6.4 Parcel generation with model
In order to evaluate the model visually, new parcels are gen-
erated with the help of the model. Given a bigger parcel, it is
divided into smaller ones with rules of the grammar. Start-
ing at hierarchy level 1, the number of subparcels is taken
from the functional relation table: n = n(level = 1) = 3; the
associated probability is P(3|1) = 1.0. Thus the parcel is cut
into three subparcels. Formally this results in a structure:
P(part)
S
PARCEL PARCEL — PARCEL — PARCEL
The production rules are now applied to the subparcels cor-
respondingly. Being in level 2 now, the number of successors
is either 5 (with probability 0.67) or 2 (with lower probability
0.33). In the example, the color indicates the probability of
the partition: bright parcels have high, dark ones low prob-
ability. The first partition (Figure 4) - which shows clear
similarities to the example (see Figure 1), is more probable
than the second one (Figure 5).
Figure 4: Aggregation structure generated by model - high
probability
7 FINAL REMARKS
A prerequisite for any interpretation of images is to have
adequate model descriptions for the expected objects in the
images. As these are normally quite complex, there is a
need for automatic extraction or generation of such models.
In general, although examples for an object can be identified
and enumerated, they can not be described in a compact
form. Learning techniques are a means to solve this problem:
they allow for making a interior structure (hidden in the
862
examples) explicit. Which technique to chose, however, is
dependent on the type of object.
The approach to automatic model acquisition from examples
given in the paper basically is a general one, since it generates
a structural representation of the data. Such descriptions,
specifying an object in terms of parts and relationships are
useful for high level image interpretation tasks, dealing with
complex real world data.
The result of the modelling is not only a description of the
object, but the model also reflects the statistic in it. This
opens the way to a compound analysis of data of different
knowledge sources basing on the MDL principle.
The paper showed that it is possible to extract a representa-
tion for parcel structures from examples. The new represen-
tation gained from the original line representation is more
compact, of generic type and thus more suitable for subse-
quent object location or recognition tasks.
LITERATURE
Cohen, P.R. and E.A. Feigenbaum [1982]: The Handbook
of Artificial Intelligence, Vol. 3. Kaufmann Inc., Los
Altos, California, 1982.
Connell, J.H. and M. Brady [1985]: Learning Shape De-
scriptions. Proceedings of the IJCAI pp. 922-925, 1985.
Fu, K.-S. [1982]: Syntactic Pattern Recognition and Appli-
cations. Prentice Hall, New Jersey, 1982.
Janssen L.L.F and H. Middelkoop [1991]: Knowledge Based
Image Classification; the use of knowledge about crop
rotations in a Thematic Mapper Image Classification
Procedure. Report 40, The Winand Staring Centre,
Wageningen, The Netherlands, 1991.
Michalski, R.S. [1984]: A Theory and Methodology of In-
ductive Learning. In: Machine Learning, Michalski,
R.S. et al. (ed.), Springer-Verlag, Berlin, 1984.
Niemann, H. [1981]: Pattern Analysis. Springer Verlag,
Berlin, 1981.
Pan, H.-P. and W. Forstner [1992]: Stochastic Polygon Map
Grammars: A Generic Model for Understanding Lan-
duse Maps and Images in Remote Sensing. In: Robust
Computer Vision, Wichmann, 1992.
Ripley, B.D. [1981]: Spatial Statistics. Wiley, New York,
1981.
Rumelhard, D.E. and J.L. McClelland [1986]: Parallel Dis-
tributed Processing: Explorations in the Microstructure
of Cognition. MIT Press, Cambridge, 1986.
Segen, J. [1988]: Learning Graph Models of Shape. Proceed-
ings 5th International Conference on Machine Learning,
pp. 29-35, 1988.
Simon, H.A. [1984]: Why should Machines learn ? In: Ma-
chine Learning, Michalski, R.S. et al. (ed.), Springer-
Verlag, Berlin, 1984.
Stier, S. [1991]: Bildverstehen als Lernprozefi aus ele-
mentarem Wissen, PhD thesis, Universitit Hamburg,
1991.
