3. Istanbul 2004 
-action of urban 
y, International 
58: 83-98. 
id roads from 
mation, G/S, 6: 
n of main-road 
igery based on 
al Conference 
nerican Society 
Yorage, Alaska), 
Extraction from 
imizing Based 
Vol 58 (1), 41- 
rain models, 
D dissertation, 
ty of Calgary, 
'ses. html. 
melis, 2001. A 
ction of Linear 
ansactions on 
9. 
, C. Steger, and 
of roads from 
Machine Vision 
. Assessing the 
r intensity data, 
near structures, 
and Machine 
ze based road 
nal Archives of 
):387-393. 
road extraction 
(4):215-224. 
icing by profile 
n-made Object 
A. Gruen, +O. 
erlag, Basel, pp. 
iowledge-based 
ian Journal of 
AUTOMATIC SCALE ADAPTATION OF SEMANTIC NETS 
K. Pakzad, J. Heller 
Institute of Photogrammetry and Geolnformation, University of Hannover 
Nienburger Str. 1, 30167 Hannover, Germany — (pakzad,heller)@ipi.uni-hannover.de 
Commission III, WG III/4 
KEY WORDS: Interpretation, Expert System, Knowledge Base, Model, Scale, Representation, Multiresolution, Generalization 
ABSTRACT: 
This paper deals with a methodology to derive object models for automatic object extraction in low resolution images from models 
created manually for high resolution images. The object models are represented by semantic nets, which describe landscape objects 
explicitly in terms of natural language. Starting from semantic nets for high resolution images the strategy is to first decompose them 
into parts, which can be handled autonomously. The object parts are then adapted, i.e. generalised, to smaller scale. The adaptation 
takes into account the object shape, radiometry, and texture. For the generalisation process “scale change models" are used, which 
describe how different types of objects evolve over scale mathematically. Finally, all object parts are fused and transferred to a 
semantic net representation. In this paper first results of the described methodology are presented. Focussing on line-type objects, 
such as streets, we describe how to create an object description with semantic nets using constraints, which have to be satisfied, in 
order to be able to adapt the nets to other scales automatically. In addition we show tests of the behaviour of some edge- and line- 
extraction operators through scale space. These tests are necessary to predict the scale behaviour of different object types. At last, we 
describe as an example for a particular object events during scale change observed in an image and their impact on a semantic net. 
This example demonstrates the suitability of the proposed kind of semantic net to follow the scale space events in digital images, and 
thus, its applicability in an automatic approach. 
1. INTRODUCTION 
Landscape objects appear differently in remote sensing images 
of differing resolution. While many object details are visible in 
high resolution images, in low resolution images many of them 
disappear or merge. Even the dimensionality can change. 
Where in high resolution images areas are observable, in low 
resolution images lines or even points might be found. This fact 
also affects an automatic extraction of landscape objects from 
digital images with different resolutions. For an automatic 
extraction from satellite and aerial images knowledge-based 
systems with an explicit knowledge representation, such as 
semantic nets, offer high flexibility and can easily be structured 
(Pakzad, 2001). This knowledge representation contains the 
object models, which describe the objects with all relevant parts 
and characteristics. As described above the models for the same 
objects have to be different depending on the resolution of the 
images. They are tailored to specific scales of aerial and 
satellite images. Decision about the best scale for object 
detection is mostly still made intuitively (Schiewe, 2003). In 
(Baumgartner, 2003), the representation of roads in a small and 
a large image scale is combined in a semantic net. However, the 
fusion of the two scales is solely used for increasing the 
reliability of the extraction results. 
Existing approaches for explicit object models do not permit an 
automatic transfer to other scales. Hence, a new model is to be 
developed for each image scale manually. For the case of scale 
reduction, a description of object behaviour is possible though, 
as investigations of features in scale-space indicate (Witkin, 
1986). The scale-space theory was formulated for a multi-scale 
representation of objects, depending only on one parameter for 
scale. Following this theory, with increasing scale parameter, 
ie. lower spatial resolution, new details will not appear, but 
existent details will disappear and merge with each other 
(Lindeberg, 1994). The object representation in image data of 
lower spatial resolution can therefore be predicted starting from 
high resolution. A methodology for an automatic adaptation of 
object models to lower spatial resolutions would make the 
manual generation of these object models for different 
resolutions redundant. Thus, a once created object model could 
be utilised for a wider range of applications and for diverse 
sensor types exhibiting a wider range of image scale. 
This paper therefore presents an approach to derive object 
models for low resolution images from models created 
manually for high resolution images. Although the contents of 
scale-space theory were widely applied to many image 
processing tasks, e.g. for edge and line detection algorithms 
(Lindeberg, 1998), the connection to semantic net object 
representation for knowledge based image analysis is new. 
Section 2 gives an overview about the general strategy of the 
procedure and briefly describes the different steps. Section 3 
focuses on the composition of the semantic nets and suggests 
some constraints, in order to be able to handle the semantic nets 
automatically regarding scale adaptation. The semantic nets 
represent the high level processing of the image interpretation 
task, but also the low level processing, which is directly 
connected to the nets, has to be observed. Section 4 describes 
tests on the scale behaviour of some feature extraction 
operators, and section 5 contains an example for scale change 
events observed in a scene and their impact on the semantic net. 
2. STRATEGY FOR SCALE ADAPTATION 
This section gives an overview of the proposed strategy for 
scale adaptation. As shown in Fig. 1, the main input of the 
process is a manually created object model, represented as a 
semantic net, with the description of that object, which has to 
be extracted from images. The details of the object description 
are adjusted to a large start scale. Object parts, which are not 
observable at that scale, are also not represented in the object