The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008
JJ f(p)dp= ¡¡np)(l-H(^p)))dp (8)
R M (C(0) n
Then, the energy function is:
E(W) = JJ (u,v,t)dudv
R,„ R oM
= ¡¡¡¡ni-Hw«mmm))dudv
n o
In this way, the problem changes to find a level set function (p
with which (6) gets its minimum. A simplified form of this
problem is depicted as below function:
d<p(v)
dt
= 8(<i>{v))
\\co(v, u)xv((/>(u))H (<f>(u)du
[-*
\\w(<l>)H(</>(u))du
u)w(<p(u))( 1 - H ((f>{u)))du
_R -
JJw(^(w))(l - H ((f>{u)))du
\-yK\S7(f>\
(9)
R
where w(.) is weight function, and then w{(p{p)) is the
dissimilarity weight of the pixel p and k is curvature of (j) .
3. RESULTS AND DISCUSSION
In Fig. 1 and 2, we applied our method on a real high resolution
remote sensing images, Fig. 1 is an image of a harbour and ships,
and Fig.2 is an image of a buildings, after 500 iteration a
reasonable good segmentation results were gotten.
a. initial curve b. 110 iteration
c. 370 iteration d. 490 iteration
Figure 1 harbour and ships
2.3 Watershed -based GPAC
The computation of the dissimilarities possesses most of
memory and CPU resource during the WGPAC method. As for
a image of M*N, a dissimilarity matrix W of size MN*MN
should be calculated. Although IF is a symmetric matrix, it
requires very large memory resource.
We introduce a watershed pre-segmentation algorithm. First, we
segment the image into m regions A k (k = 1,...,m) by
watershed methods. Let F{x, y) is the property of pixel (x, .
As for region A k , the average property F k is:
K =
1
Area(A k )
£ F(x,y)dxdy
(10)
Where Area(A k ) is area of A k . The dissimilarities matrix
W becomes a new dissimilarities matrix W' of ATNxm ,
where the dissimilarity between pixel (x,^) and region i is:
co\x,y,i) = \F{x,y)-F.\ (11)
In order to reduce the number of regions, first we minify the
original image n times and get a result image /,, second /, are
segmented into S x by watershed method, last we magnify S { n
times and the magnification result S 0 is the target regions.
c. 380 iteration d. 510 iteration
Figure 2 buildings
In this paper, we introduce a weighted and hybrid active
contour model. The problem of image segmentation is
converted into minimization problem of the total dissimilarities
between the pixels by defining different dissimilarity metrics.
Our algorithm is flexible because the dissimilarity is easy to
