£ 
  
   
  
  
  
  
  
  
  
  
  
eo 
Figure 1: Example streetl: intensity image (top left), hue image (top middel), binary image (top right), extracted image lines (bottom 
left), map (bottom middle), landmark model (bottom right) 
2.1 Classification 
Two different image types were used: true colour images in scale 
1:8000, in which we classified grain fields, and near infrared im- 
ages in scale 1:6000, in which we extracted streets in rural areas 
and water surfaces. The images were digitized by scanner result- 
ing in red, green and blue images (RGB). These RGB images were 
transformed into a hue, saturation and intensity (HSI) represen- 
tation. The hue channel was used for classification. Figure 1 top 
middle shows an example of a hue image, which was calculated 
from a RGB image using the equation below [Frey 1990]. 
0.5- ((R — G) 4- (R — B)) 
VEG E-G) (E BC- 
H = arccos 
The only exception was the water surface, which was classified 
by thresholding the red channel. 
A region growing was performed on the used band before clas- 
sifying the pixels in order to stabilize the results and to avoid 
small regions. Pixels belonging to the object surface were ex- 
tracted by thresholding the average hue values of the extracted 
regions. The thresholds were fixed in advance, according to the 
object and image type. Although the images had poor spectral 
quality, the results of the classification showed to be robust to 
changes in the thresholds. 
2.2 Line extraction 
The thresholding was followed by a line and node extraction to 
transform the raster image (figure 1 top right) into a vector im- 
age. In the cornfield example this was done by tracing the con- 
tours of the selected regions. For images with linear objects, like 
the roads and rivers, a thinning algorithm [Arcelli and Baja 1985] 
was used to obtain a skeleton of the selected objects. Figure 1 
bottom left shows a line image, derived from the skeleton image 
by a line following algorithm. In this step lines, nodes and en- 
closed regions were extracted. Features caused by image noise 
like nearby nodes representing the same points, close parallel 
lines, and short lines were eliminated afterwards. 
The relational description of the landmark (figure 1 bottom mid- 
dle) was obtained by digitizing a map (figure 1 bottom right). 
2.3 Primitives and relations 
A structural description consists of a primitive part and a re- 
lational part. The primitive part contains geometric primi- 
tives like points, lines, and regions, which represent the ob- 
ject parts. The primitives are characterized by a set of at- 
tribute values like line length, line type or region size. E.g., 
a straight line of length 30.2 is represented by the primitive 
P1 : {(length 30.2) (shape straight)}. The second part describes 
the interrelationships between these primitives. Possible rela- 
tions are angles between lines, connections between points and 
lines and between lines and regions. These can be characterized 
by attribute values, too. E.g., an angle between lines p; and ps 
is represented by the relation tuple 7; : (pi ps (angle7?)]. 
Following [Shapiro and Haralick 1981], we use the symbols 
D4(P, R) and D3(Q, S) for the image and the landmark descrip- 
tion, respectively . P represents the set of primitives p;, Q rep- 
resents the set of primitives qj. R and S are the set of relation 
tuples r; and s;. 
970 
3 
In or: 
mark 
betwe 
select 
h:F 
qualit 
Intuit 
simile 
and r 
in th 
of pri 
more 
of pri 
A me 
prope 
3.1 
The 1 
a of 
the s 
I (a( 
The : 
tribu 
on th 
tion. 
of an 
Weav 
by 
So, af 
tion, 
The 
attril 
ity th 
value 
tion . 
a(pi) 
The s< 
attril 
Thes 
betw 
other 
As t] 
map] 
comp 
attril 
then 
this 1 
3.2 
'The 
from 
tive ' 
be m 
corre 
the | 
of th