTECTION USING MULTIRESOLUTION WAVELET-BASED WATERMARKING AND FUZZY LOGIC (20)

PDF
Wavelet Based Image Watermarking
PDF
Psnr value of digital image watermarking by using
PDF
1674 1677
PDF
A Review of Digital Watermarking Technique for the Copyright Protection of Di...
PDF
A Wavelet Based Hybrid SVD Algorithm for Digital Image Watermarking
PDF
Dual-layer Digital Image Watermarking for Intellectual Property Right Protection
PDF
ROBUST IMAGE WATERMARKING METHOD USING WAVELET TRANSFORM
PDF
Reversible color video watermarking scheme based on hybrid of integer-to-inte...
PDF
A New Technique to Digital Image Watermarking Using DWT for Real Time Applica...
PDF
A New Wavelet based Digital Watermarking Method for Authenticated Mobile Signals
PDF
DIRECTIONAL BASED WATERMARKING SCHEME USING A NOVEL DATA EMBEDDING APPROACH
PDF
SVD Based Robust Digital Watermarking For Still Images Using Wavelet Transform
PDF
K42016368
PDF
[IJET V2I4P2] Authors:Damanbir Singh, Guneet Kaur
PDF
Feature Based watermarking algorithm for Image Authentication using D4 Wavele...
PDF
Gh2411361141
PDF
Towards Optimal Copyright Protection Using Neural Networks Based Digital Imag...
PDF
A Hybrid Digital Watermarking Approach Using Wavelets and LSB
PDF
Novel DCT based watermarking scheme for digital images
Wavelet Based Image Watermarking
Psnr value of digital image watermarking by using
1674 1677
A Review of Digital Watermarking Technique for the Copyright Protection of Di...
A Wavelet Based Hybrid SVD Algorithm for Digital Image Watermarking
Dual-layer Digital Image Watermarking for Intellectual Property Right Protection
ROBUST IMAGE WATERMARKING METHOD USING WAVELET TRANSFORM
Reversible color video watermarking scheme based on hybrid of integer-to-inte...
A New Technique to Digital Image Watermarking Using DWT for Real Time Applica...
A New Wavelet based Digital Watermarking Method for Authenticated Mobile Signals
DIRECTIONAL BASED WATERMARKING SCHEME USING A NOVEL DATA EMBEDDING APPROACH
SVD Based Robust Digital Watermarking For Still Images Using Wavelet Transform
K42016368
[IJET V2I4P2] Authors:Damanbir Singh, Guneet Kaur
Feature Based watermarking algorithm for Image Authentication using D4 Wavele...
Gh2411361141
Towards Optimal Copyright Protection Using Neural Networks Based Digital Imag...
A Hybrid Digital Watermarking Approach Using Wavelets and LSB
Novel DCT based watermarking scheme for digital images
Ad

More from gerogepatton (20)

PDF
International Journal of Artificial Intelligence & Applications (IJAIA)
PDF
Performance Evaluation of Block-Sized Algorithms for Majority Vote in Facial ...
PDF
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
PDF
International Journal of Artificial Intelligence & Applications (IJAIA)
PDF
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
PDF
Augmented and Synthetic Data in Artificial Intelligence
PDF
3rd International Conference on AI, Data Mining and Data Science (AIDD 2025)
PDF
July 2025 - Top 10 Read Articles in Artificial Intelligence and Applications ...
PDF
6th International Conference on Natural Language Processing and Computational...
PDF
From Insight to Impact: The Evolution of Data-Driven Decision Making in the A...
PDF
6th International Conference on Artificial Intelligence and Machine Learning ...
PDF
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
PDF
International Journal of Artificial Intelligence & Applications (IJAIA)
PDF
AI-Driven Vulnerability Analysis in Smart Contracts: Trends, Challenges and F...
PDF
International Journal of Artificial Intelligence & Applications (IJAIA)
PDF
6th International Conference on Artificial Intelligence and Machine Learning ...
PDF
A Thorough Introduction to Multimodal Machine Translation
PDF
International Journal of Artificial Intelligence & Applications (IJAIA)
PDF
6th International Conference on Advanced Machine Learning (AMLA 2025)
PDF
OWE-CVD: An Optimized Weighted Ensemble for Heart Disease Prediction
International Journal of Artificial Intelligence & Applications (IJAIA)
Performance Evaluation of Block-Sized Algorithms for Majority Vote in Facial ...
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
International Journal of Artificial Intelligence & Applications (IJAIA)
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
Augmented and Synthetic Data in Artificial Intelligence
3rd International Conference on AI, Data Mining and Data Science (AIDD 2025)
July 2025 - Top 10 Read Articles in Artificial Intelligence and Applications ...
6th International Conference on Natural Language Processing and Computational...
