the 
rthe 
is of 
1 the 
line 
ding 
tion 
and 
, the 
»rian 
w of 
The 
> the 
reels 
epts 
e no 
sym- 
con- 
LBS. 
ts is 
: the 
gra- 
The 
| ac- 
(lar- 
ance 
f the 
fthe 
then 
ated 
À is 
The 
concept BL consists of the concepts polygon 0.35mm 
and BS. The concept polygon 0.35mm can be instantia- 
ted by analysing the thickness of the instances of the 
concept polygon. If the thickness of a polygon is beyond 
athreshold the concept polygon 0.35mm is instantiated. 
If it is below this threshold the concept TL is instantia- 
ted. 
The bottom-up instantiation of the concept BS is com- 
plicated. The concept BS consists of arcs (the small 
circles in Fig. 6). Hence one possibility to find instances 
of the concept BS is to search for arcs that resemble 
circles. However, due to distortions in the input image 
and because of errors introduced by scanning or line 
extraction the arcs are deformed and can be broken up. 
For this reason we decided to use template matching in 
the image to find the BS. 
4.2. Top-down instantiation of the concept 
boundary stone 
  
Top-down instantiation is a way to gain evidence for the 
existence of the BS. An instance of a concept poly- 
gon_0.35mm initialises the instantiation of a concept BL. 
By means of the neighbourhood relation instances of the 
concept BS are found, if they already exist. If they do not 
exist, instances of the concept arcs that could be parts of 
the BS are searched for. If they are found these instances 
are transferred to the concept BS for instantiation via 
top-down processing. 
The concept BS is now tried to be instantiated with one 
of these arcs as spatial reference. A compound opera- 
tion is used for the instantiation. A binary template of 
the circle symbolising the BS is matched in a specified 
area with the binary image. If the score of this matching 
lies beyond a threshold, points on the centre line of the 
template are matched into the image graph and the arcs 
found are marked as parts of the BS. Other meanings of 
these arcs are deleted. Furthermore, the topology of the 
lines ending in the BS is changed so that the lines inter- 
sect in the centre of the BS. 
The determination of the arcs is carried out to integrate 
the knowledge gained by the matching into the instances 
of our semantic network to keep it consistent. 
After the BS are found they are integrated as parts in the 
instance of the concept BL from which the top-down 
processing was initialised. 
583 
Because all the necessary parts are now available the 
concepts PA and GA can be instantiated. The parts of 
the concepts are found by an operation that calculates 
the neighbouring parts to a given part. Thus, the parts of 
the concept are found incrementally. If no more parts 
are found, it is checked if the related area is closed and 
if this is the case, the concept PA or GA is instantiated. 
S. IMPLEMENTATION AND RESULTS 
We use a small part of the map (512 pixels x 512 pixels 
scanned with 400 dpi; see Fig. 4) to test the interpretation 
capability of our implementation. 
The map is scanned and binarised. The result is a raster 
image in level 4. 
In this level disturbances and the screen symbolising the 
buildings are removed from the image which results in 
the cleaned image (see Fig. 5). This cleaning operation 
is based on the area of the connected components in the 
image. The remaining information will be used later for 
the recognition of buildings. 
On the cleaned image a topology preserving thinning 
operation and a distance transformation are performed. 
The lines are tracked, and a graph building process and 
a line approximation are performed [Maderlechner and 
Jeppson 1988]. The results are graphs in the image graph 
(see Fig. 6). For every graph also the sum of the length 
of all arcs and the area of the bounding rectangle are 
computed. According to these values the graphs are 
specialised into small graphs and large graphs. 
The instantiation is carried out according to chapter 4. 
Fig. 7 shows the result for the instantiation of the poly- 
gon_0.35mm. Obviously most of the labelling was cor- 
rect, but two short lines of the boundary stone in the 
lower left were classified as polygon 0.35mm. 
In Fig. 8 the six instances of the parcel areas can be seen. 
The boundary stones have all been found and the poly- 
gons that were wrongly classified as polygon 0.35mm 
(see above) have been recognised by matching the centre 
line of the boundary stone into the image graph and then 
deleted. 
Fig.9 shows the general areas. They partially correspond 
to the parcel areas. These areas are marked with the 
same numbers as in Fig. 8. The areas that are split by the 
topographic lines are marked by characters. Obviously 
all the areas that can be seen in Fig. 4 have been found. 
However, some of the general areas still have to be