(3.2)
e change in
|. and 0 and
ariables with
certain frac-
1
ayis./—e *
'ge for strong
figs. 3.2 and
At each grid
by state ac-
on the curva-
ing count for
propriate di-
of the count.
V
=0.02,
1 1 1 1 L 1 1 L 1 1
e 1 g 3 4 5 6 7 8 9
Fig.3.3. Random walk simulation with K,=0, t=15, op=0.5, 6, =0.01.
For the computation of the energies H (eo) a neighbor-
hood system of two-site cliques is defined (Koch & Schmidt,
1994). Each site has neighbors of varying order forming a
clique with each of those neighbors. Fig 3.4 shows the neigh-
borhood system of site s up to order 5. For an element e, of the
state space Eg, the counts of the random walk model are
summed clique by clique. Each count depends on the parame-
ter values £r and €, i.e. on direction and curvature in the
neighboring site f and the direction and curvature proposed at
site s, as well as on the location of / with respect to s. If Er is
"no line", the counts at s are 0 independent of £,. High counts
for £, indicate a high probability of £,, as the presence of a
neighboring line site supports the presence of a line with a
certain direction and curvature.
514130 475
4121112
4
3 1 S 1 3
d | 2/31 | 2, 4
S4 4/3144 $5
Fig. 3.4. Neighborhood system for a two-site clique Gibbs field. Site s and
another site form a two-site clique of neighborhood order n shown in
the graph.
We now extend our two-site clique neighborhood model, as a
line should also make certain neighboring lines improbable.
This is because line sites parallel to a directly neighboring line
site do not conform with the elongatedness of lines. It can be
modeled with the same type of random walks. We only
imagine a different type of particles, called inhibiting particles,
diffusing perpendicularly to the direction of a line site. The
particles inhibit the presence of lines perpendicular to the
direction of propagation in the same way the particles used
before supported the presence of lines in the direction of
propagation. Therefore, the corresponding counts make the
presence of those lines improbable and are subtracted from the
supporting counts.
Introducing a one-site clique containing only s, we can control
the overall probability of a line independent from the state of
neighboring sites.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Hence, in agreement with (2.8), H,Íe,|àe;) is computed from
a o -In(c)-» i£c2z1 (2
AENOE = :
SNESIENST Vd In(1) if C<1
C= Cy, (e,) SY Ca, (eles: sie A;)
A,
where C, ] is the count-equivalent of the one-pixel clique:
C if €, ="line(6,, K;}"
C. (e,) 4 l 5 ( 3 J i
1 " : " ; "
à if €, ="no-line
cj and c,, are empirically chosen "basic currents" which control
the overall probability of line and no-line sites. »,C 4, is the
A»
sum of the counts of the two-site cliques containing s.
3.2 Specific Knowledge from GIS Data
The intention is to use GIS data to support the extraction of
linear structures. It may for instance be known that a road is
crossing the imaged area, and an approximate registration of
the SAR scene and the GIS data may be given. Around the
projection of the road center line into the SAR scene the prob-
ability to detect a line with the direction and curvature of the
road center line should be increased. These facts have to be
used to compute the energy of the prior PDF.
It is known that the registration of SAR and GIS data can only
be accurate to a limited degree. What is more, the decision
about the exact location of the linear structure in the results of
the algorithm has to depend on the SAR data and not on the
given geographic information. Therefore, a corridor symmetri-
cal around the object center line is defined inside of which the
probability of the object class e, = line(0;,x;) is uniformly
increased. 0;, K; are the direction and curvature of the object
center line at the point i closest to site s (Fig. 3.5).
Center ] ine
a | corridor
Fig. 3.5. Corridor around an object center line in which the detection of
lines with direction and curvature of the object center line is
increased.
The parameters of the algorithm are the width of the corridor
which depends on the accuracy of the registration, and the
amount by which the probability of line detection is increased.
The increase in probability is taken into account by changing
the computation of Ca; in (3.3) to
315