NT, 
„In 
lote 
irks 
ient 
sed 
the 
for 
; of 
) in 
In 
the 
the 
IY 
the 
rity 
vith 
ily 
for 
DY, 
s is 
ove 
the 
an 
for 
ely 
the 
and 
rea 
nly 
rks 
ge. 
this 
ing 
the 
are 
rks 
ge, 
so 
ote 
age 
ive 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
  
embedding algorithm of the image content-based localized 
watermarks; in section 3 and 4, we respectively discussed the 
choice and detection of the localized watermarks and 
experimental results and oùr conclusion were presented 
respectively in section 5 and 6. 
2. Embedding of Localized Watermarks 
In our algorithm we exploited the spacial feature points to 
orient watermarks, then adaptively embedded watermarks of 
different strength into the wavelet domain of the only 
sub-image which corresponds with the feature point and 
detected watermarks to realize localizing of watermarks. We 
considered to choose those angle points (including crossed 
points, Y-shape points and T-shape points) as the feature points, 
because those points have good stability. After finding those 
feature points, we can embed the localized watermark into the 
D XD sub-image which is centered by each feature point. 
2.1 Strategy of Secrete Watermarks Embedding Based on 
DWT 
According to the distributed character and the qualitative and 
quantitative characters of breadth of wavelet coefficients, we 
exploited the new strategy presented in Document (Huang 
Daren, 2002a), namely watermarks should be firstly embedded 
into the low-frequency wavelet coefficients and then embedded 
into those high-frequency coefficients according to their 
important orders if there are still surplus watermarks. 
Furthermore watermarks should be embedded into the 
low-frequency and high-frequency coefficients by different 
embedding strength. 
2.2 Adaptive Watermarking Algorithm Based on HVS 
In addition to the position of watermark embedding, the 
robustness of watermarks also depends on the embedding 
strength of watermarks. So we should make adequate use of 
vision characters, namely under the condition of invisibility, we 
should rationally distribute the embedded watermarking energy 
and improve the strength of the local embedded watermarking 
weight as large as possible. And adaptive watermarking 
algorithms are just based on this idea. 
Given the above embedding strategy, in this paper we exploited 
the adaptive watermarking algorithm which introduced vision 
system characters into watermark embedding procedure, 
namely organized anew the wavelet coefficients into wavelet 
blocks, then according to the texture-hiding characters of vision 
system, classified those wavelet blocks and by the classification 
results, embedded the watermarking weights of different 
strength into different wavelet coefficients. 
2.2.1 Characters of Wavelet Blocks and the Relation with 
Low-frequency Coefficients: We made wavelet decomposition 
of the original remote sensing image, then each pixel is 
corresponding with some blocks in the wavelet domain, the 
relation among which can be denoted by a four-branch tree, 
namely a wavelet sub-tree. The root of a wavelet sub-tree lies 
on the low-frequency domain and we can organize all the three 
four-branch wavelet sub-trees of different orientations and the 
same root to form a block of the fixed size. All the blocks of the 
fixed size are called as wavelet blocks. 
2.2.2 Classification of Wavelet Blocks: In order to improve as 
193 
high as possible the watermark embedding strength, we 
classified wavelet blocks into 2 classification: classification 1 is 
the wavelet blocks of weaker textures, marked as s/ and 
classification 2 is the ones of stronger textures, marked as s2. 
The wavelet coefficients of large breadths are corresponding 
with the breaking pixels in the original remote sensing image, 
so the textures of a wavelet block would be stronger if it has 
more wavelet coefficients of large breadths, otherwise the 
textures would be weaker. Namely, if 
number( | Fu. vy > Tl(uv)O W.O) 5 T2, then the wavelet 
block Wy [0S Cotherwise "x €52 0 and TI T2are the 
preliminarily designed fields. In this paper, we chose 
T1-0,T2-45. 
2.3 Embedding Algorithm of Localized Watermarks: 
Ruizhen Liu has presented CBWM (Content-Based 
Watermarking Model) in Document (Ruizhen Liu, 2001a) 
‘Image Content-based Watermarking Model’. CBWM makes 
no hypothesis to the original remote sensing image and is an 
addition watermarking model and because in CBWM the 
chosen watermark is orthogonal to the feature vector, it’s also 
an image content-based watermarking model. CBWM is a 
universal watermarking model based on frequency domain and 
can be combined with any watermarking algorithm based on 
frequency domain. The image content-based localized adaptive 
watermarking algorithm presented in this paper is just an 
adaptive watermarking algorithm which applied CBWM to 
DWT of the local area (sub-image). 
The watermarking algorithm presented in this paper is remote 
sensing image content-based, namely each chosen local 
watermarking sequence is orthogonal to the feature vector of 
the corresponding sub-image, so we chose the low-frequency 
coefficients of DWT as the choosing space of the feature 
vectors and according to formula (1), adaptively embedded the 
watermark H' into the feature vector in the low-frequency 
domain. And we can exploit the texture-hiding character of 
vision system by adjusting the factor (*. By a large amount of 
experiments, we found: letting 
dk=d, +aw, » k=1,2,,K 
0.02 WkEs2 
a= (1) 
0.005 WkESsI 
the watermarked remote sensing image would have a good 
robustness. F(u,v) is the wavelet low-frequency coefficient, and 
where 
in our algorithm it is the weight d , of the feature vector in the 
low-frequency domain and Wk is the wavelet block in which 
d, lies. 
Let A={ai j) € FP be the original remote sensing 
sub-image and B={bi j) € FP be the adaptively 
watermarked remote sensing sub-image, then the procedure of 
the CBWM localized adaptive watermarking algorithm is as 
follows: