The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
434
d(x) is the distance between x and C(p, t=0), and the plus
(minus) sign is chosen if the point x is outside (inside) the initial
curve C (p, t=0).thus we have build the initial level set function.
The equation for the evolving function which contains
the embedded motion of C (p, t) as the level set {0 = 0) is given
out by equation^ = V ;; ГС . V(k) is the speed function and
we refer to this as a Hamilton-Jacobi “type” equation.
There are many advantages to this Eulerian Hamilton-Jacobi
formulation, the most important are that it avoid the evolution
of the parameters of the curve equation, makes curve topology
changes(split or merge) become very natural and as evolving
function 0$C. ?> always remains a function, it can be easily
carried out by using numerical approximation methods.
2.2 C-V model
In 2001, Chan and Vese proposed an simplified model based on
Munford-Shah segmentation model. Suppose I(x,y) is the image,
there are a few homogenous regions in an image, Q is the edge
of region Rj ,the DN value of very region is a constant, that is to
every region Rj , I(Ri)=constant(Ci). What the C-V model to do
is create an minimize energy function to find the optimal
segmentation, which makes the distinction between segmented
image
. _ f вднешкОД ifll Inside Cl
w t£p ¡f Ri sutiri» Ci
and I(x,y) is the smallest.
We use an simple example that given by Chan and Vese in their
paper to explain this idea(Tony F.Chan, Luminita A. Vese,
2001)(Fg.l). Define the evolving curve C in O, as the boundary
of an open subset of П. inside(C) denotes the region ,
and outside(C) denotes the region Q\<k). The image u 0 is
divided into two regions by curve C: Uq and Ug. The average
DN value of very region is C| and C 2 .Then consider the
following “fitting” term:
F- CO ~ F; it) = I lii-Cx V-) - ?,J : ¿жф
In this simple case, it is obviously that C 0 , the boundary of the
object, is the minimize of the fitting term
WWMQ+ F-CCJ9 * 0 * F a CC«H F,(C„) For
instance, if the curve C is outside the object, then F5 (CJ > 0
and F e (Q*o. If the curve C is inside the object, then
F S <CJ*0 and F; C If the curve c is both inside and
outside the object, then MC»& and F; CO > ^.Finally,
the fitting energy is minimized if C=C 0 , if the curve C is on the
boundary of the object.
Based on this idea, Chan and vese create the energy function
F(C- defined by
FCffc, sv. C) = u>L«n$&.0 v ■
+¿4 * [ h*c0s-> r 5 -
Where
}t ^ 0, V 2 0, 2 € are fixed parameters.
L$ngtfi(C) is the length of the curve C
AmClnsJdeCC)) is the area of the region inside C
LtngthfjC) Ar«&(insld«(0) are regularizing terms.
By solving the minimization problem iwig^(<B| t Vy, O,
we can get the final curve C.
Figure 1 .all possible cases in the position of the curve
3. EXPERIMENTAL PROCEDURES
The experiment procedure of this paper is directly and in a
semi-automatic method. At first, the program will get initial
edge by Human-Computer Interaction way. Usually, initial edge
was simply given by using circular or rectangular, because it
would be very easy to build signed distance function. But when
we use this way to process remote sensing image, it will cost
very long time to make the curve move to the real edge. Though
the algorithm of building signed distance function will become
a little more complicated and time costing, because the initial
edge is much more accurate, the total time will be shorter.
After getting the initial curve, the program will build signed
distance function and evolve the curve automatically.
We provide two ways to stop the evolving of the curve. The
first way is that the program judges if the curve has moved to
the right place and determines to continue or stop by itself. The
second way is to stop the curve by human sending commands,
this way is more effective because we use our experience to
determine if the curve has moved to the right place.