ternal
mage
in of
by a
alyse
10des
pour
plus
sont
ns le
GE
wn as à
les. The
with the
an areas,
global
ken into
and the
tem is a
produces
ext step.
complex
hanisms.
stand the
one (for
all these
ze aerial
ution to
external
ssing to
3. THE USE OF EXTERNAL DATA FOR
IMAGE ANALYSIS
Such data are available in several fields of image analysis.
They may be anatomical sketches in the field of medical
imagery, plans in the field of robotics. In our domain of
aerial images analysis applied to cartography, external
data may be maps, plans or geographic databases.
Cartographic external data provide a description of the
scene, that is :
e an inventory of the cartographic objects which are in
the scene (roads, buildings, forests...)
e a description of each of these objects: their shape,
their inner features...
e a description of the links between the objects that
constitute their geographic context.
These external cartographic data may be very useful to
guide the image analysis. On one hand, they allow to
focus the research on some objects and on some areas in
the image. This first contribution is essential since it
reduces drastically the processing time. On the other
hand, the specific knowledge about the inner features of
the objects and about their context allows to regulate the
parameters of the low level algorithms or better still to
select the most adapted algorithm. This contribution is
also promising insofar as most detection algorithms are
efficient in a precise context. Lastly, the external data may
be used to validate the objects detected in the imagery.
Thus, the number of false detections can be reduced.
Through these three contributions, it appears that using
external data is really a good way to improve the
efficiency and the reliability of aerial image analysis.
4. DIFFERENT TYPES OF EXTERNAL
DATA AND DIFFERENT METHODS TO USE
THEM.
In the field of aerial image analysis, the external data may
be provided by various cartographic sources such as
scanned maps, cartographic databases... The diversity of
data sources leads to different methods to use the data in
image analysis processes. These methods are closely
linked to the following four characteristics of the data
source :
e its semantic content,
e its scale,
e its geometric accuracy,
® its representation mode : raster or vector.
The description of the scene provided by the external data
is more or less complete according to these 4 parameters.
The more this description is accurate, the more the image
analysis can rely on it.
Depending also on the image characteristics (resolution,
landscape typology, ...), different ways to integrate the
external data in the image analysis process have been
developped in previous works.
135
4.1 Data fusion processes
These processes are based on data fusion methods in
which external data and images are used without
distinction. This kind of method may be used when the
accuracy of the external data is compatible with the image
accuracy. In (Roux, 1995) satellite images are classified
by merging the images and external data such as DTM
and several distance maps. In each document, the
possibility for a pixel to belong to the different classes is
computed. These different possibilities are merged and
provide the final classification. This method is very
useful to take advantage of different data sources
(including symbolic data) which are complementary.
4.2 Matching and readjustment processes
These processes are used when the external data are
enough accurate and when the image itself may provide a
good description of the objects to be detected. The image
processing is computed independently of the external
data, then the external data are matched with the objects
detected in the image. This matching allows to validate
the objects detected and eventually to readjust the
external data on these objects. In (Servigne, 1993), this
method is used to update a cadastral database. Firstly, an
edge detection is processed on the whole image then the
segments detected are matched with the objects of the
database. It allows to detect changes and thus to update
the database.
4.3 Algorithms guiding processes
In these processes, external data are used to guide the
detection, that is to say to choose the proper algorithm to
process on the image and to adapt its parameters. The
external data have a direct effect on the image processing.
In (Strat, 1995), the external data are used as contextual
data. A general architecture for contextual interpretation is
introduced. The principle is to process an algorithm only
when the conditions (the context) in which it is known to
be efficient are verified. The external data provide
contextual information on each object to be detected in
the image. The other contribution of these processes is
the regulation of the parameters. In (Yu, 1994), the
objects of a map are projected on the image, they define
learning areas for the parameters of the classification. This
kind of processes are suitable for external data with a rich
semantic content.
However, even when the data source is as accurate as a
topographic map the external data cannot be considered as
an absolute reference. In any cartographic document, there
are constraints which induce geometric distortions of the
objects compared to the ground reality and to the image
one.
Thus, before using external data, the data source must be
studied :
e the information which is significant in the image must
be selected.
e the geometric and semantic accuracy of the data must
be evaluated in order to know what can be expected
from them.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996