Académique Documents
Professionnel Documents
Culture Documents
algorithm algorithm based based based based on on on on wavelet wavelet wavelet wavelet transform transform transform transform
Zheng Xiong-bo
School oI Science
Harbin Engineering University
,Harbin 15000, China
zhengxiongbo(hrbeu.edu.cn
Zhang Xiao-wei
School oI Science
Harbin Engineering University
,Harbin 15000, China
zhangxiaowei(hrbeu.edu.cn
Sun Ming-iian
School oI Science
Harbin Engineering University
,Harbin 15000, China
sunmingiian(hrbeu.edu.cn
$EVWUDFW $EVWUDFW $EVWUDFW $EVWUDFW----In In In In recent recent recent recent years, years, years, years, digital digital digital digital watermark watermark watermark watermarking ing ing ing technology technology technology technology is is is is aaaa
frontier frontier frontier frontier research research research research area area area area in in in in academia, academia, academia, academia, which which which which is is is is proposed proposed proposed proposed to to to to deal deal deal deal
with with with with copyright copyright copyright copyright protection protection protection protection problem problem problem problem in in in in a aaa network network network network environment environment environment environment. ...
In In In In the the the the paper, paper, paper, paper, a aaa blind blind blind blind digital digital digital digital watermark watermark watermark watermarking ing ing ing algorithm algorithm algorithm algorithm based based based based on on on on
wavelet wavelet wavelet wavelet transform transform transform transform is is is is proposed. proposed. proposed. proposed. After After After After wavelet wavelet wavelet wavelet transform, transform, transform, transform, we we we we get get get get
the the the the wavelet wavelet wavelet wavelet coefficient coefficient coefficient coefficientssss of of of of aaaa image. image. image. image. According According According According to to to to the the the the Euclid Euclid Euclid Euclid
distance distance distance distance between between between between every every every every two two two two coefficients, coefficients, coefficients, coefficients, all all all all wavelet wavelet wavelet wavelet coefficients coefficients coefficients coefficients
are are are are classified classified classified classified into into into into several several several several classes. classes. classes. classes. Then Then Then Then watermark watermark watermark watermark is is is is
embedded embedded embedded embedded into into into into some some some some fit fit fit fit coefficients. coefficients. coefficients. coefficients. It It It It needn needn needn needn` ```t ttt the the the the original original original original
image image image image in in in in the the the the process process process process of of of of extracting extracting extracting extracting watermark watermark watermark watermarking ing ing ing, ,,, so so so so it it it it is is is is aaaa
blind blind blind blind watermarking watermarking watermarking watermarking algorithm. algorithm. algorithm. algorithm. The The The The experimental experimental experimental experimental results results results results show show show show
that that that that it it it it is is is is a aaa high high high highly ly ly ly robust robust robust robust algorithm algorithm algorithm algorithm against against against against 1PEG 1PEG 1PEG 1PEG compress compress compress compress,,,, the the the the
addition addition addition addition of of of of noise, noise, noise, noise, median median median median filtering filtering filtering filtering and and and and cropping cropping cropping cropping....
.H\ZRUGV .H\ZRUGV .H\ZRUGV .H\ZRUGV GLJLWDO GLJLWDO GLJLWDO GLJLWDO LPDJH LPDJH LPDJH LPDJH ZDWHUPDUN ZDWHUPDUN ZDWHUPDUN ZDWHUPDUN ZDYHOHW ZDYHOHW ZDYHOHW ZDYHOHW WUDQVIRUP WUDQVIRUP WUDQVIRUP WUDQVIRUP ELW ELW ELW ELW
GHFRPSRVLWLRQ GHFRPSRVLWLRQ GHFRPSRVLWLRQ GHFRPSRVLWLRQ
I. INTRODUCTION
In recent years, with the rapid development and widely
application oI network technique and digital multimedia
technology, the requirement oI an eIIective copyright
protection method oI digital production such as digital video,
audio, images becomes more and more pressing|1-2|. As an
eIIective method to protect the security oI digital inIormation,
digital watermark has attained much interest and attention in
recent years. According to the place where the inIormation is
embedded, the method can be classiIied into spatial domain
and transIormation domain algorithm. The disadvantage oI
spatial domain algorithm is that the inIormation can be
embedded into is very little, otherwise the quality oI image
will decline obviously. According to the algorithm on
transIormation domain, such as Discrete Fourier TransIorm
(DFT) |3|, Discrete Cosine TransIorm (DCT) |4-5|, Discrete
Wavelet TransIorm (DWT) |6| and so on, image is
transIormed Iirstly, and then watermark is embedded. Using
these methods, more inIormation can be embedded, and the
security is higher. According to whether the original image
or watermark image is needed, watermark algorithm can be
divided into blind watermark, semi-blind watermark and
non-blind watermark. For wavelet transIormation has good
time-Irequency property, it is widely used in digital
watermark technology. Thus, in the paper, cluster analysis is
used to embed the watermark inIormation into suitable
wavelet coeIIicients in wavelet domain. Because it needn`t
the original image in the process oI watermark extract, so it
is a blind watermarking algorithm. The experimental results
show that it is a highly robust algorithm against JPEG
compression, the addition oI noise, median Iiltering and
cropping.