From Insight to Impact: The Evolution of Data-Driven Decision Making in the A...
6th International Conference on Artificial Intelligence and Machine Learning ...
3rd International Conference on Artificial Intelligence and IoT (AIIoT 2025)
International Journal of Artificial Intelligence & Applications (IJAIA)
AI-Driven Vulnerability Analysis in Smart Contracts: Trends, Challenges and F...
International Journal of Artificial Intelligence & Applications (IJAIA)
6th International Conference on Artificial Intelligence and Machine Learning ...
A Thorough Introduction to Multimodal Machine Translation
International Journal of Artificial Intelligence & Applications (IJAIA)
6th International Conference on Advanced Machine Learning (AMLA 2025)
OWE-CVD: An Optimized Weighted Ensemble for Heart Disease Prediction

Recently uploaded (20)

PPTX
UNIT 4 Total Quality Management .pptx
PPTX
Recipes for Real Time Voice AI WebRTC, SLMs and Open Source Software.pptx
PPTX
Welding lecture in detail for understanding
PPTX
CYBER-CRIMES AND SECURITY A guide to understanding
PPTX
Construction Project Organization Group 2.pptx
PDF
Mohammad Mahdi Farshadian CV - Prospective PhD Student 2026
PPTX
web development for engineering and engineering
PPTX
Infosys Presentation by1.Riyan Bagwan 2.Samadhan Naiknavare 3.Gaurav Shinde 4...
PDF
Well-logging-methods_new................
PDF
R24 SURVEYING LAB MANUAL for civil enggi
PPTX
MET 305 2019 SCHEME MODULE 2 COMPLETE.pptx
PDF
Model Code of Practice - Construction Work - 21102022 .pdf
PDF
Enhancing Cyber Defense Against Zero-Day Attacks using Ensemble Neural Networks
PDF
Automation-in-Manufacturing-Chapter-Introduction.pdf
PDF
composite construction of structures.pdf
PPTX
CH1 Production IntroductoryConcepts.pptx
PPTX
IOT PPTs Week 10 Lecture Material.pptx of NPTEL Smart Cities contd
PDF
PPT on Performance Review to get promotions
PPTX
UNIT-1 - COAL BASED THERMAL POWER PLANTS
PPTX
Engineering Ethics, Safety and Environment [Autosaved] (1).pptx
UNIT 4 Total Quality Management .pptx
Recipes for Real Time Voice AI WebRTC, SLMs and Open Source Software.pptx
Welding lecture in detail for understanding
CYBER-CRIMES AND SECURITY A guide to understanding
Construction Project Organization Group 2.pptx
Mohammad Mahdi Farshadian CV - Prospective PhD Student 2026
web development for engineering and engineering
Infosys Presentation by1.Riyan Bagwan 2.Samadhan Naiknavare 3.Gaurav Shinde 4...
Well-logging-methods_new................
R24 SURVEYING LAB MANUAL for civil enggi
MET 305 2019 SCHEME MODULE 2 COMPLETE.pptx
Model Code of Practice - Construction Work - 21102022 .pdf
Enhancing Cyber Defense Against Zero-Day Attacks using Ensemble Neural Networks
Automation-in-Manufacturing-Chapter-Introduction.pdf
composite construction of structures.pdf
CH1 Production IntroductoryConcepts.pptx
IOT PPTs Week 10 Lecture Material.pptx of NPTEL Smart Cities contd
PPT on Performance Review to get promotions
UNIT-1 - COAL BASED THERMAL POWER PLANTS
Engineering Ethics, Safety and Environment [Autosaved] (1).pptx

PERCEPTUAL COPYRIGHT PROTECTION USING MULTIRESOLUTION WAVELET-BASED WATERMARKING AND FUZZY LOGIC

  • 1. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 DOI : 10.5121/ijaia.2010.1304 45 PERCEPTUAL COPYRIGHT PROTECTION USING MULTIRESOLUTION WAVELET-BASED WATERMARKING AND FUZZY LOGIC Ming-Shing Hsieh Department of Computer Engineering and Information Science, Aletheia University, Damshui, Taiwan 251 sms.sms@msa.hinet.net ABSTRACT In this paper, an efficiently DWT-based watermarking technique is proposed to embed signatures in images to attest the owner identification and discourage the unauthorized copying. This paper deals with a fuzzy inference filter to choose the larger entropy of coefficients to embed watermarks. Unlike most previous watermarking frameworks which embedded watermarks in the larger coefficients of inner coarser subbands, the proposed technique is based on utilizing a context model and fuzzy inference filter by embedding watermarks in the larger-entropy coefficients of coarser DWT subbands. The proposed approaches allow us to embed adaptive casting degree of watermarks for transparency and robustness to the general image-processing attacks such as smoothing, sharpening, and JPEG compression. The approach has no need the original host image to extract watermarks. Our schemes have been shown to provide very good results in both image transparency and robustness. KEYWORDS digital watermarking, discrete wavelet transform, fuzzy inference filter, adaptive quantization, entropy. 1. INTRODUCTION The increasingly easy access to digital media and increasingly powerful tools available for manipulating digital media have made media security a very important issues. Due to the open environment of Internet downloading, copyright protection introduces a new set of challenging problems regarding security and illegal distribution of privately owned images. One potential solution for declaring the ownership of the images is to use watermarks to embed an invisible signal into multimedia data so as to prove the owner identification of the data and discourage the unauthorized copying. In the study, we focus our research on digital image watermarking, but the method could be modified for other applications as well. In general, there are two common requirements of watermark. First requirement is the watermarks must be perceptually transparency. They should not noticeable to the viewer. The second requirement is the watermarks must be robust to intentional or unintentional attacks and common signal processing. Particularly, the watermark should still be detectable even after attacks have been applied to the watermarked image. These attacks include image compression, linear or nonlinear filtering, image enhancements, etc. The watermarking techniques can be divided into two different classifications. One is applied to spatial domain, and the other is applied to the frequency domain. The spatial domain watermark techniques are developed early [1, 2], simple but not robust is their obvious weakness. They can’t against intentional or unintentional attacks and image processing. The embedded watermark signals are easily disfigured, distorted, or removed. The frequency domain approach has some advantages because most of the signal processing operations can be well characterized, and many good perceptual
