building
sand
1 be reduced by using
| McKeown Jr., 1994).
'e often available. In
n two different views.
ut the combination of
tation with additional
ing (Haala, 1994).
view 2
different views
à is a digital elevation
manual or automatic
ke a laser scanner. A
which are higher than
). A popular operator
y opening. In case of
as an extension of the
1 and Munkelt, 1995).
rank operators. The
1 rank value while the
iximum - rank value).
opening, and the value
ig. In the case of n/2
ink value thus controls
1996
the behaviour of the operators. For the extraction of high objects
a rank value a few percents above 1 is used. The increase of the
rank value should be chosen proportional to the amount of noise
pixels.
As an example for the segmentation of a DEM figure 5 shows
a gray image of a hilly landscape with buildings and trees and the
corresponding DEM.
gray image digital elevation model
Figure 5: Hilly landscape with the corresponding DEM
At first the dual rank is applied to the DEM. The result corre-
sponds to the ground without objects (ground DEM). Subtracting
this ground DEM from the original DEM yields the result in figure
6, which is similar to the top hat of a gray opening. Only those
objects remain which are higher than their surroundings. The
extraction of these objects is now simply a threshold operation,
where the parameter is chosen according to the desired height.
The right image shows the results of the threshold operation. The
detection of the road (left below the intersection) is a bit surpris-
ing, but the road is actually higherthan the surrounding meadows.
Another problem in this example are the trees, which are too small
to be significant for the resolution of the DEM.
extracted objects
top hat of dual rank
Figure 6: Normalized DEM and the extracted high objects
Shadows are another class of "objects" that can be extracted
using the DEM. This is useful, because shadows cause a lot of
problems during the interpretation of images since they change
the gray values of objects drastically and add egdes or texture-
like structures. In the image of figure 7, for example, the road
is divided into light and dark areas. The extraction of shadow
pixels cannot simply be done by selecting all dark pixels because
other dark objects may be present. Instead the illumination of the
sun is simulated using the DEM (figure 7 center and right). The
segmentation of this image gives the raw shadows.
Due to the low resolution of the DEM the segmentation has to
be improved in the gray image using the following steps:
]. Elimination of small areas.
2. All those pixels of the remaining areas are selected which
have the intensity of shadows (i.e., their gray values are
inside a given range).
3. These pixels are used as seed areas for regiongrowing: Bor-
der pixels are added as long as the difference between their
gray values and the mean value of the area is below a given
threshold. In addition the number of iterations is limited
according to the maximal error of the DEM.
The result of this post-processing can be seen at the right of figure
8. These areas can be used to support the interpretation (Lin et
al., 1994), for example, to extend small areas of road hypotheses.
raw segmentation shadows
Figure 8: Segmentation of the illuminated DEM and refinement
using the gray image
3 NOISE CLEANING
In many cases some kind of preprocessing has be applied to the
images before segmentation. One reason is the grain when using
maximum resolution of the film, another is the elimination of
texture which complicates the segmentation. Different theories
and algorithms have been developed to solve these problems.
Some important classes are:
Lowpass Filter: The assumption of these filters is that noise has
a high frequency. The elimination of noise is therefore done
by suppression of high frequencies. Popular representatives
of this class are the average and Gauss filter. These filters are
very fast but the noise suppression is poor, especially using
the average filter, and important image structures like edges
are blured.
Rank Operator: These nonlinear operators take linear combi-
nations of the sorted values of all the neighborhood pixels.
Conceptually we can visualize the operator as sorting the
gray values from the smallest to the largest and taking a lin-
ear combination of these sorted values. The most common
rank operator for noise cleaning is the median. The median
suppresses small lines and points but edges are preserved.
Other types of rank operators have been proposed, like vari-
able size and shape of the neighborhood depending on the
noise, or the weighted-median filter which adds gray values
more than once (depending on the weight) to the sorting list.
Further information can be found in (Haralick and Shapiro,
1992).
Wiener Filter: Like inverse and pseudoinverse filters the Wiener
filter is used in the field of image restoration. The Wiener
filter, which models noise explicitly, makes use of the fol-
lowing model assumptions:
167
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Xe AE