International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
  
  
  
  
Table 1. Comparison of Experimental Results 
Testing Type Testing COX method DCT Entire algorithm Our algorithm 
: Number 
Filter JMedian, Gauss, FMLR, Sharpening! 5 5 5 
JPEG compression 10 10 10 10 
Symmetrically and non-symmetrically 5 5 5 5 
moving rows and lines 
Current linear geometrical : 5 3 5 5 
transform 10 10 10 10 
Changing x-y axes display scale > > 5 5 
Scale transform 5 3 5 
Cropping in x-y directions 15 8 10 13 
Circumvolving with cropping and no scale 
transform 15 7 9 10 
Circumvolving with cropping and scale 
transform 10 2 3 8 
Center cropping 2 2 2 A 
StirMark random bend 
Summation 87 59 69 78 
  
  
  
  
  
  
  
6. Conclusions 
In this paper we proposed CBWM localized adaptive 
watermarking algorithm base on DWT. In this algorithm, we 
firstly extracted edges and chose feature points in the original 
remote sensing image, then established each sub-image, in the 
center of which is the feature point, and embedded the local 
watermarks into the feature vectors in the three-level DWT 
low-frequency space (LL3) of sub-images and the local 
watermark is orthogonal to the feature vector, 
watermark is sub-image content-based. In addition the 
embedding procedure was adaptive, namely we firstly 
organized wavelet coefficients into wavelet blocks, classified 
those wavelet blocks by texture intension and according to the 
classification results, adaptively embedded watermark weights 
of different strength into different weights of the feature vector. 
Furthermore the algorithm exploited the Neymann-Pearson 
criterion to detect local watermarks. And we exploited StirMark 
as the testing instrument to verify robustness of the algorithm. 
And experiments testified that the algorithm has a strong 
robustness, simultaneously a good resistive ability against filter, 
noise, geometrical transform, remote sensing image 
compression, x-y direction cropping and StirMark attack and 
has a certain robustness against remote sensing image rotation, 
which needs to improve further. In addition, after embedding 
watermarks, there is little influence on such applications of the 
remote sensing image as edge detection and image classification. 
Therefore the watermarking algorithm is a practical one for 
remote sensing images. 
Moreover, we could improve robustness of CBWM localized 
adaptive watermarking algorithm based on DWT for remote 
sensing images from the following 2 aspects: 
1) theoretically as long as part of the remote sensing image 
after cropping contains at least one feature point and the 
corresponding sub-image, the watermarks can be detected. But 
after compression and other image processing (such as filter, 
noise adding), the feature points in the remote sensing image 
can't all be detected, which would lead to failure of 
namely . 
196 
watermarking detection. So when detecting watermarks, we 
should detect not only from the sub-image of the feature point, 
but also from all candidate points and the points in the near 
areas around them. As long as in one sub-image corresponding 
to some point, the watermark can be detected, we think there be 
localized watermarks in the image. But that would increase 
complexity of the computation; 
2) the algorithm is more complicated in computation, so we can 
consider making heuristic search by heuristic knowledge to 
decrease the computation during the course of optimization 
computation. 
In addition, we can consider to improve robustness of the 
algorithm from 8 aspects: 
1 )making feature points possess stronger stability; 
2)exploiting the algorithm which is more suitable to embed 
local watermarks in remote sensing sub-images, consequently 
improves robustness of local watermarks; 
3)choosing more appropriate size of local area to embed 
watermarks, therefore further increasing the ability to resist 
cropping and StirMark attack and the robustness of local 
watermarks; 
4)choosing more suitable scale gene ‘a’ to better control 
embedding intensity and energy of local watermarks; 
5)classification of wavelet blocks can also be done in the spacial 
area (Huang Daren, 2002a). Each wavelet block all corresponds 
with a sub-image of the same size in the spacial area and those 
sub-images are not overlapped and formed a kind of division of 
the original remote sensing image. In addition edge points 
represent those breaking points of image pixel grays, therefore 
if a sub-image has more edge points, it would have stronger 
texture; and so the texture of the wavelet block corresponding 
with the sub-image would be stronger. Therefore we can 
consider to classify the wavelet blocks by exploiting the density 
of edge points; 
6)managing to make watermarks resist cropping as well as 
zoom and general affine transform; 
7)developing the digital watermarking technology based on 
features (Yi Kaixiang, 2001a). Watermarking technology based