  • 2. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 46 models are developed in the frequency domain [3]. Cox et al. [4] proposed an image watermarking method based on spread spectrum theory, which shows good performance among invisibility and robustness to signal processing operation and common geometric transform. Hsu et al. [5] proposed DCT based image watermarking technique by embedding a visually recognizable watermark. An image-content-based adaptive embedding scheme is applied in discrete Fourier transform (DFT) domain of each perceptually high textured subimage to ensure better visual quality and more robustness [6]. On the other hand, several methods used the discrete wavelet transform (DWT) to hide data to the frequency domain to provide extra robustness against attacks. DWT is known that the wavelet image/video coding, such as embedded zero-tree wavelet (EZW) coding [7], and play an important role in image/video compression standards, such as JPEG2000 and MPEG4 due to its excellent performance in compression. In [8], they propose watermarking schemes where visual models are used to determine image dependent upper bounds on watermark insertion. This allows users to provide the maximum strength transparent watermark, which is extremely robust to common image processing and editing such as image compression. Dugad et al. [9] picked all coefficients in the subbands, which are larger than a given threshold, and watermark is added to these coefficients only. The reason to set the threshold is to embed the watermarks in the edge regions of the host image, so that the watermarked image is not distinguishable from the original image. Hsu et al. [3] used DWT other than DCT and the same method proposed in [5] to embed signal into each wavelet subbands. The empirical results showed that the watermarked image was more robustness and imperceptibly then they early did in [3]. Inoue et al. [10] proposed a method based on zerotree [11], which classified wavelet coefficients as insignificant or significant using zerotree, and select proper position to insert watermarks. Embedding watermarks into host image using DWT are both in transparency and robustness. In [4] the watermark is a symbol or a random number, which comprises of a sequence of bits, and can only be detected by employing detection theory. In [12], [13], the watermark is a visually recognizable pattern. This kind of watermark is more intuitive for representing one’s identity than a random sequence, and also can be measured the correlation between the detected visual pattern and original watermark during the verification phase. Originally, the determination of a wavelet coefficient to be an embedded target is done by comparing subband’s coefficients uniformly with a fixed sorting order. If all coefficients of a wavelet subband are less than the threshold value, they are set “insignificant” and had no chance to embed watermarks. Actually, the determination is uncertain and mutual influence. For example, all coefficients of a subband being slightly less than the threshold value should be not always set “insignificant”; on the other hand, all coefficients of a subband being slightly greater than the threshold value should be not always set “significant”. Therefore, a fuzzy filter is necessary to provide reasoning with the vague and uncertain information. The use of fuzzy filters in image processing is based on the idea that pixels or coefficients are not uniformly fired by every fuzzy rule. The fuzzy set theory has the potential capability to efficiently represent input and output relationships of dynamic systems, thus it has gained popularity. For example, the usage of fuzzy algorithms for vector quantization has been proposed by Munteanu et al. [14]; fuzzy clustering was employed to preserve the textually important image characteristics while a compression algorithm was proposed by Karras et al. [15]; Yang and Toh [16] applied heuristic fuzzy rules to improve the performance of the traditional multilevel filter; Farbiz et al. [17] applied a new fuzzy logic filter to improve the performance of image enhancement. Hsieh et al. [18] applied a new fuzzy logic filter to improve the performance of image coding. In [19], an adaptive watermarking algorithm is presented which exploits a biorthogonal wavelets-based human visual system (HVS) and a Fuzzy Inference System (FIS) to protect the copyright of images in learning object repositories. The FIS isutilized to compute the optimum watermark weighting function thatwould enable the embedding of the maximum-energy and imperceptiblewatermark. In this paper, we propose a wavelet-based watermarking approach by adding visually recognizable image to the larger entropies of coefficients calculated by the fuzzy inference filter of selected wavelet subband. The proposed approach has the following advantages: (i) the extracted watermark is visually
