e concepts are
re realizations
ene. Thus, the
ige analysis, is
n a symbolic
nstances. The
ges or links of
'epresented by
bject can be
yrmation of an
entation in the
viated con-of.
ze in different
ind an image
| of particular
neration and
representation
)bject can be
1e goal of this
to a smaller
generation of
the following
h enables to
rding scale
an automatic
suitable parts
letely with all
important. for
1g scale.
tic extraction
semantic net
le object into
] directly by
E
sing standard
e experts in
omplicate it.
s and expert
epresentation,
ids and road
es on objects
xis along the
ine markings,
e are able to
periodic and
s, which we
ect types.
re stripe (the
a road, the
)bjects at the
2.7 the object
Ve define for
butes:
object by its
ect type have
e and can be
re extraction
Istanbul 2004
International Archives of the Photograninctry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
operators. For the description of streets we define four
object types: continuous stripe, periodic stripe,
continuous line and periodic line. This also affects the
group of feature extraction operators, which are used
to extract the object type. The difference between
stripe and line is given by the width of the object. We
define that lines are not wider than 2 pixels. For the
extraction of lines in raster images other feature
extraction operators will be used than for the
extraction of stripes. Stripes would be extracted by
finding the edges. The specification periodic or
continuous indicates dashed or continuous lines and
stripes respectively.
e Grey Value: This attribute describes the radiometric
characteristic of the object.
e Extent: This value describes the width of the object
and is independent from resolution, i.e. is stated in
meters, not in pixels.
e Periodicity: For periodic object types this attribute
expresses the ratio of the length of the lines/stripes
and the extent. For continuous object types this
attribute has no meaning.
The relations, which are used in the presented net in Fig.6, are
part-of and spatial relations. Some of the part-of relations have
additional attributes, as “central” or “left/right boundary". The
attribute “central” can be used to guide the object extraction.
The attribute “left/right boundary” has an important function
regarding the scale adaptation: part-of relations with these
attributes describe the boundary parts of an object. These parts
are important, if neighbouring objects exist. In that case the
border parts of both neighbouring objects have to be
considered, because they can affect cach other as scale varies.
The spatial relations play a key role in the object description,
because they directly affeet the necessary modifications for
scale adaptation of the nets. It is essential that the position of all
object parts are clearly specified by spatial relations. For the
examined objects, the position perpendicular to the street axis
has to be described. We use the relations left-of and right-of for
this description. Furthermore, attributes are important for the
description of the distances. Usually, it is not possible to
describe the exact distance to another object. We therefore use
ranges for distances here. The magnitude of the distances is also
expressed in meters, independently from scale.
All nodes of the semantic net are connected to feature
extraction operators, which are able to extract the represented
objects. But the strategy of object extraction is to call the
feature extraction operators only for the nodes without parts,
corresponding to the nodes at the bottom of the semantic net. In
Fig.6 the node “Roadway” contains a connection to a feature
extraction operator, which is able to extract stripes with the
given constraints. But as long as markings on the stripe are
extractable in the target scale, the extraction of “Roadway”
would be realized by the extraction of the markings.
Based on these semantic nets the scale behaviour of the defined
object types can be investigated. Taking into account single
object parts, the following behaviour is possible:
Unfortunately, these four possibilities are not sufficient for an
automatic scale adaptation of the semantic nets. It is also
possible that different parts affect each other during scale
change. As an example, two stripes with a small distance apart
will merge at a certain scale. Hence, neighbouring parts have to
be analysed simultaneously in this case.
Taking into account object pairs, which might affect each other,
the following possibilities can be found:
Nr | Before Scale Change
5 Cont. Stripe — any
6 Per. Stripe — Per. Stripe
Per. Stripe — Cont. Line
Per. Stripe — Per. Line
7 Cont. Line — Cont. Line
Cont. Line — Per. Line
8 Per. Line — Per. Line PL or CL or Invisible
After Scale Change
CS or CL or Invisible
PS or CS or PL or
CL or Invisible
CL or Invisible
Nr | Before Scale Change After Scale Change
Continuous Stripe (CS) | CS or CL or Invisible
Continuous Line (CL) CL or Invisible
WIN —
Periodic Stripe (PS) PS or CS or PL or CL or
Invisible
4 | Periodic Line (PL) PL or CL or Invisible
Table 1. Object Type Scale Behaviour of Single Objects
Table 2. Object Type Scale Behaviour of Object Pairs
This scale behaviour of the object types has to be used for the
adaptation of the semantic concept nets. It is possible, that a
given range in a concept net for a distance between two objects
will lead in combination with a given target scale after scale
adaptation to different possibilities for the new object type. In
that case different possibilities have to be included in the new
concept net as representation of one object.
The question, at which target scale the object type changes is
directly connected to the scale behaviour of feature extraction
operators. This problem is addressed in the next section.
4. SCALE BEHAVIOR OF FEATURE EXTRACTION
OPERATORS
As described in section 3, the object type of a node in a
semantic net is also determined by the feature extraction
operator, which is bounded to the nodes of a certain object type
and is used for its extraction. Objects of different types use
different feature extraction operators. But as scale varies the
object type may change, because the same operator is not able
to extract the same object type successfully at all resolutions. In
order to be able to predict from which scale on the object type
has to change, an analysis is necessary about the scale range, in
which the feature extraction operators are usable. The
performance of three commonly used higher developed
operators for edge and line detection are exemplarily examined
- Canny, Deriche and Steger. The goal of this investigation is to
analyse the behaviour of the operators in sensor data of
different resolutions. For that in a first step the different sensors
are simulated by creating synthetic images of different
resolutions and applying the operators on them.
The Canny edge detector was developed as the “optimal” edge
detector (Canny, 1986). Its impulse response shape closely
resembles the first derivative of the Gaussian. The Deriche edge
detector is based on the Canny operator, but uses recursive
filtering and thereby reducing computational effort (Deriche,
1987). The Steger operator extracts lines in sub-pixel accuracy
by using the first and second order derivative of the Gaussian
(Steger, 1998).
The generated synthetic grey value image has a size of
1000x1000 pixels and also is composed of a bright line with
two pixel width stretching over the entire image on dark
background. An image pyramid was created from this synthetic
image by Gaussian interpolation. The pyramid comprises 100
levels from the original image (labelled as pixel size 1.0) to the
smallest image with the largest pixel size of 1000-fold,
