ıtomation
he german base map
e increasing demand
werful techniques to
ildings and separates
ited by their outlines
nition, a raster based
e used to detect the
d columns, a kind of
standing task can be
ten am Beispiel der
steigende Bedarf an
istungsstarker Meth-
h Gebäude lokalisiert
strichkarte durch ihre
| im Bereich der kar-
an der schwarzen und
rasterbasierten Tech-
s Regionenwachstum
aufgabe bis zu einem
904), operational sys-
fying quality are still
erpretation using im-
automatic raster-to-
h complexity of map
jf the map as a whole
. stage of technology.
buildings. Spatial in-
n many fields such as
irban areas. Further-
he shape of buildings
port automated anal-
hange detection and
own by (Quint/Bähr,
rte 1:5000 (DGK5) is
map of Germany cov-
DGK5 represents the
other objects as black
e determined by their
Fig.la shows a subset
ined DGK5.
a 1996
This type of map was chosen for the following reasons: Due
to its large scale the objects remain true to scale and shape.
The mapped buildings are characterized by a high positional
accuracy ranging from 0.1 to 0.3 meters. Furthermore, the
data are more complete (every building covering more than 15
m? is mapped) and more actual than German cadastral maps
which merely contain 7096 of all existing buildings whereas in
the DGK5 95% of all existing houses are included.
In the DGK5 a building is represented by its outline filled with
a hatched pattern, an optional house number and, in few
cases some additional explanatory characters. The hatched
pattern consists of parallel and equidistant straight lines. The
types of buildings are characterized by different hatched pat-
terns varying in the slope of the hatching line with respect
to the longer edge of the building. Residential buildings are
marked by diagonal hatching lines (angle of 45?) whereas
outbuildings like garages are depicted with parallel hatching
lines (angle of 90?), see Fig.1.
residential building outbuilding
Figure 1: Different Types of Buildings of the DGK5
However, a robust recognition of these objects is difficult be-
cause the objects have manifold shapes reaching from simple
single rectangles to complex building agglomerations in city
centres. In addition, the hatched patterns can occur in any
direction and a wide range of spatial extensions, whereas ir-
regularities often arise. The most common irregularities are
broken hatching lines because of overlaying of other carto-
graphic objects, widened cross points at merging areas of
hatching lines with corners or edges of the building-outlines,
interrupted lines, varying equidistances of hatching lines and
non-parallelism.
2 OUTLINE OF THE NEW RASTER BASED
APPROACH
Existing approaches of automated map interpretation, e.g.
(Illert, 1990), (Hori/Okazaki, 1992), (Ablameyko et al.,
1993), (Mayer, 1994), mostly start with a vectorization of
the entire image content and only as a second step, they
try to find structures representing map objects. Vectoriza-
tion at this early processing stage always leads to a loss of
original information. These problems are caused by steps like
thinning, distance transformation and skeletonisation. As a
result, errors in position or topology occur (Klauer, 1993).
Another disadvantage of this approach is the large amount of
obtained vector data which is difficult to handle during the
various processing steps.
In contrast, this paper describes a new method for hatched
building recognition using raster based techniques until an
advanced recognition stage.
The developed approach suits the following main principles:
At the beginning hatched patterns only serve as characteristic
features to locate buildings and separate them from the re-
maining map objects. For that purpose the spatial extension
of hatching is recognized in the raster environment.
In a second step the identified hatched patterns are removed
so that only the building-outlines remain in form of a raster
representation. This way the following vectorization process
only deals with the empty outlines of the buildings.
The hatching is eliminated because it carries no additional
information to identify the location of the buildings. Fur-
thermore, the hatching introduces 'noise' in the vectorization
process and causes an unnecessary large amount of vector
data.
For the described recognition task different raster based tech-
niques are combined. Since mathematical morphology is one
of the most important ones, the following section gives a brief
introduction to this method. After that, the developed raster
based approach is explained and first results are shown.
3 MATHEMATICAL MORPHOLOGY
Mathematical morphology is the name of a specific collec-
tion of set theoretic operators defined on an infinite lattice,
see e.g. (Haralick/Shapiro, 1992). These operators, which
were first examined systematically by (Matheron, 1975) and
(Serra, 1982) in the 1960's are an extension of Minkowsky's
set theory. The operators are especially useful for image ana-
lysis and image enhancement. They represent an interesting
alternative to classical linear filter convolutions. In contrast
to the classical ones, the morphological operators are shape-
dependent and nonlinear image transforms. They can be de-
fined in binary or grayscale images for any number of dimen-
sions. Since the presented work is based on binary images,
the following explanations are limited to the topic of binary
mathematical morphology.
A morphological operator is governed by a small pseudo im-
age, a so called structuring element. When applied to an
image, the operator returns a quantitative measure of the
image's geometrical structure in terms of the structuring el-
ement. This measure can be used to decompose complex
shapes into their meaningful parts and separate them from
their extraneous parts.
The primary morphological operations are erosion and dila-
tion. The erosion — as its name already indicates — removes
pixels from border areas in the foreground region. In contrast,
the dilation adds pixels to border areas. The extent to which
the original shapes of the image are changed, depends on the
shape of the structuring element. The structuring element is
an operator mask of any shape. Frequently used structuring
elements are e.g. disk shaped masks.
The erosion follows the principle: If all pixels of a foreground
region in the original image I are covered by the structur-
ing element S, the considered central pixel will be set to 1
(foreground) in the resulting image. The erosion operation is
denoted by 1 © S.
Dilation, denoted by 7 & S, is the morphological opposite of
erosion: In case the structuring element covers at least one
pixel of a foreground area in the original image, the actual
pixel will be set to 1.
On the base of dilation and erosion the morphological opera-
tions of opening and closing can be composed. Opening con-
sists of an erosion followed by a dilation: ToS = (I©S)®S.
Vice versa closing is a dilation followed by an erosion: IeS =
(I ® S) o S. Opening mainly leads to smoothing of fringed
borders and the elimination of small areas. Closing can be
83
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
