The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008
606
image are calculated and recorded in the Table 1. The expected
of the road’s gray value is much lower than that of the
background in despite of the speckle noise effect, and it is
always below 80. Meanwhile, the variance of the road is also
much more lower than that of the background. Thus, another
conclusion can be made that the road area is more homogenous
than its background.
statistical property we discussed in the previous section . An
almost automatic algorithm is devised with the assumptions that:
1 ) There is obvious contrast between road and its background; 2)
The road itself is homogenous; 3) Road area is more
homogenous than its neighborhood when taking the road
candidate into account.
According to the first assumption, the local structure model of
the road area is formed by three paralleling parts shown in
figure 3 from which we can see that B represents the road area
whose center is the target pixel while A and C represents the
surrounding background area(Touzi, 1989; Geling, 1993). By
computing the summation of the pixel gray value of each area
and the average ratio of the road area to the background area
according to formula (1)
(b) Vertical direction
Figure 2. Relation of the gray value sum and coordinate
Road
1
2
3
4
5
E
30.6
54.52
13.12
19.24
41.16
V
6.29
10.25
3.22
4.54
7.68
Non
1
2
3
4
5
E
173.8
134.9
195.2
107.7
154.1
4
6
4
2
2
V
32.77
24.94
35.88
20.29
28.73
X
/= 1
m
R 2
X B,
(i)
we can get Ri and R 2 . If R]<1 or R 2 <1, then the central pixel of
area B is a road candidate point. This operator can be used in 4
or 8 directions, the minimum value is chosen to be the best
response. The first operator replaces the target pixel’s own gray
value with its neighborhood gray value and thus controls the
speckle noise influence to the road extraction effectively.
While on the other hand, the replacement of target pixel’s gray
value by its neighborhood’s will certainly lead to wrong
detection, regarding the non-road pixel along the road edge as
road candidate which also lead to the road fracture. In the
previous part of this article, we discussed the statistical property
of the road in SAR images, and concluded that the road area is
homogenous in SAR images and is more homogenous than the
background area in the meantime. In order to solve the previous
problem, a statistic is introduced to describe the homogeneity of
the imagery. Chose a 3X3 or 5X5 window with the target
pixel at the center and calculate the variance(D) of this area,
then we can get R3 by formula (2).
Table 1. The expectation and variance of the road and
non road area(E: expectation, V: variance)
3. ROAD LINEAR FEATURE DETECTOR
In order to reduce the influence of the multiplicative speckle,
the method proposed in this paper is based on the gray
«3 =
(2)
As there are some homogenous areas in parts of the background,
we take the pixel gray value(P) into consideration and get R4
by formula (3).