(with a certain
be verified. If
] to these non-
symbolic level
cessing general
out the topo-
s system must
dge. Different
own like pred-
ics or semantic
à schemes con-
and modelling
and the actual
EST (Erlanger
is work ((NIE-
93)).
ge: declarative
vledge consists
ledge contains
ncepts as well
cial algorithm,
fig. 11).
image arca
|
lysis
jt the scene to
t step two SN
ut the database
generic models
d on a compar-
ural knowledge
ge still present.
itabase domain
e model. This
nsformed to the
image domain. Combining it with the generic model 'image'
automatically a specialized SN for analysis of the scene is
created. Now the image analysis based on this model (spe-
cialized to the actual scene) is carried out (instantiation). A
description of the scene in image domain together with the
specialized model for the processing of the concerning scene
will be the result of image analysis. Verification and classifica-
tion is not the comparison of two processes executed parallel,
but the result of the more error-tolerant analysis procedure
(database analysis).
4.2 Data Analysis
The main topic of analysis process will be the verification and
classification of image segments (procedural knowledge pro-
cessing). To achieve the above mentioned structured storage
of data the symbolic image information was introduced to a
SN. An overview of the SN 'image' is shown in fig. 12, where
the different components of the SN can be seen. Beside of
this image information — declarative knowledge realized by
concepts and links — also the results of feature extraction
(additional procedural knowledge) have to be introduced in
the SN to enable a successful verification of all image objects
based on the 'learned' information.
This analysis is realized inside the SN by creation of suitable
valuation- and analysis-functions. The procedural knowledge
for analysis of the generic models consists of functions for
determination of attributes and parameters of concepts as
well as for valuation of the achieved attributes, of the links
and the structure relations, where the database (DLM200)
acts as an additional information. This functions are different
for the database and image domain (resp. for the diverse
objects). Therefore, they are determined process-specific.
KOT
= | Wasser ]
ne
Gruenflaeche
Kul cuxrw
zer meme
Sondexrflaeche
Verkehisftl
Y |
E
Fluss
Ss __uenstliach
Sondexrkul cure
d. Strasse
Ueber bau tt
Un” ers iegelt
Wiese
Pict
Landwirtschaft
«XN.
<
Ne
Mischwald
Mo ox
Nadelwalda
Natuex lich
Rasen
S_mnmatuerliaach
Schiene
Str u oc bs tt
Versiegel ©
Weiner oo
Wohnflaeche
Zierteich
[9] SPEZ
Figure 12: Overview (components) of SN model ’image’
Fig. 13 shows the description graph of the SN 'image'. The
connections can be recognized by the the diverse links. On
757
the lowest level the SN is conected to the concrete data,
which are stored as segment polygons.
(KON, BST)
0,0]
(2,0)
Sonderflaeche
(1,0) ; kuenstlich
(0,2)
(0,1)
(0,0) ÿ |S_natuerlich
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Figure 13: Description graph of SN 'image'
5 CONCLUSION
First experiences with an extented feature base and a spe-
cial segmentation confirm the efficiency of this concept by
achieving a better separability of object classes. The system-
atic structuring of knowledge in semantic networks is a basic
precondition for this integrated Knowledge processing. But
the creation of concrete SN has proved to be a very com-
plex and time consuming task. Therefore, specific software
has to be developed for automation of this procedure. In the
next phase of realizing our concept of semantic modelling the
determination of suitable valuation functions in the decision
process and the treatment of uncertainty will be the most
important aspect.
REFERENCES
F. Kummert, H. Niemann, R. Prechtel, G. Sagerer: Control
and explanation in a signal understanding environment.
Signal Processing, 3, 1993, S. 111-145
H. Niemann, G. Sagerer, S. Schróder: ERNEST: A semantic
network system for pattern understanding. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
12, 1990, S. 257-269
T. Voegtle, K.-J. Schilling: Wissensbasierte Extraktion
von Siedlungsbereichen in der Satellitenbildanalyse.
Zeitschrift für Photogrammetrie und Fernerkundung,
Heft 5/95, 63. Jahrgang, Sept. 1995, Wichmann Verlag,
S.199-207
K.-J. Schilling, T. Voegtle, P. MüBig: Knowledge based anal-
ysis of satellite images. ISPRS Comm. Ill Symposium:
Spatial Information from Digital Photogrammetry and
Computer Vision, Munich, 5.-9. Sept. 1994, Proceed-
ings p.732-736
