3.2.2 Object Relations: Knowledge about structures can be
represented efficiently by semantic nets. Semantic nets con-
sist of nodes and edges in—between. Here the nodes of the se-
mantic net represent the scene objects or their sensor specific
realization respectively. The nodes are implemented as
frames. The edges or links of the semantic net form the rela-
tions between the objects. Different relations describe the de-
composition of an object into its parts (part—of), the special-
ization (is—a), and concrete realizations in the image data
(con—of). The relations are exploited for object recognition.
The part— of relation states that the object is composed of
parts. Thus object search can be reduced to a more simple
task, the detection of its components. Objects linked via
cdpart— of appear only in a certain context. Thus these ob-
jects are only searched for when the context, i.e. superior ob-
ject, has been detected. Finally the optpart—of relation
points out objects that might be present.
Objects can often be detected based on their geometric or
photometric appearance, that can directly be segmented in
the image data. This transformation of an abstract concept to
a concrete realization is represented by the concrete — of link.
The is—a relation describes a specialization of an object. The
specialization inherits automatically all relations and attrib-
utes of its more general concept.
The instance —of relation is used durung interpretation and
connects instances with their prototypes.
3.2.3 Sensors: Sensors like cameras project the 3D objects onto a
two dimensional target. They are sensitive to certain wave-
lengths. The different radiometric and surface properties of the
materials are mirrored by corresponding colours and textures in
the sensor image. E.g. the asphalt of roads appears bright in the
visual spectrum and dark in SAR (synthetic aperture radar)
images. In the semantic net the sensor transformation is modelled
by the con—of relation (fig. 3). Propagate methods restrain the ex-
pected range of attributes top down. Compute methods obtain the
measured value bottom-up from the sensor. To model uncertain-
ties the attributes are described by minimum and maximum val-
ues,
3D— Geometry
Layer
Asphalt 3D —Stripe
Width [m]:
Range: 5 ...10
Value: 6...8 LS
-—
! Compute
Fig. 3: Representation of sensor transform characteristics
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
3.2.4 Strategy: Strategy knowledge is represented by rules.
These rules exploit the knowledge represented in the seman-
tic net to control interpretation. For the relations exist ac-
cording rules which propagate new information over the links
of the semantic net. A rule is composed of a condition and an
action part. The condition checks for a new interpretation
state of neighboured nodes in the semantic net. If a situation
formulated in the condition is detected, the action is executed
to adapt the interpretation state of the focussed node accord-
ingly. The knowledge, that an object is detected when all its
parts n; € are detected, i.e. are complete instances, is repre-
sented by following rule:
CONDITION: If state (node nj) = complete instance
Vn; EP P ={n;| n; = part—of(no)}
ACTION: Then state (node no) = complete instance.
Different strategies are represented by various sets of rules.
3.3 Knowledge for Landscape Modelling
Figure 4 shows a simplified semantic net for landscape mod-
elling. The knowledge base distinguishes three conceptual
layers. The top layer, called scene layer, describes the scene
specific semantic. The middle layer represents the objects
based on their 3D —geometry and material. The bottom layer
is sensor related and describes the sensor specific photomet-
ric and geometric appearance of the objects. If more than one
sensor is present the sensor layer is multiplied accordingly.
Each layer uses a common appropriate vocabulary. E.g. the
attribute size is measured in meter at the 3D — geometry layer
and in pixel at the sensor layer (fig. 3).
In the context of landscape modelling roads, forests, and
grassland shall be distinguished. The forest is composed of a
forest roof and a forest edge which have to be modelled sepa-
rately. For recognition only the forest roof has to be visible in
the image data. Roads appear as homogenous stripes in the
aerial images. The initial concepts 'textured 2D —region' and
'"homogenous 2D —stripe' posses methods for segmentation
of textured regions and homogenous stripes respectively.
4. INTERPRETATION
4.1 Control
The aim of the interpretation is to match the objects of the
analyzed scene with the corresponding nodes of the semantic
prototype net. Image interpretation exploits the knowledge
base to instantiate hypotheses of objects expected in the
scene. According to the state of the object recognition three
different types of instances are distinguished: hypotheses,
partial instances and complete instances. Hypotheses are not
yet verified in the sensor data. Partial instances contain all
concretes and context independent parts. Complete
instances possess all concretes and obligatory parts and con-
text dependent parts.