interpretation using the generic model. In a practical 
situation, this is performed testing primitives in the 
image by scanning a set of rules involving properties 
and relations of object parts. 
Several subtopics have been identified when 
automizing the process of image interpretation: 
* Specification of generic models which define the 
objects and thus limit what can be and is to be 
described. 
* Segmentation of the image according to the generic 
models in order to obtain a well defined set of 
primitives for future parsing. 
* Parsing the segmentation output in terms of object 
types given by the generic models and recon- 
struction of objects in 3-D space using this parse. 
* Least squares adjustment of the obtained parse to 
the original grey level image or the segmentation 
when quantitative object description and 
localization is desired. Description of the global, 
regional and local quality of the image as well as the 
performance of image operators such as 
segmentation, when run on the images used. 
The description of complex objects present in digital 
images is thus approached using a procedure with 
several steps. In the first step, a segmentation 
procedure is run on the raw image data producing a 
set of closed polygons with interior descriptions. This 
procedure presents the segment borders in a language 
that is intelligable to a generic model for the objects to 
be described. In the case investigated here, the objects 
are buildings with the borders between roofs and walls 
assumed to be straight line segments, this making the 
output set of closed polygons and other geometrically 
simple figures in the segmentation procedure a 
natural choice. In a second step, the segmentation 
output from the first step is interpreted as an element 
in object space using a parser. Quantitative 
information on the images can be obtained from the 
parse. If necessary, high quality information as well as 
quality aspects can be assessed in a third step. A more 
detailed elaboration of the second step is given in 
(Gülch, 1992) and of the third step in (Zielinski, 1992). 
3 IMAGE REPRESENTATION USING 
SEGMENTATION 
Due to the methods existing today for acquisition of 
digital images, these are almost invariably represented 
by giving grey levels in a regular grid. Here, a strategy 
is proposed where the image representation is goal 
dependent, i.e. the image is represented in terms of 
primitives which are suitable for describing the objects 
possibly contained in the image. 
Limiting the scope to objects in aerial images, the 
objects can be described in terms of specified parts in 
order to be able to keep to a limited set of primitives. 
The description of these primitives as well as their 
possible internal relations are given by a generic 
model. The representation of the image, chosen to 
suit to the parser, is given in terms of the same 
primitives. In digital aerial images the expected objects 
728 
are buildings, forests, fields, lakes, etc. These are all 
describable in terms of internally homogeneous 
regions with boundaries made up of straight lines, 
smooth curves and a limited set of corners. A suitable 
image representation is then given in terms of such 
regions together with a stochastic model for 
description of texture. This representation is in 
principle complete, i.e. the grey level image can be 
reconstructed from it, disregarding white noise. This 
kind of representation will here be called a 
segmentation. 
There is an abundance of different methods to 
segment digital images including edge closure, region 
growing and other methods. Many methods are of the 
ad hoc type leaving the user with no information 
about the quality of the result. No method is known 
to us where possible region boundaries and other 
properties are restricted by object related model 
requirements. In a rigorous approach, the 
segmentation should be performed under such 
restrictions. Segmentation procedures using the 
principle of minimum description length (Leclerc, 
1989), (Dengler, 1991) have recently been developed 
and appear quite suitable for this task. Region 
boundaries could be described using strip trees, which 
also give the description length. Such a procedure is 
under development. Until results are available, an 
existing procedure using region growing is used. The 
boundaries of the resulting segmentation are 
described using strip trees as described in section 3.1. 
Having passed the image through such a segmen- 
tation procedure, the data set to be parsed consists of a 
set of closed polygons made up of straight line or 
curve segments, a description of segment interiors in 
terms of trends and texture parameters and finally a 
window coarsely defining what part of the 
segmentation is to be considered. 
3.1 Boundary descriptions using strip trees 
The boundaries sought for in aerial images are 
assumed to be smooth with a limited number of 
corners. Linear segments are used to approximate this 
case. Strip trees are used to describe these boundaries 
following a method developed by (Ballard, 1981). 
Letting a strip tree represent the boundary to a 
neighbour, each region segment refers to K strip trees. 
The following notation is chosen to represent the 
region boundary, see also figure 1: 
a number of strip trees in boundary (=K). 
ax k-1,..,K. Pointers to the root strips S, of the 
K strip trees generating region boundary. 
Sk 7 (Xp, Xe, Wy, Wy, y P) Strip descriptions. 
Xp, Xe coordinates of start and end points for strip. 
Wy W, leftand right width of strip. 
W = W+W, 
PvPr pointers to left and right substrip. 
The 
ima 
infc