  • 3. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 47 recognizable to claim one’s ownership, (ii) the approach is hierarchical and has multiresolution characteristics, (iii) the embedded watermark is hard detected by human visual perceptivity. Our experimental results show that the proposed watermarking approach is very robust to image compression and image operations. This paper is organized as follows. Wavelet transform of images, context-based method and fuzzy filter is described in Section 2. Section 3 describes the watermark embedding approach. In Section 4, the experimental results are shown. The conclusion of this paper is stated in Section 5. 2. PRELIMINARIES 2.1. Wavelet transform of images The wavelet transform is identical to a hierarchical subband system, where the subbands are logarithmically spaced in frequency. For an input sequence of length N, DWT will generate an output sequence of length N. The 1-D DWT can be implemented by the Direct Pyramid Algorithm, which was developed by Mallat [20] as follows. Begin {Direct Pyramid Algorithm} For k = 1 to K For n = 1 to 2K-k Y[k, n] = ∑ − = −− 1 0 21 N m ngmnkX )(],[ X[k, n] = ∑ − = −− 1 0 21 N m nhmnkX )(],[ End where k is the current octave, K represents the total number of octaves, 1 ≤ k ≤ K, n is the current input, and N is the total number of inputs, 1 ≤ n ≤ N. Mallat’s direct pyramid algorithm is executed by down sampling. For a 2-D image, a wavelet Ψ and a scaling function Φ are chosen such that the scaling function ΦLL(x, y) of low-low subband in a 2-D wavelet transform can be written as ΦLL(x, y) = Φ(x) Φ(y). Three other bi-dimensional wavelets can also be obtained using the wavelet associated function Ψ(x) as follows. ΨLH(x, y) = Φ(x) Ψ(y) ; horizontal ΨHL(x, y) = Ψ(x) Φ(y) ; vertical ΨHH(x, y) = Ψ(x) Ψ(y) ; diagonal where H is a high-pass filter and L is a low-pass filter. The basic idea in the DWT of a 2-D image is as follows. An image is firstly decomposed into four parts of high, middle, and low frequencies (i.e., LL1, HL1, LH1, HH1) subbands, by cascading horizontal and vertical two-channel critically subsampled filter banks. The subbands labeled HL1, LH1, and HH1 represent the finest scale wavelet coefficients. To obtain the next coarser scale of wavelet coefficients, the subband LL1 is further decomposed and critically subsampled. This process is continued an arbitrary number of times, which is determined by the application at hand. Fig. 1 shows layout of the image subbands from three-level dyadic decomposition and an example of DWT decomposition of the Lena image using a wavelet filter set. In the figure, Lena image is decomposed into ten subbands for three scales. Each level has various band-information such as low-low, low-high, high-low, and high-high frequency bands. Furthermore, from these DWT coefficients, the original image can be reconstructed. The reconstruction process is called the inverse DWT (IDWT). Let I [m, n] represent an image. The DWT and IDWT for I [m, n] can be similarly defined by implementing the
  • 4. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 48 DWT and IDWT for each dimension m and n separately: DWTn [DWTm I [m, n]]. The wavelet transform is identical to a hierarchical subband system, where the subbands are logarithmically spaced in frequency and represent an octaveband decomposition. An example of three scales wavelet transform is show in Fig. 1. The image is first decomposed into for subband LL1, LH1, HL1, HH1 by using horizontal and vertical two channel critically subsampled filter banks. Each coefficient represents a spatial area corresponding to approximately a 2×2 area of the original image. The LH1, HL1, HH1 represent the finest scale wavelet coefficient. After decomposed and critically subsampled the subband LL1, we can obtain the next coarser scale of wavelet coefficients. Note that for each coarser scale, the coefficients represent a larger spatial area of the image but a narrower band of frequencies. The decomposed process continues until some final scale is reach. LL3 HL3 LH3HH3 HL2 LH2 HH2 HH1LH1 HL1 (a) (b) Fig. 1. (a) Layout of the image subbands from three-level dyadic decomposition. (b) An example of DWT decomposition of the Lena image. 2.2. Context-based method The main function of the context-based method is to find the relationship of one pixel/coefficient and its adjacent pixels/coefficients in the image. The relationship among the context is to determine characteristics of the target pixel/coefficient. Context-based method can be applied in image compression, which is based on wavelet transform. Triantafyllidis et al. [21] proposed that current wavelet coefficient could be estimated using the coefficients that lie in the current band, scale bound(s), parent band, and the pyramid structure. Yoo et al. [21] introduce a context-based classification technique, which classifies each subband coefficient based on the surrounding coefficients, and different quantizer is then used for each class. In our approaches, we introduce context-based method on watermark embedding. Unlike previous works that only use the larger coefficients to insert watermarks, we use the current coefficient and its surrounding coefficients to calculate the entropy of each coefficient in selected subband, and choose the larger entropy to embed watermarks. If a coefficient has higher entropy denotes the violent variation of spatial domain in the host image, to embed watermarks in the center coefficient of the context could improve the transparency of the watermarked image and robustness to extracted watermarks. 2.3. Using fuzzy filter to calculate entropy of wavelet coefficient In this section, the use of context and fuzzy filter to calculate the entropy corresponding to each coefficient in one subband are described. General block diagram of a fuzzy inference system is shown in Fig. 2.
