ND
n, fil-
scal-
Je &
Jres,
aliza-
opol-
ch is
veral
d as
es &
/eals
rma-
> im-
lirec-
ilues
oigh-
ency
tions
t ori-
od is
on of
data
tree
k ac-
1 ap-
cept
oach
r the
sec-
s he
1etric
r the
eral-
n the
actly
e re-
lines.
topo-
;t ob-
Japs,
the now obsolete areas have to be merged to neighboring
objects. The author introduces the concept of importance to
determine, which objects to drop and how to aggregate them.
The importance of an object is defined as a function of its size
and type. Thus the aggregation relies on spatial relationships
(proximity) and importance of the areal objects. In this way a
successive selection and aggregation of the objects can be
achieved which finally results in a object hierarchy. The most
general (important) object is on the top and the other objects
reside in the corresponding levels of the hierarchy. Opera-
tions defined on the tree data structure are the calculation of
the area and perimeter of any polygon in the hierarchy. This
approach proves to be very fast - which was the primary fo-
cus it was developed for.
The fact that only one (quite simple) rule is responsible for the
aggregation of the areas is considered a drawback. Also the
notion of importance seems to depend on certain parameters
which have to be tuned in advance.
The concept presented in this contribution aims at a deriva-
tion of the relevant rules and parameters directly from the
data - and thus independent of the user. He only is responsi-
ble for the control and for the provision of meaningful exam-
ples.
4 LEARNING RULES FOR RELATIONS
The approach bases on an object-oriented system named
FLAVOURS. FLAVOURS is embedded in a programming en-
vironment called POP11 (cf. [Barrett, Ramsay & Sloman
1985]) and is an implementation of the MIT Flavors package.
POP11 is interpreter-based which allows for a self-generation
of program-code. This feature is exploited to a great extent in
the prototype learning program, since the rules learned can
immediately be applied to the data and thus verified.
The basic assumption is that the rules are either complicated
or not easily formulated. Furthermore, deriving rules from the
data is easier and more reliable. Still the teacher is there to
finally verify and refuse or modify the rules if necessary.
4.1 Learning Structural Object Models
Image and map interpretation needs models in order to in-
terpret the visual information. The problem is to describe
appropriate object models - including objects attributes and
their relations. Sester [1995] presents an approach of using
Machine Learning techniques to derive a model description
from given examples. Learning techniques make use of the
fact that examples can be named and pointed at quite easily,
however it is often not known what the classifying attributes
are. So it is left to the learning procedure to determine them.
In order to define the learning task, the objects and their pos-
sible relations have to be identified. The actual manifesta-
tions of these relations will be learned based on the system
“vocabulary”, namely the properties and methods of the ob-
jects and relations, which is provided. These properties in-
clude geometric and topologic functionalities (e.g. size, form,
adjacency, relative position, ...). The teacher specifies the
concepts to be learned and selects the examples. The task
!The figure also visualizes the learning environment.
771
of the learning system is to determine the relevant properties
for a given concept and also their values.
In order to solve the task of interpreting given visual infor-
mation, the system starts with an identification of elementary
objects (namely polygons). For the further classification of
these objects, the system acquires adequate criteria in the
subsequent learning step. The basic idea of the system is
to control the process directly by the objects: each objects
checks, which methods it has available and whether they
can be applied to the given data - namely to the other ob-
jects. After all objects have applied their methods and no
more changes occur, the teacher can take over the control
by starting the learning procedure to acquire new objects or
a new object functionality. This in turn can extend the object
methods, which they apply to the data. Also there is the pos-
sibility of correcting and extending an automatically derived
description.
As an illustrative example consider the situation in the picture
on the right hand side of Figure 2'. Humans can immediately
recognize it as an extract of a map with fields and traffic ob-
jects. The system however in the first step distinguishes lines
and constructs polygons from it.
Eee 6 EE FI
Figure 2: Learning of concepts traffic and field
In order to learn discriminating criteria for individual object
classes the teacher points at different objects and gives a
classification. This leads to an automatic creation of new ob-
ject classes. The characterizing and classifying attributes of
the objects are derived by ID3. Thus after the learning step,
the systems knowledge has been extended: now the poly-
gons have a new method to apply, namely to differentiate into
the new object classes. A successive application of this pro-
cedure leads to the recognition of the objects given in Figure
3 (left), namely fields (feld), streets (strasse) and cycle tracks
(radweg).
Learning relations allows for the discrimination of so-called
part-of-relations and associations. In order to learn relations,
two objects are pointed at, and the teacher indicates whether
the relation is valid for these objects or not.
A part-of-relation results in the aggregation of objects that
share the learned relation. Associations have the effect of in-
cluding more context into the object description. An example
for the learning of such a relation is the association between
a cycle track and a neighboring street. The teacher points
at examples and counter examples for the relation which are
stored in an attribute-value list. The first line is a comment
line describing the concept to be learned, followed by the at-
tributes; each example with its corresponding attribute values
is stored in the subsequent lists.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
