nposition will be ex-
iple: a set of n filters
f a specific frequency
"hus we have a vector
ts neighborhood with
ut for n-dimensional
1);
ppropriate filters have
)sed by (Laws, 1980).
structed from vectors
+ vj.
fe)
] 7.9. 1)
150 4-1)
jo 4. 1)
| —2 1)
(Shao and Fôrstner,
e and have some nice
, multiscale property,
spatial and frequency
le gauss shaped filters
ct to rotation and are
ion. Typical filters of
aws filters is given in
e calculated from the
is calculated using a
termask to generalize
circular mask (diam-
yy images can be used
ures like those of sec-
ich emphasize special
, Or corners.
, is defined by (Harris
2
> Gaussian filter with
s invariant with respect
k is 0.04. In this case
have negative values.
on is given in formula
Go * (9293) — @
y
ind not to the original
' highly curved edges.
naximums of the filter
a cross. Most of the
nd as well as corners
other filters for corners
|. Some of these can
TE
fit NNT ITT
ES ARN ENT TH
HT NN TIN NR
IAN NN NITE
NINN IE
/ THAN J] N UTI ' AN , n h
HON IRN [AS IH WW AN) NT
NN INN AR
Hh i ANN ND 7] NOY Jf | NS lij ! 1} \ ARN S
77 LIN NUN i PONE
TERN TEC ^ TUN NN
ONTO ZAHN NOR
HITT NAH fl UNS Ws AT i INN NN JS
Er TELLE] i D) RO
LES
Laws Gabor Gauss
Figure 22: Frequency response of different texture filter classes
original image filter ee filter rl median of ee median of rl
Figure 23: Result of two different Laws filter
roofs of similar width are selected. These have to be eliminated
using context information.
image 1 image 2
Figure 24: Result of extended corner response filter
original image = bandpass filter
be found in: (Dreschler, 1981), (Deriche and Giraudon, 1993), Figure 25: Emphasizing linear structures of a given with
(Tabbone, 1994), (Fôrstner and Gülch, 1987). In (Rohr, 1993) a
comparison of different operators is given.
As a second example for specialized operators the extraction 7 CONCLUSION
of lines will be discussed. The aim is to construct a filter, whose
emphasizes linear structures of a given width and suppresses struc-
tures beeing smaller or wider. We define the frequency f of a line Is was shown that the segmentation of aerial images needs task
as the frequency of the sinus wave which half period is equal oriented segmentation procedures. Depending on the class of ob-
to the the width of the line. Given the minimum and maximum jects, appropiite resolutions and procedures have to be selected.
width of the line and thus fmaz and fmin we define a bandpass This selection has to be done according to the object model. The
segmentation becomes more stable if additional data, like color
filter with the following properties: or a DEM, i, iSS used. One open question is how to merge
: - : segmentation results when processing different object classes si-
1. Suppression of frequencies below fmin. multaneously.
2. Suppression of frequencies above fmaz. All examples of this paper were programmed with the im-
age analysis system HORUS using the interactive user inter-
The frequency fax Can be choosen higher if sharper edges of the face HORU SDevelop (Eckstein and Steger, 1996). The im-
lines are required. In figure 25 an example of such a filter for the age examples have mainly been taken from the ETH-Zürich
extraction of roads can be seen. fmax has been choosen higher, and the ISPRS testset (Fritsch et al., 1994). Al images
so the edges of the road are fairly well defined. The extration of are available via ftp from: £tp://ftp.informatik.tu-
the road is simply a threshold operation. Besides the roads some muenchen.de/pub/rec/images/space/
173
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
WARF
TEE
