ce shuttle orbit deter-
lelling of attitude ma-
)2/D2 mission; Flight
ry Symposium, God-
-19, Greenbelt.
(1994): A simulation
or the MOMS-02/D2
unctional model; Geo-
taler Gelandemodelle
rste Ergebnisse; Geo-
T., Schmidhuber M.
orbit adjustment us-
gs of the International
une 19-23, Toulouse,
A. (1996): The Eval-
ree-Line Scanner Im-
ilts; Photogrammetric
2(3), 293-299.
CORSO Putz E. (1996):
Stereoscopic CCD-
d paper at the XVIII.
mission I, Vienna.
hof T. (1994): A
94 Orbit Determina-
le Adjustment; DLR-
und globale Punktbes-
und Bahninformation
Geodätische Kommis-
MS-02: An advanced
reo scanner for Earth
tems 6(1), 4-11.
ına 1996
SEGMENTATION AND TEXTURE ANALYSIS
W. Eckstein
Technische Universität München
Forschungsgruppe Bildverstehen (FG BV), Informatik IX
Orleansstr. 34, 81667 München, Germany
eckstein @informatik.tu-muenchen.de
Commission III, Working Group 2
KEY WORDS: Segmentation, Aerial Imagery, Algorithms, Vision, Feature Extraction, Recognition
ABSTRACT
This paper describes the state of the art in segmentation algorithms of aerial images. Different approaches and object classes are
described and their advantages and limitations are shown. First the advantage of multiple input data (e.g., color, infrared, DEM) and
the information that can be derived from these sources is discussed. Besides sensor data, "synthetic" input images (e.g., using texture
filters) are generated to support the segmentation process. After an optional noise cleaning, primitives are extracted in scale space.
This offers the possibility of selecting an optimal resolution depending on the size and shape of an object. Using this resolution, the
raw segmentation will be stable and conflicts with other object classes will be reduced. Depending on the class of the object the final
extraction has to be selected: Compact artificial objects can be segmented using primitives like areas, lines, or points. Linear objects
like roads are similar but the borders are curves and the size is not limited. Arbitrary areas like meadows, forests, or fields have an
arbitrary border and are mainly defined by their specific texture. Objects like trees or cars have to be treated in a very specific manner.
Finally, different base algorithms for segmentation are discussed: Pixel classification is very simple but lacks the use of context. The
extraction of primitives (egdes, lines, area, points) can be used as a basis for a wide class of objects. Texture analysis can be used for
a rough segmentation of the image. Specialized operations are useful for the extraction of objects like single trees or to support the
interpretation process.
1 INTRODUCTION
Before describing the topics of segmentation we have to discuss
one important question: Is there a known algorithm to extract all
objects in aerial images? The answer to this question is no and
will remain no for many years and it is not even clear if there
exists any. Segmentation is not just applying one sophisticated
procedure and thus extracting all desired objects. On the other
hand there exist a lot of more or less specialized algorithms.
These have to be selected, depending on the classes of objects to
be extracted, the resolution of the image, and the type of sensor.
The reason for this is the complexity of an aerial image. There
are completely different classes of objects, like, buildings, roads,
rivers, trees, meadows, fields, rocks, ice, hills, cars, poles, bridges,
ships, waves, to name but a few (see figure 1). These classes have
different extensions (e.g., cars and roads), specific or indifferent
shapes (e.g., trucks and forest), uniform or textured surface (e.g.,
roofs and forest), which also depends on the resolution and can
be extracted locally or only globally (e.g., trees versus rivers).
In addition, the appearence of objects changes depending on the
point of view, the weather, the time of day, and the season.
On the other hand there is a lot of information about the ob-
ject classes. This knowledge can be used in multiple ways: As
the basis for the interpretation, but also to design the segmenta-
tion procedure in two directions. Firstly, the selection of sensors
and procedures operating on their data defines the static (pro-
cedural) knowledge incorporated in the system. Second, shape,
topology, and radiometry of object classes can be used to control
the segmentation process during runtime (dynamic knowledge).
Constructing a system for a “complete” segmentation of an image
(ie, with different object classes) the following points have to be
observed:
e Use all input sources that ease the task.
e Select the optimal resolution for every object class.
e Select an optimal strategy for the extraction of every object
class.
Neglecting one of these points will limit the system significantly
or at least adds a lot of work for the developer.
2 SOURCES
In the case of aerial image analysis a lot of data sources are
available. Different sensors which allow a more stable extraction
of a special class of objects can be used. Additional information,
like the position of the sun (for shadows) or the angle of view (for
the interpretation of 3D objects), can be used.
2.4 Color
Most interpretation of aerial images is done based on black and
white pictures. The reason is mainly the availability of these
pictures, and lower cost for digitizing and the required computer
equipment. Many problems can actually be solved using this
kind of images. Nevertheless, additional channels, like color or
infrared, can ease the task (Ford and McKeown Jr., 1992). Given
the task of interpreting suburb regions, for example, green areas
are probably lawns, red rectangular areas are candidates for roofs,
and small red, yellow or blue rectangular areas on the road are
probably cars (see figure 2). Using infrared, the extraction of
vegetation is even more stable.
The advantage of color is the simple algorithms for segmenta-
tion which are well known from multispectral analysis in the field
of remote sensing. In some cases even a simple color transforma-
tion like the HSI, HSV, or CIE space, with a successive threshold
suffices. But besides the pixel classification region oriented post
processing must be used to combine groups of pixels to areas.
Morphological operators, like dilation, closing, or binary rank,
are very useful in this context. At the left of figure 3 left an exam-
ple for a pixel classification can be seen. At the right the modified
regions after filling of small holes, applying an opening operation
165
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
LEER MEN EEE
Pu
EX.
