reference to the semantics of the descriptions (see
the paragraph on object based information fusion). An
example is given in [89] using relational description for
locating control points, derived from a topographic map.
In case no map data are available any type of seg
mentation and the corresponding relational description
will have to cope with the problem of observability of
complete segments: many object parts do not show in
image segments due to reason of lighting, common use,
common or similar reflection properties or weak radio
metric resolution of the sensor. Other problems are
caused by textures where boundaries are even more
unstable/uncertain. In case a generic model about the
region boundaries is available (e.g. land-use units have
polygon-like boundaries) this might be used to improve
segmentation and thus fusion of different information
sources.
Higher level aggregates
The containment hierarchies of most objects to be
extracted from images give rise to several levels of
grouping, lines and regions representing the lowest level.
Often such groupings can be defined without explicit
reference to the semantics of the objects. Examples are
field structures, ribbons of streets, symmetries of buil
dings, sequences of trees at avenues or in plantations.
The feasibility of such groupings has been shown in [64]
for matching and for building extraction. An example for
extracting the structure of land-use areas is given in [70].
It is essential that such structures can be derived without
explicit object model but of course could be guided by
such models. Obviously the boundary to techniques
which perform the fusion on the object level is not clear
cut, as it is the geometric and topological structure of the
objects which is mapped to the structure of the image
features.
Object based information fusion
The strength of fusing information on the signal and the
feature level results from the rigour of the used geo
metrical, physical and/or structural models. The fusion is
performed bottom up possibly leading to hypothesis of
high level aggregates. The semantic relations between
the parts of the objects in concern are only used impli
citly, e.g. when searching for antiparallel edges for road
extraction in case this search is part of the standard
feature extraction procedure.
In case the semantics of the objects, not only of their local
properties, is used explicitly for fusing information we
refer to object based fusion. We may distinguish different
knowledge sources, namely the human analyst, inter
preted map data and generic models. They obviously are
different with respect to flexibility, adaptability, auto
mation potential, quality or suitability. All of them may be
used within the analysis process either individually or
integrated.
In addition we therefore have to distinguish two cases:
there is only a single knowledge source containing the
object model. Then this model normally is used to
guide the image analysis top down;
there are several knowledge sources containing the
object model or parts of it. Then besides guiding the
image analysis it is necessary to fuse the parts of the
object model into one coherent model.
Both cases are discussed separately.
image analysis with a single object model
In case we only use a single object model as the basis for
image analysis we in a first instance can assume the
model defining the context of the analysis in a unique
way. This of course does not imply the resultant inter
pretation to be unique as e.g. sensor information may not
be sufficient.
Having a human analyst defining the context of the analy
sis still provides highest flexibility and adaptability when
integrating different information sources. This is due to
the background and common sense knowledge which is
used, especially when coping with long term monitoring
where explicit object models are not available, e.g. in
case political, social or historical knowledge is required,
or this knowledge is just too weak. Humans are also very
flexible in selecting analysis strategies, especially when
coping with incomplete or seemingly contradicting infor
mation. The coding and use of common sense knowledge
is a classical topic in Al but still waits for results appli
cable in practice [34 and 53],
The least flexible but useful and available object models
are coded in maps. We assume a map interpretation is
performed, either off-line in advance like in a GIS or
available during image analysis, such that the necessary
attributes and relations are accessible. Whereas the leg
end of a map contains information about object classes,
the map itself describes an instance of a complex object.
It does not contain generic information about topographic
features, say.
This special model, however, may already be very useful,
as sensor and map data are in some way complementary:
maps describe mainly geometric aspects of the objects,
including topological relations. They only give access to
the highest level of the specialization hierarchy and con
tain names, i.e. user readable keys which may provide
links to other information sources. On the other hand
sensor data provide up to date information about object
parts and in case of multi-temporal data time series
of object attributes which may directly be linked to the
objects in the map [43 and 44], Thus using maps as the
only knowledge source containing an object model may
very well be useful for monitoring in case the attributes of
well defined object (instances) have to be observed over
time.
The inflexibility of maps reveals when changes of the
objects occur, be it geometrical, physical, or semantic
changes, or when the context the map was made for is no
longer valid. Then a new map is required, at least par
tially. Thus the problem of map update cannot be solved
without reference to an object model containing semantic
relations between classes and their change over time,
which either requires a human analyst or generic models.
The term generic model here is used for object models
which are formalized statements about object classes
and possibly instances. Generic models thus describe the
objects, their possible attributes and their mutual rela
tions (see the paragraph on object models).
Typical representations of generic models are the entity-
relation-model [83] in the area of data bases, semantic
networks [68] and frames [61 in the area of Al, grammars
in formal languages [25] or object oriented techniques
[15]. Rules may be used to represent actions, changes
or other operations on the above mentioned represen
tations.
NGT GEODESIA 93 - 8
379