  • 5. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 49 Fig. 2. General block diagram of a fuzzy inference system (FIS). In an arbitrary wavelet subband S, where we determine entropy of each coefficient in S is based on current coefficient and its surrounding coefficients. Template made of nine coefficients form the context is shown in Fig. 3. Let x0 be the current target coefficient to estimate its entropy, xi, 1≤i≤8, is a x0’s surrounding coefficient as shown in Fig. 3. Fuzzy rules shown in Table 1 are used as inference filter to calculate the entropy corresponding to each coefficient in S: x 1 x 2 x 3 x 4 x0 x 5 x 6 x 7 x 8 Fig. 3. Template made of nine coefficients form the context. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Ratio Degreeofmembership Very Low Low Medium High Very High Fig. 4. Membership functions depicting the degrees of normalized fuzzy wavelet coefficients. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fig. 5. Typical membership function for the fuzzy function “more”. Fuzzifer Fuzzy Inference Engine DeFuzziferFuzzy Rule Base Crisp input x, y Crisp ouput z µ(x, y) µ(z) Fuzzy Fuzzy
  • 6. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 50 Table 1. Fuzzy Inference rules Rule Rule Description R1 R2 R3 R4 R5 R1:If more of xi are Very High Then En(xi) is Very High. R2:If more of xi are High Then En(xi) is High. R3:If more of xi are Medium Then En(xi) is Medium. R4:If more of xi are Low Then En(xi) is Low. R5:If more of xi are Very Low Then En(xi) is Very Low. Note that in the above production rules in Table 1, xi’s, named normalized fuzzy coefficients (NFCs), are the normalized fuzzy value (NFV) of wavelet coefficients. NFV of each coefficient is normalized by dividing some of T to emphasis the importance of coefficients in its context. The predicate’s and conclusion’s part, ‘Very High’, ‘High’, ‘Medium’, ‘Low’, and ‘Very Low’ are the degrees of fuzzy membership as illustrated in Fig. 4. The output En(xi) denotes the entropy of xi, is given by the target coefficient and its surrounding coefficients which yield the result by Equ (3). The term more denotes an S-type fuzzy function whose typical shape is shown in Fig. 5. The curve of this function enables the non-uniform firing of the basic fuzzy rules. This more function may be described by the following formula: )( 1 1 )( βα µ −− + = zmore e z (1) where α=7.80375, β=3.29596. The curve of the function enables the non-uniform firing of the basic fuzzy rules. NFV of each coefficient is defined as follows. 0 0 0 T x x = , 1 7,5,4,2| T x x n nn == , 2 8,6,3,1| T x x m mm == , where T0 = Avg(|S|), T1= T0+1/16×Std(|S|), T2= T0+1/8×Std(|S|), Avg, Std denote the average, stand deviation, respectively. The activity degree of rule Rk is computed by the following relationship: ×∈= )}sup(:)(min{ qualityxx iiqualityk µλ       ∈ i ii more xofnumbertotal qualityxwhichxofnumber )sup( µ (2) where 1≤ k ≤ 5, and the corresponding quality ∈ {Very High | High | Medium | Low}. After all λk’s are calculated, the entropy of coefficient xi is computed by equation: ∑= = 5 1 )( k kki CxEn λ (3) where Ck represents the center point of the membership function Rk. 3. WATERMARKING IN THE DWT DOMAIN The proposed digital watermarking approach can hide visually recognizable patterns in images. The main study task of digital watermarking is to make watermarks invisible to human eyes as well as
  • 7. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 51 robust to various attacks. In the proposed approach, watermarks are embedded in the host image by modifying the coefficients with larger weighted energy. The block diagram of embedding watermarks in the host image is shown in Fig. 6. Host image DWT Select embedded subband Calculate entropy coef.'s Select embedded coef. Sorting/ Quantization Embedded position Sorting map Calculate weight T0-T2 Embedding Watermark Sorting Sorting map parameter v IDWT Key Watermarked image Calculate casting weight v Fig. 6. The block diagram of embedding watermarks. 3.1. Watermark embedding method The algorithm for embedding a watermark in a host image is described as follows: Step 1: Sort the gray levels of watermark W in ascending order to generate the sorted watermark SW, Assume that the size of watermark image W is n. Step 2: Decompose a host image H into three levels with ten subbands of a wavelet pyramid structure. Choose a subband S, for example HL3, to embed watermark W. Step 3: Calculate the weighted entropy En of coefficients in subband HL3 using the proposed method approach described in Section II.B. Step 4: Let the preset interval be τ, t be the number of referenced coefficients used as a key to extract watermark W without the host image. p coefficients with larger entropy are chosen from subband S, where p = n + t. The larger entropy of coefficients makes the watermarked image more robustness and transparency. Obviously, if n/τ=0, the value of t is fix(n/τ) + 1, otherwise t is fix(n/τ) + 2, where function fix is to get the integer part of its argument. Let {Cp} be the set of referenced coefficients and the coefficients to be embedded watermarks; {Cp} is called the alternative coefficients. Sorting {Cp} to