II. THE CLUSTER OF WAVELET COEFFICIENTS
$ :DYHOHW WUDQVIRUPDWLRQ RI JUD\VFDOH LPDJHV
Due to the linear phase and compact supported
properties oI the biorthogonal wavelet (such as 9/7 wavelet)
|7|, the distortion oI the image that transIormed by
biorthogonal wavelet, and reconstructing signal stability is
also very good. So it is suitable Ior watermark embedding.
Thus, the 9/7 biorthogonal wavelet is used to decompose
gray image oI Lena (256
___________________________________
978-1-4244-8728-8/11/$26.00 2011 IEEE
similarity measurement between the vector oI the L group
and the cluster center
M
4 . In the paper, the clustering
dataset is the ensemble : which contains all the wavelet
coeIIicients oI a image except that oI the smooth part. Here
Euclidean distance is
M L M L
4 Z G =
,
.
Let 0 as the maximum coeIIicient and P as the
minimum coeIIicient in : . II 0 > P , we iust take the
maximum coeIIicient 0 . And it is not necessary to
calculate the part related to P in the Iollowing algorithm,
while the rest process is same to that oI 0 < P .Cluster oI
the coeIIicients in : as Iollowing:
1) Set the initial cluster centers. Let 1 is a positive
integer. Divide | , 0 | 0 and | 0 , |P into 1 equal intervals.
Let
) 1 , 1 ( , , 0 ,
0
= = = =
1 L P
1
L
4 4 0
1
L
4
L L
"
(1)
Here
L L
4 4 4
, ,
0
is the initial clustering centers.
2) ClassiIy the coeIIicients according to the non-
similarity measurement between two groups, each class
correspond to one cluster center. There are 1 2 1 classes.
They are
L L
$ $ $
, ,
0
( 1 , 1 = 1 L " ).
3) Regulate the clustering centers. Let
L L
4 4 =
0
and c
a very small positive integer (in the paper,
2
10 c
= ).
Calculate the average value oI the coeIIicients oI
L
$ and set
itas
L
4 .Let
L
4 as the new cluster center oI
L
$ .II
c <
0
L L
4 4 , then it is end, or turn to the second step. II
it still does not converge aIter several cycles, then iump out.
Through clustering, the wavelet coeIIicients oI a image
except the smooth part are divided into 1 2 1 classes.
III. WATERMARK EMBEDDING
$ :DWHUPDUN JHQHUDWLQJ
The 8-bit and 256 gray images are used as watermark
signal. Here the bit decomposition method oI gray images
which was proposed by Niu Xiamu is adopted to decompose
images into 8-bit planes|9|. Each pixel oI every plane only
contains 0 or 1.A image with size oI 3 3 can be
decomposed into nine 8-bit panes, as shown in Fig.1.
In Iigure 1, (a) is the gray value oI original image, (b)
is the highest bit plane aIter decomposing, (h) is the lowest
bit plane. Part oI the decomposed inIormation can be used
as watermark, the rest are used as the key Ior watermark
extraction.
(a) original image (b)
7
E (c)
6
E
(d)
5
E (d)
4
E (e)
3
E
(I)
2
E (g)
1
E (h)
0
E
Fig.1 bit decomposition Iigure Ior 8 bit gray-level image
% (PEHG ZDWHUPDUN LQWR ZDYHOHW GRPDLQ
II watermark is added at the position with large
coeIIicients, image will be easily distorted. But iI watermark
is added at the position with little coeIIicients, the
watermark will have the shortcomings oI weak resistance to
image compression, Iiltering and cropping etc. By the
clustering process introduced above,
) 1 ( 1
,
1 1
$ $ etc are
the classes with lager absolute oI wavelet coeIIicients.
Wavelet coeIIicients included in
0
$ are near to 0. So
watermark is added to those coeIIicients oI the classes
between
1 1
$ and
0
$ as well as
0
$ and
) 1 ( 1
$ .
