I
I
modeling
world model
Instance
symbolic
world
description
real world
sub8ymbolic
analysis
model
image
model
interpretation
sensor
model
object
model
modeI
analysis
description
interpretation
image
description
sensor
description
object
description
description
analysis
interpretation
object
image
sensor
Fig. 4.
The different compo
nents of a system for
image interpretation.
The models (upper
row) give rise to instan
ces, i.e. descriptions
(middle row) of the ob
jects and processes in
the real world (tower
row).
human analyst
which allows to construct new powerful models at every
level, at the expense of weak possibilities of formali
zation.
The following generic model for the interpretation process
wants to clarify the role of the different components of an
image interpretation scheme. The image analysis part will
be discussed in details below.
Generic model for image interpretation
The components of image interpretation are shown in
fig. 4. There are three levels of description:
real world, which is to be described. This level is
sub-symbolic, in the sense that it is independent on
whether we describe it by some means or not. Of
course what is shown in fig. 2 is already a symbolic
description on a meta-level as it describes the struc
ture of the used models;
symbolic description of the real world. Such descrip
tions are fixed numbers, symbols, relations, etc. which
within a certain context defined by the models have an
exact interpretation i.e. relation to certain objects in
the real world.
models of the real world. The symbolic descriptions
are instances of models of the real world. The models
thus can be seen to give the generic structure of the
symbolic descriptions.
The observation process which is part of the loop of
surveying and mapping activities discussed above now
consists of the following three steps:
1. data taken with suitable sensors are collected in
images. The design of sensors usually is based on
some kind of model, especially with some application
in mind. On the other hand especially in remote
sensing often an analyst may be confronted with
sensors or sensor data not specifically designed for
his/her application, moreover where the description of
the sensor may not sufficiently well be described. In
this case the models of the sensor designer and the
models of the analyst may not be coherent causing
3) The notion „interpretation" is ambiguous: meaning both, the
process of labelling as well as the result of this process. We use
the „analysis" to name the process, though it suggests a too
narrow and one sided process seemingly omitting the larger
context. We use „interpretation" for denoting the result of the
analysis process, however, sometimes refer to both, the proce
dures as well as the whole analysis as „interpretation process".
large efforts to find out the value of the sensor data for
the specific task in concern.
The images in all cases are analogue representations
(in contrast to symbolic representations, see below),
as the geometric relations between the objects are
implicitly contained in the image and no whatsoever
interpretation, i.e. labelling takes place.
2. the image data are processed to obtain a symbolic
image description. Symbolic here stands for „symbo
lic representation" [76] in contrast to the analogue
(raster) representation of the sensor data. Any pro
cedure for image processing or pattern recognition,
often termed segmentation in fig. 4, may be used for
feature extraction. In principle the symbolic descrip
tion completely replaces the image. The structure and
the complexity of the symbolic description depends on
the image model which has to be derived from the
object model and the sensor model. Therefore the
image model is governed by the up to now not
specified object model and the usually comparably
simple sensor model.
3. the central part of the analysis consists of labelling
the symbolic image description yielding an interpre
tation of the image3). The analysis process usually is
hidden, both, when performed by a human or a
machine, unless the process itself is analyzed. This
analysis yields a description of the image analysis
process whose structure is defined by the analysis
model. This model, in addition to the object model,
and the sensor model depends on the application
defined by the user and the regularities which the
developer or analyst assumes the system to follow.
Modelling the observation process thus consists of
setting up the five models of fig. 4 and their mutual
relations, namely the:
object model;
sensor model;
image model;
analysis model;
interpretation model.
Object models
Object models contain geometrical, physical, biological,
structural, semantical and possibly other elements. Rela
tions between objects may be modeled by specialisation
(is-a) or containment (is-part-of) hierarchies. Other rela
tionships refer to time, or to causality between objects.
Object models must be object centred in order to be
invariant to the observation process used to infer the
presence, form, class etc. of objects. This can only be
374
NGT GEODESIA 93 - 8