generate {SCp} called the sorted alternative coefficients. Step 4: Quantize {SCp} using a preset interval could extract watermark W without the host image. Step 6: Embed watermark SW into subband HL3 by the watermark embedding strategy, SCj = SCj + vi×SWindex, where SCj is a sorted alternative coefficients except the referenced coefficients, vi = Eni × (T0+T1+T2)/3 = Eni × T1. The value of scaling factor vi is the maximum variances to modify the embedded coefficients for robustness while embedding watermarks to the target coefficients and SWindex is a sorted waremark. A coefficient with larger weighed entropy could embed watermark with larger scaling factor for robustness without obviously degrading the host image. The embedding procedure can be implemented by the proposed Casting watermarks Algorithm as follows. Begin {Casting Watermarks Algorithm}
  • 8. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 52 index=1 /* Embedded watermarks for each intervals with τ coefficients. */ For i = 1 to p-τ step τ+1 For j = i+1 to i+τ SCj = SCj + vj×SWindex index = index + 1 EndFor EndFor /* Embedded watermarks for the rest alternative coefficients. */ If (n mod τ) ≠ 0 For j = p-(n mod τ) to p-1 SCj = SCj + vj×SWindex index = index + 1 EndFor EndIf End {Casting Watermarks Algorithm} Step 7: Save scaling factor vi, the symbol of embedded subband, symbol table of SCi, corresponsive map of Ci and SCi, and corresponsive map of Wi and SWi. To form watermarked image, one can take the IDWT of the modified DWT coefficients and the unchanged DWT coefficients. 3.2. Watermark extracting method The embedded watermark can be detected based on the stored parameters after the wavelet decomposition of the watermarked image as follows: Step 1: Decompose a watermarked image into three levels with ten subbands using DWT. Step 2: Restore the scaling factor vi, the symbol of embedded subband, symbol map of SCi, corresponsive map of Ci and SCi, and corresponsive map of Wi and SWi. Step 3: Extract the sorted watermarks by the proposed Extracting Watermarks Algorithm as follows: Begin {Extracting Watermarks Algorithm} index=1 /* Extracting watermarks for each intervals with τ coefficients. */ For i = 1 to p-τ step (τ+1) For j = i+1 to i+τ value = (|SCi|+ |SCi+τ+1|) / 2 SWindex = (|SCj|– value) / vj index = index + 1 EndFor EndFor /* Extracting watermarks for the rest alternative coefficients. */ If (n mod τ) ≠ 0
  • 9. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 53 For j = p-(n mod τ) to p-1 value=(|SCp-(n mod τ)-1+ |SCx’p|) / 2 SWindex = (SCx’j – value) / vj index = index + 1 EndFor EndIf End {Extracting Watermarks Algorithm} Step 4: Rearranging watermarks from corresponsive map of Wi and SWi to get the extracted watermarks W’. In our scheme, the extracted watermark W’ is a visually recognizable image. A subjective measurement based on the standard correlation coefficient, ∑∑ ∑ −− −− = 22 )()''( ))(''( WWWW WWWW ncorrelatio (4) is used to evaluate the quality of the extracted watermarks by measuring the similarity of the original watermark W and extracted watermark W’. The value of correlation is between zero and one. A larger value of correlation represents more similarity of the original watermarks and extracted watermarks. 4. EXPERIMENTS A 512×512 Lena image was taken as the host image to embed a 32×32 binary watermark image with “NCU CSIE” characters following a sequence of attacks. The proposed approach emphasizes the local characteristics in the experiments, we also examined the quality of watermarked images and detectability of watermarks. 4.1. Image quality The proposed perceptual watermarking framework was implemented for evaluating both properties of transparency and robustness. The two conflicting characteristics are referred to the embedded positions and casting degree of watermarks. For the embedded positions, as we chose the larger entropy of coefficients to embedded watermarks, which stands for the violent variance of spatial domain in a host image, the transparency and robustness of host image could be hewn. For the casting strength, the larger casting degree makes it more robust to the watermarked image. The side effect is that the lower transparence makes the watermarks the perceptual break. In the experiments, an adaptive casting strength was proposed to adjust casting degree of the coefficients, which makes the balance of transparency and robustness. Fig. 7 shows an example of embedding results, where Lena is used as the test image, a binary image with “NCU CSIE” is used as the watermark. The performance of the proposed image watermarking approach was evaluated. The common-used 512×512 gray-scale Lena image was used as original image and the peak signal-to-noise ratio (PSNR) was used for objective comparison, MSE PSNR 2 10 255 10log= , (5) where MSE is the mean square error between a watermarked image and its original image. Fig. 7 also shows the watermarked image, in the figure, the PSNR of the watermarked image is 45.68, correlations of the original watermarks and extracted watermarks is 1. Table 2 shows an example of extracting results from Fig. 7(b) without any attacks using the proposed method.