Assume the classes which watermark will be added to
are
2 1
,...,
N N
$ $ . To describe conveniently, these classes are
remarked as
N
% % % ,..., ,
2 1
, 2 ) ( 2
2 1
+ = N N N . The
clustering center is also remarked as
N
T T ,...,
1
. According to
a certain
L
% , let
L L L L L L
T % U % T U = = ) max( ), min(
2 1
.
The clustering center and radius are used as key Ior
watermark extraction. Assume D is an integer, to the
class
L
% N L , , 2 , 1 " =
|) ,
1
| (
|)
1
, | ( )| max( ), |min(
2 2
1 1
L L L L
L L L L L L
U
'
G
T U
'
G
T
U
'
G
T U
'
G
T % %
+
=
*
* *
,
here ' G , , 2 , 1 " = : every coeIIicient oI
L
% must belongs
to one oI | ,
1
|
2 2
L L L L
U
'
G
T U
'
G
T +
+ and
|
1
, |
1 1
L L L L
U
'
G
T U
'
G
T
+ then
iI 1 =
M
S then
iI 1 ) 2 , ( = G 0RG then , , =
iI 0 ) 2 , ( = G 0RG then
2
1
L
U
'
, , =
else iI 0 =
M
S then
iI 0 ) 2 , ( = G 0RG then , , =
iI 1 ) 2 , ( = G 0RG then
2
1
L
U
'
, , =
else iI
1 1
1
L L L L
U
'
G
T , U
'
G
T
s s then
iI 1 =
M
S then
iI 1 ) 2 , ( = G 0RG then , , =
iI 0 ) 2 , ( = G 0RG then
1
1
L
U
'
, , + =
else iI 0 =
M
S then
iI 0 ) 2 , ( = G 0RG then , , =
iI 1 ) 2 , ( = G 0RG then
1
1
L
U
'
, , + =
3II / M < , then turn to the step 2, otherwise, stop
calculating and complete the process oI insetting watermark.
According to the original image , , through adding
water in wavelet domain and inverse wavelet transIormation,
the image , ' with watermark will be attained.
IV. EXTRACT WATERMARK
To extract the watermark, the image , ' that with
watermark is decomposed three times into wavelet domain,
then check the wavelet coeIIicients. The process oI
watermark extraction is shown as Iollows:
1) Let initial-value 1 , 1 = = M L .
2) Judging the present coeIIicient whether belongs
to
L
% , iI it is true, then we can get the conclusion that we
have inserted watermark into this point.
iI
2 2
1
L L L L
U
'
G
T , U
'
G
T + s s
+ then
iI 1 ) 2 , ( = G 0RG then 1 =
M
S
iI 0 ) 2 , ( = G 0RG then 0 =
M
S
else iI
1 1
1
L L L L
U
'
G
T , U
'
G
T
s s
iI 1 ) 2 , ( = G 0RG then 1 =
M
S
iI 0 ) 2 , ( = G 0RG then 0 =
M
S
Then 1, 1 L L M M + + . II N L = , then 1 = L , at
the same time checking the next wavelet coeIIicient:
otherwise, it shows that we haven`t added watermark, and
then check the next wavelet coeIIicient directly.
3) II / M < , then turn to step2, otherwise, stop
calculating, and completing the process oI extracting
watermark.
AIter the above steps, the ensemble }
M
S ( / M " , 1 = )
which contain watermark inIormation is Iormed. Then
combining the key, revert the ensemble which contain
watermark to gray image, thus the watermark image is
obtained.
V. EXPERIMENTAL RESULTS
In the experiment, the 256 grayscale image Lena
( 256 256 ) is used as a carrier (Fig3(a)), the watermark is
a 32 32 grayscale image, 4 = '
,
4 = 1 , then aIter the
clustering, wavelet coeIIicients are divided into
7 1 4 2 = class. In the paper, the peak value signal-to-
noise ratio (PSNR) is adopted to evaluate the diIIerence
between the original image and watermark embedded image.
In order to test the robustness oI the algorithm, adding
Gaussian, median Iiltering, average Iiltering, JPEG
compression and cutting are made to the image with
watermark,, and checking the images.
(a) image without watermark (b)image with watermark
Fig.2 original image and watermark
The experiment indicated that the peak value signal-
to-noise ratio (PSNR) oI the original image and the
image with watermark is 38.1678 dB. AIter adding
watermark to the image, there is almost no decline oI
image quality, and the watermark can be extracted
without distortion.