corresponding to the lowest resolution. Each pyramid level
image is then smoothed by a 3x3 Gauss filter. White noise with
certain amplitudes is added afterwards. In these pyramid
images, the feature extraction operators were run, varying the
values for contrast (between line and background) and noise.
For the edge extraction algorithms, the results of the Canny and
Deriche operators were followed by a non-maxima-suppression
and thresholding. Finally, the skeleton was derived. The edge
detection sequence is depicted in Fig. 2. The response of the
Steger operator in the line detection algorithm was simply
thresholded.
un
Edge , Non-Maxima-
Detector ^ ^ "| Suppression | 1 Thresholding L3 Skeleton
Figure 2. Edge Detection Algorithm
The edge and line detection algorithms were optimised for the
smallest pixel size of the image pyramid, i.e. the resolution of
the creation stage. The described parameters were maintained
for the edge and line detection in all pyramid images. The
images were all processed with the same procedure (same
operators with the same parameter values) to ensure
comparability of the results. In this paper we call one and the
same operator with different parameters as different operators.
Performance of the operators was obtained by recording the
ratio of the actual edge and line length of the operator in the
image to the expected well known length of the edges and line.
All operators yield 100% performance in the first stage.
125,00
Usability Thresholds —— — —
10000 Joe rer t rrr en
8 7500 Canny : \
E —- Deriche tA
S 5000 ~ Steger ; \
t J
© \
n. |
25,00 |
ee
0.00 : iEn TRIN té Erro
Qu aM dh am C qus A quas d Gd A ud
SPP DRS SPD PR PT gx
SNS Sa a oF oF A PS 4^ 4°
Pixel Size [x-fold]
Figure 3. Performance of the Canny, Deriche and
Steger Operator
The dependence of the feature extraction operator’s
performance on image resolution was analysed. The
performance is gradually decreasing for lower resolutions. Fig.3
shows the performance curves of the operators in the highest
resolution image of pixel size 1.0 with a grey value difference
between the dark background and the white line of 240 and a
noise amplitude of 3, corresponding to approximately 1% of the
grey value range. This noise level was chosen to simulate a
realistic noise impact on digital images.
As can be seen, the performance curves of the three examined
operators behave quite differently. The difference in the shape
of the performance curves is not only due to the operator itself,
but is dependent on the chosen parameters in the
implementation as well. The choice of the threshold values
mainly determines at which pixel size the good performance
breaks off. While the Deriche operator’s best performance more
or less ends abruptly, the performance of the Canny and Steger
operators oscillates over a certain range of resolution before
failing to detect features. The results of these two operators in
these resolution ranges must be regarded as unreliable. Feature
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
extraction at these and lower resolutions cannot be carried out
with these operators under the preference of the chosen
parameters. The derived usability thresholds are marked in
Fig3.
25
-—Canny
7 Deriche
20 Steger
S
o
15
X
o
D »
o 7
3 10 /
a npr T
ks
5 p a erem
E
exu
pci
Q ee
0 50 100 150 200 250
Grey Value Difference
5
Figure 4. Usability Range of Canny, Deriche and Steger
Operators with varying Contrast
Contrast in the image plays an important role in feature
extraction and influences the operator’s performance. To
determine the resolution up to which an extraction with the
presented operators with a given contrast is reliable, the limit
for the operator performance was set to 98%. If the output of
the extraction algorithm falls below this limit, the operator was
regarded unusable for the respective image resolution on to any
lower resolution, at least with the implemented parameters.
Fig.4 depicts the performance limits for the three operators with
a given noise level of 1% depending on the grey value
differences between the dark background and the lighter line. A
pixel size of zero means there were no operator responses even
in the highest resolution of 1.00 because of insufficient contrast.
25 ;
! *- no Noise
| —+ Noise 3 frt
| Noise 5
20 |
v i
o |
15.
= |
o i
A
o
3 10
=
n. i
I
5 |
|
0 Lane
0 50 100 150 200 250
Grey Value Difference
Figure 5. Usability Range of the Steger Operator with varying
Contrast and Noise
Furthermore, feature extraction is also susceptible to prevailing
image noise. Analyses were carried out to the operator’s
performance with three white noise amplitudes — 0, 3 and 5,
corresponding to 0% - 2% of the 8-bit grey value range. With
increasing noise level in the image the extraction performance
declines. The sensitivity of the Steger operator to noise is
exemplarily presented in Fig.5. The Canny and Deriche
operators exhibit a very similar behaviour in varying noise.
The performance of the Canny, Deriche and Steger operators is
degraded by the influence of low contrast and high noise. The
smaller the grey value differences and the higher the noise
amplitude, the lower the image resolution at which the trustable
performance of the operator breaks off and the smaller the
resolution range for which the operator can be used.
Internai
[Roadv
| Margin
| Cont.S
[vs
ext
| ie
i
|
|
i
|
|
Figu
For an «
created
describe
from ar
goal of
semanti
digital
automat
In the
modelle
values ¢
various
the roac
line, th
Attribut
the obje
spatial 1
scale ac
Here, th
correspx
the roa:
marking
connect
only the
addition
search
objects
conjunc
The foll
adaptati
The ada
of large
seen in
smaller
necessit
appropr