  • 10. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 54 (a) (b) (c) Fig. 7. Example of digital watermarking using visual watermarks, where (a) is the host Lena image, (b) is the watermarked image with PSNR= 45.68 after embedding, (c) the watermarks. Table 2. The extracted watermarks from Fig. 7 are visually recognizable Embedded watermarks PSNR (dB) after embedding watermarks 45.68 Extracted watermarks Correlation 1 Error bits 0 4.2. On the robustness against JPEG lossy compression Table 3 shows the extracted results from JPEG compressed version of the watermarked images with different compression quality. The quality of watermarked images is still in good situation even under the high compressed ratio. The extracted watermarks and the original watermarks are with high correlation. In the proposed method, error rate of the extracted watermarks is rarely small (<0.5%) even under the situation of compression ratio, 16.59. From the experimental results, we can find that the proposed watermarking techniques yield satisfactory results in terms of transparency for watermarked images.
  • 11. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 55 Table 3. Changes of correlation and error rate values of the proposed method under JPEG lossy compression. Attack Extracted Watermark Correlation Error rate Quality=100, CR=1.6810 1 0 Quality=90, CR=4.6681 1 0 Quality=80, CR=7.1466 1 0 Quality=70, CR=9.1314 1 0 Quality=60, CR=10.9222 1 0 Quality=50, CR=12.4298 0.9972 0.09% Quality=40, CR=14.1922 0.9972 0.09% Quality=30, CR=16.5903 0.9864 0.49% Quality=20, CR=20.5619 0.8022 7.8% *CR=compression ratio 4.2. On the robustness against JPEG lossy compression Sharpen operations are used to enhance the subjective quality. Table 4 shows the extracted results of applying enhanced operation to a watermarked image. The extracted results are highly similar to the original watermark. Smoothing operations such as median filter are used to decrease spurious effects that may be present in images from a poor transmission channel. Table 4 shows the extracted results of applying median filter to a watermarked image. The extracted watermark is still visually recognizable. Table 4. The extracted of watermarks, correlation and error rate under sharpen and smoothing operation. Attack Extracted Watermark Correlation Error rate Sharpen 0.9945 0.19% Median filter 0.9488 1.8% 5. CONCLUSION We have introduced a watermarking framework for embedding visually recognizable watermarks in
  • 12. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 56 images, which can resist image-processing attacks, such as JPEG compression, smoothing, sharpening, blur, etc. The proposed new techniques are based on the context ideas and fuzzy filter. In the experiments, the fuzzy filter is employed to conclude the entropy of each coefficient in the center of the context. The watermark embedding strategies of the proposed methods consider the local characteristics by choosing the larger-entropy coefficients of DWT subbands to embed watermarks. The experimental results show that the proposed methods provide extra robustness against JPEG-compression and image processing compared to the traditional embedding methods. Moreover, the embedded coefficients in the proposed approaches have their own embedding degrees, which are calculated automatically in accordance with the entropy contributed from the context. Finally, the proposed approaches have no need the original host image to extract watermarks. ACKNOWLEDGEMENTS The authors would like to thank Mr. Y.-H. Huang, Prof Tseng D-C, and the anonymous reviewers of this paper. REFERENCES [1] R. G. van Schyndel, A. Z. Tirkel, and C. F. Osborne, (1994) “A digital watermark,” Proc. of IEEE Int. Conf. Image Processing, Vol. 2, pp. 86-90. [2] R. Wolfgang and E. Delp, (1996) “A watermark for digital image,” Proc. Int. Conf. on Image Processing, (3), pp. 211-214. [3] C.