E
MEE IE
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
model. The automatic creation of the semantic net requires
some constraints in order to be able to adapt the net to another
scale. These constraints are described in section 3. The second
input of the process is the target scale, which has to be smaller
than the start scale.
The first step is a decomposition of the semantic net. The goal
of the decomposition is to identify parts of the semantic net,
which can be processed separately. These parts can consist of
single object parts or blocks of them, if the object parts
influence each other during scale adaptation: E.g. two objects
with a small distance to each other possibly have to be adapted
together (depending on the target scale), because during the
scaling process the small distance can disappear and objects can
merge.
The next step is the scale adaptation of the decomposed parts
themselves. The selected object parts and blocks have to be
generalized. In this process different aspects have to be taken
into account: It is possible that the object type changes as well
as the object attributes. All objects of the same object type
exhibit a comparable behaviour in scale change and can be
extracted by the same group of feature extraction operators.
Scale Change
Model 1
Target Scale: 4
E f(x
A m
i-i = +
Y Scale Change i-um wu
, Decom- |. Model 2 dFusion: =. La Ti =
position | ns + = n 5
Object Model for
Large Scale Small Scale
Scale Change
1 Model n
ei.) >
Object Model for I
Figure 1. Strategy for Scale Adaptation
For the scale adaptation of the elements we intend to use scale
change models. Scale change models describe the kind of
change of a certain object type depending on the value of scale
change. Different attributes, such as the grey value contrast of
the object to the neighbours or the spatial measurements, are
used as input parameters for the scale change models. The
decision, of how to change the object parts or the blocks, has a
direct connection to the question of how they can be extracted
after scale change by using feature extraction operators. If it is
necessary to change the feature extraction operators after scale
change, the object type has changed. Hence, the scale behaviour
of feature extraction operators has to be taken into account. In
section 4 we describe first results of some examinations, aiming
to determine the suitable scale range for certain feature
extraction operators.
The last step is the fusion of the adapted object parts to a
complete semantic net, which describes the object in the smaller
scale and which can be used for an automatic object extraction
in low resolution images.
3. COMPOSITION OF SUITABLE SEMANTIC NETS
The knowledge representation in this approach uses the form of
the semantic nets of the knowledge based system AIDA
(Liedtke et al., 1997, Tónjes, 1999). This system was developed
for automatic interpretation of remote sensing images. Semantic
nets (Niemann et al., 1990) are directed acyclic graphs. They
consist of nodes and edges linking the nodes. The nodes
represent the objects expected in the scene while the edges or
links of the semantic net model the relations between these
objects. Attributes define the properties of nodes and edges.
Two classes of nodes are to be distinguished: the concepts are
generic models of the objects and the instances are realizations
of their corresponding concepts in the observed scene. Thus, the
knowledge base, which is defined prior to the image analysis, is
composed of concepts. During the interpretation a symbolic
scene description is generated consisting of instances. The
relations between the objects are described by edges or links of
the semantic net. Objects are composed of parts represented by
the part-of link. Thus, the detection of an object can be
simplified to the detection of its parts. The transformation of an
abstract description into its more concrete representation in the
data is modelled by the concrete-of relation, abbreviated con-of.
This relation allows for structuring the knowledge in different
conceptual layers, for example a scene layer and an image
layer.
The concept of semantic nets for the extraction of particular
objects enables many possibilities for the generation and
composition of a particular object model, i.e. the representation
of an object model for the extraction of an object can be
realized with different semantic nets. Based on the goal of this
research — adapting semantic nets automatically to a smaller
scale — it is necessary to find rules for the generation of
semantic nets.
The semantic nets to be created should satisfy the following
constraints:
e They should have a structure, which enables to
analyse them automatically regarding scale
behaviour. The structure should enable an automatic
decomposition of the semantic net into suitable parts
in order to treat them separately.
e They should describe the objects completely with all
characteristics and attributes, that are important. for
the extraction of the objects in the starting scale.
eo They have to be suitable for an automatic extraction
of the objects from digital images. The semantic net
should contain a refinement of the whole object into
suitable parts, which can be extracted directly by
using certain feature extraction operators.
e They should be easy to create by using standard
language. The rules should direct the experts in
creating the semantic nets, not complicate it.
Otherwise advantages of semantic nets and expert
systems, like the explicit knowledge representation,
would be lost.
In this stage of our research, we focus on roads and road
markings. As a further specification, the focus lies on objects
which run for a long distance parallel to the road axis along the
road. That means we deal with objects such as lane markings,
but not with objects such as zebra crossings. We are able to
describe such objects with the object types periodic and
continuous stripes and lines. The semantic nets, which we
create for such objects, contain only these four object types.
We represent roads first by describing the entire stripe (the
pavement) as the basis object and, as part of a road, the
markings on it (see Fig.6). As a rule, only the objects at the
bottom layer will be extracted. That means for Fig.7 the object
will be extracted by finding its line markings.
The nodes of the semantic net represent objects. We define for
every object of the semantic net the following attributes:
e Object Name: The description of the object by its
name in standard language.
eo Object Type: All objects of the same object type have
a comparable behaviour in scale change and can be
extracted by the same group of feature extraction
Interna
The re
part-of
additio
attribu
The at
regard
attribu
are im
border
consid
The sp
becaus
scale a
object
examir
has to
this de
descri
descrit
ranges
expres:
All ne
extract
objects
feature
Corres|
Fig.6 t
extract
given
extract
would
Based
object
object
