unts 
This 
n to 
om- 
iper 
| for 
s to 
)) is 
ality 
d as 
ster 
ctor 
ta is 
aps. 
the 
rlay 
| the 
:'k of 
Isive 
ator. 
the 
uted 
> the 
One possibility is to use techniques from pattern recog- 
nition and line drawing analysis. For maps of low com- 
plexity (eg cadastral maps of scale 1 : 500 or 1 : 1000, 
forest or water plates of a topographic map) commercial 
software packages are available which use these tech- 
niques (eg STRUVE from M.O.S.S., Germany, GEO- 
REC from SysScan, Norway). They operate in batch 
mode accompanied by manual pre- and postprocessing. 
The amount of manual editing is closely related to the 
complexity and the quality of the maps. For instance for 
cadastral maps of scale 1 : 500 of very good quality for 
the GEOREC system time savings of 50 % compared to 
manual digitising were reported [Schmitz 1991]. How- 
ever, these techniques fail, if more complex maps of 
lower quality are to be converted to GIS data. 
In this paper we introduce a multi-level model for a 
large-scale map as the key to improved and robust auto- 
matic map interpretation. The levels of the model are 
called semantic objects, graphics and text, image graph 
and image. Objects, operations to be performed on the 
objects and relations between the objects are presented 
for each level. We first shortly review some work done 
in the knowledge based interpretation of maps. Then, we 
introduce our model for the interpretation of a large- 
scale map and its representation by a semantic network. 
The flow of the interpretation is explained next, followed 
by first results. 
2. FORMER WORK IN MAP INTERPRETATION 
The basic operation for processing line drawings is line 
extraction from a raster image. Often this is done by 
thinning or by a distance or medial axis transformation 
of the binarised image [Smith 1987] followed by line 
tracking and line approximation. Another possibility for 
obtaining the lines is the extraction from grey-valued 
images [Joseph 1989]. It returns lines and line crossings. 
A first step in processing the lines (eg [Domogalla 1984]) 
is to attempt to reconstruct the elements of the graphics 
(eg points, straight and curved lines, circles, arrows, 
hatched and screened areas and other kinds of map 
symbols). It should be pointed out that in nearly all cases 
this reconstruction can only be partially successful, be- 
cause of errors in the input data (in the map itself or 
introduced by scanning or preprocessing) and because 
of a lack of higher level knowledge. 
579 
After the partial reconstruction of the graphics the in- 
terpretation of objects of the domain (eg buildings, 
roads), denoted by the graphics is attempted using hig- 
her level knowledge. The work we will present in the 
following is ordered according to an increasing use of 
high level knowledge. 
In the system CAROL [Illert 1991] designed for the 
interpretation of large-scale maps (eg German base map 
1:5000) emphasis is laid on the improvement of different 
interpretation modules. The text and the symbols are 
recognised by classifying the contours expanded in a 
Fourier series. The system also recognises hatched areas 
and dashed lines and combines symbols into strings (eg 
digits to numbers). At last, strings and symbols are as- 
signed to spatial features (lines, points). 
MARIS [Suzuki and Yamada 1990] is a system for the 
interpretation of large-scale maps of Japan. At first, 
special algorithms for tracing closed polygons are util- 
ised to find buildings. Afterwards, height contour lines 
and lines representing railways, roads and water areas 
are recognised. 
T. Kilpeldinen [Kilpeldinen 1988] tries to recognise 
buildings in city maps. She only describes the buildings 
(local knowledge) on several levels of detail, using the 
programming language PROLOG. Although she does 
not model the surroundings and the relations among the 
buildings and the surroundings, she obtains high recog- 
nition rates. 
Another direction is chosen by J.A. Mulder [Mulder 
1988]. He investigates topographic objects and con- 
straints between them. Then he analyses simple line 
drawings containing these topographic objects using a 
constraint satisfaction technique. 
M. Ilg [Ilg 1990] uses domain dependent knowledge to 
analyse road maps, using bottom-up and top-down pro- 
cessing for hypothesis generation and missing parts pre- 
diction respectively. He uses the exo-skeleton (skeleton 
of the background of the image) to find parallel lines. 
CIPLAN [Antoine 1991] is an experimental system to 
interpret French city maps. It uses a model containing 
"real entities" (eg buildings, roads), their graphical rep- 
resentations and the relations between them together 
with a procedural network for the extraction process.