-T. Hsu and J.-L. Wu, (1998) “Multiresolution watermarking for digital images,” IEEE Trans. Consumer Electronics, Vol. 45, No. 8, pp. 1097-1101. [4] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, (1997) “Secure spread spectrum watermarking for multimedia,” IEEE Trans. Image Processing, Vol. 6, No. 12, pp. 1673-1687. [5] C.-T. Hsu and J.-L. Wu, (1999) “Hidden digital watermarks in images, ” IEEE Trans. Image Processing, Vol. 8, No. 1, pp. 58-68. [6] Xiaojun Qi, and Ji Qi, (2007), “A robust content-based digital image watermarking scheme”, Signal Processing, Vol. 87, pp. 1264–1280. [7] J. M. Shapiro, (1993) “Embedded image coding using zerotrees of wavelet coefficients,” IEEE Trans. Signal Processing, Vol. 41, No. 12, pp. 3445-3462. [8] Manabu Shinohara and Fumiaki Motoyoshi, (2007), “Wavelet-Based Robust Digital Watermarking Considering Human Visual System”, Proc. 2007 WSEAS International Conference on Computer Engineering and Applications, Gold Coast, Australia, pp. 177-180. [9] R. Dugad, K. Ratakonda and N. Ahuja, (1998) “A new wavelet-based scheme for watermarking image,” Proc. Int. Conf. on Image Processing, pp. 419-423. [10] H. Inoue, A. Miyazaki, A. Yamamoto and T. Katsura, (1998), “A digital watermark base on the wavelet transform and its robustness in image compression,” Proc. IEEE Int. Conf. Image Processing, 2, , pp. 391-395. [11] R. Dugad, K. Ratakonda and N. Ahuja, (1998), “A new wavelet-based scheme for watermarking image,” Proc. Int. Conf. on Image Processing, 2, pp. 419-423. [12] X.-M. Niu, Z.-M. Lu and S.-H. Sun, (2000) “Digital watermark of still image with gray-level digital watermarks,” IEEE Trans. Consumer Electronics, Vol. 46, no.1, pp. 137-144. [13] M.-S. Hsieh, D.-C. Tseng, and Y.-H. Huang, (2001) “Hidden digital watermarks using multiresolution wavelet transform,” IEEE Trans. Industrial Electronics, Vol. 48, No. 5, pp.875-882. [14] A. Munteanu, J. Cornelis, G. V. d. Auwera, and P. Cristea, (1999) “Wavelet image compression-the quadree coding approach,” IEEE Trans. Information Technology in Biomedicine, Vol. 3, No. 3, pp. 176-185. [15] D. A. Karras, S. A. Karkanis, and B. G. Mertzios, (1998) “Image compression using the wavelet transform on textural regions of interest,” in Proc. IEEE Int. Conf. Euromicro, Vol. 2, pp. 633-939.
  • 13. International Journal of Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010 57 [16] X. Yang, and P. S. Toh, (1995) “Adaptive fuzzy multilevel median filter,” IEEE Trans. Image Processing, Vol. 4, No. 5, pp. 680-682. [17] F. Farbiz, M. B. Menhaj, S. A. Motamedi, and M. T. Hagan, (2000) “A new fuzzy logic filter for image enhancement,” IEEE Trans. Systems, Man, and Cybernetics, Vol. 30, No. 1, pp. 110-119. [18] M.-S. Hsieh and D.-C. Tseng, (2003), “Image subband coding using fuzzy inference and adaptive quantization,” IEEE Trans. Systems, Man, and Cybernetics B: Cybernetics, Vol. 33, No. 3, pp. 509-513. [19] Nizar Sakr, Nicolas Georganas, and Jiying Zhao, (2006) “Copyright Protection of Image Learning Objects using Wavelet-based Watermarking and Fuzzy Logic,” Proc. I2LOR 2006 - 3rd annual e-learning conference on Intelligent Interactive Learning Object Repositories Montreal, Quebec, Canada, 9–11 November. [20] S. Mallat (1989), , “Multifrequency channel decomposition of images and wavelets models,” IEEE Trans. Acoustics Speech and Signal Processing, vol. 37, no. 12, 1989, pp. 2091-2110. [21] G. A. Triantafyllidis and M. G. Strintzis(1999), “A context base adaptive arithmetic coding technique for lossless image compression,” IEEE Signal Processing Letters, vol. 6, no. 7, pp. 168-170. [22] Y. Yoo, A. Ortega, and Bin Yu (1999), “Image subband coding using context-base classification and adaptive quantization,” IEEE Trans. Image Processing, vol. 8, no. 12, pp. 1402-1714.