gray image
an
DEM : illuminated DEM
Figure 7: Gray image with corresponding illuminated DEM
e The image distortions can be modelled by a linear
system.
e The original image and the observed image can be
considered as stationary random.
e There is no correlation between noise and image.
e The noise in the amplitude only (not in the phase) of
the Fourier spectrum.
The implementation usually is done via convolution in the
Fourier domain. To build the impulse response of the Wiener
filter, the spatially invariant impulse response of the image
distortion, and the power spectral density of the original
image and the noise are needed. Wiener filtering of an
observed image produces an estimation of the undistorted
original image that is optimal in the sense of a minimal mean
square error between the estimated and original image.
Others: Hysteresis smoothing can remove minor fluctuations
while preserving the structure of all major transients. The
anisotrope diffusion is an iterative, anisotrope smoothing
operation on the basis of physical diffusion. Low edges are
suppressed while step slopes remain. The sigma filter is an
average filter with gray values in the neighborhood that are
close in value to the center value. Thus we have a simple
local adaption to noise or texture.
In general it is very difficult to decide which of the operators
mentioned above has to be selected for a given task because
no theory exists which allows a comparison. To have a short
impression of the effects one example is given. In figure 9 a roof
with noise due the grain of the film and some texture and the
elimination using a Gauss filter and the anisotrope diffusion (left
to right) can be seen. In this case it is obvious that the result of
the anisotrope diffusion is better, because edges are preserved and
noise is eliminated.
4 OPTIMAL RESOLUTION
For the extraction of a class of objects one has to select the res-
olution depending on its shape and radiometric properties. If
the resolution is too high the details complicate the segmentation
and interpretation. The advantage of a lower resolution is the
reduced number of pixels (runtime) and the reduced size of the
objects (locality and generalization). If the resolution is too low
the objects cannot be extracted at all. In german there is a proverb
which explaines the problem: “You can’t see the forest due to
many trees.” If you want to extract a forest there is no need to
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
see the leaves. The selection of the optimal resolution simplifies
the problem significantly. Roads, for example, can be detected
as lines in a lower resolution (Fischler et al., 1981), (Aviad and
Carnine Jr, 1992), (Barzohar and Cooper, 1993), (Barzohar and
Cooper, 1995), (Berthod and Serendero, 1988). In figure 10
a house in high resolution (one pixel corresponds to 25cm) is
shown. Applying a Sobel filter yields many edges due to the tex-
ture of the tiles. Using a lower resolution (one pixel corresponds
to zz 0.75 m) results in the edges of the roof.
Looking at the selection of the correct resolution more closely,
one finds out that a single resolution does not suffice in many
cases: A higher resolution is needed to refine the segmentation.
1. There is an optimal resolution to extract the raw shape of an
object called initial resolution.
2. In many cases a higher resolution (refined resolution) is
needed to extract the exact shape of the object and to distin-
guish it from other similar looking objects.
This leeds to a multi resolution approach to segmentation: Seg-
mentation starts with the initial resolution. The results of this
step are verified and improved using the next refined resolution.
If necessary, the process is continued with further refined resolu-
tions (Heipke et al., 1995). In figure 11 the extraction of a road in
the initial resolution can be seen. Here roads are extracted as lines
of a given width. In the refined resolution edges are extracted.
These results are used for a precise detection of the road bound-
aries and the elimination of false candidates. For final results a
further refinement is needed to extract road marks.
uv me
inital refined twice refined
Figure 11: Extraction of primitives of roads in different resolu-
tions (three different scenes)
The resolution hierarchy might be invalid in some case. This
can be illustrated by the roads in figure 12. The gray value of the
asphalt is simular to its surrounding. Therefore the road cannot
be extracted as a line in low resolution. The road is defined only
by the marks found in the refined resolution. In this case the
168
texture of
interpretat
hypothese
resolution.
Figure 1:
More ir
(Gauch ar
As we hav
in an aeri
used to ex
of all clas
in the foll
broader cl
But no pr
