AUTOMATIC MODEL ACQUISITION BY LEARNING *
Monika Sester
Institute of Photogrammetry
Stuttgart University
Keplerstraße 11, D-7000 Stuttgart 1
Commission III
Abstract
Recognition presumes having a model of what to recognize.
This especially holds true for the recognition of objects in
digital images. Such a model is usually formulated explic-
itly by humans. With the help of techniques from Machine
Learning however, it is possible to automatically construct
models from given examples.
The paper reviews several learning techniques and focuses
on the automatic model construction with formal grammars.
Both the requirements and the potential of such techniques
are demonstrated with an application in the domain of lan-
duse classification. A model for an agricultural parcel struc-
ture is used as one component in a system to recover land
use maps from remotely sensed data.
Keywords: Artificial Intelligence, Machine Learning, Im-
age Interpretation
1 INTRODUCTION
In order to derive landuse information from airborne images
usually multispectral classification is used. But there is more
than radiometric data in the images: both texture and geom-
etry contain information about different landuse types. Each
of these information sources needs a model to extract the rel-
evant knowledge. This paper concentrates on the analysis of
geometric models.
As shown by Janssen et al. [1991] using maps for classifica-
tion greatly improves the classification result. When however
no maps are available, a model for the parcel aggregation
has to be provided. Such a model has to be a very general
description, since it is impossible to represent any kind of
possible parcel aggregation. Thus there is a request for a
so-called generic model, where not only the object param-
eters but also the structure is free to a certain degree. An
object parameter in the case of the parcel aggregation struc-
ture is e.g. the size of an individual parcel; the structure is
reflecting the relations among the object parts (the number
of neighbors of a parcel). , Polygon“ is a generic description
for a parcel, in contrast to a list of n coordinates of the n
points of the polygon.
Normally the models are formulated explicitly by humans.
This is adequate as long as the objects are clearly definable
and have distinct features. Often a collection of prototype
*This research is supported by the Deutsche Forschungsgemeinschaft
in project SFB 228 „High Precision Navigation“
856
objects is available, but still it is not clear in advance, which
are the relevant parts of the object and which are its features
and relations. In the terminology of knowledge representa-
tion the examples denote the extensional description of the
objects. The task of Machine Learning techniques is to make
this implicit knowledge explicit, thus end up in an intensional
description.
In the paper the special problem of the parcel aggregation
structure is analyzed and a strategy to extract a parcel model
from examples is presented. The prerequisites for the auto-
matic model acquisition are briefly sketched: given the exam-
ples, the internal structure of the data has to be extracted.
The structure is revealed in a clustering process by grouping
objects which are similar in some sense. The resulting graph
is represented with the help of formal grammars, where each
node is coded by a grammar rule. In order to give respect to
the possible variety of the structure and also to noise effects,
each node is now considered as the outcome of a random
experiment. In a subsequent statistical analysis the node
parameters are estimated, and so the statistic inherent in
this structure is revealed.
The theoretical background thus lies on Machine Learn-
ing techniques, Knowledge Representation and Spatial Pro-
cesses. After an overview of the project to which this con-
tribution belongs, a short review of Machine Learning tech-
niques is given, with special focus on model acquisition for
Computer Vision purposes. The subsequent section is con-
cerned with statistics. Finally an example for the model
extraction from examples is given and the feasibility of this
model is demonstrated.
2 KNOWLEDGE BASED
LANDUSE CLASSIFICATION
Knowledge based image interpretation is performed by us-
ing any kind of information source available. A program to
extract landuse information from aerial images can base on
radiometric information in a traditional multispectral classi-
fication, but also on information about the geometry of ob-
jects. Thus not only mere grayvalues, but furthermore struc-
tural information about the objects is made use of. In order
to integrate different sources of knowledge, the Minimum De-
scription Length-principle (MDL) can be applied. MDL pri-
marily allows to treat structural and numerical pieces of in-
formation within one compound process. In [Pan and Forst-
ner 1992] a strategy for this task is presented.
Applying MDL presumes knowledge about the probability of
the influencing factors. Knowing the whole functional chain
