A St siae
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
for all partitions
selecting the data points
if the set is not empty
reading pixel coordinates
for all ordering epochs
calculating learning rate and d(x)
for all data points
for all neurons
calculating distance
endfor
for all neurons
if in calculated neighborhood
modifying weights
endif
endfor
endfor
endfor
endif
endfor
The above algorithm has strong nested loop structure. The
method is therefore computation intensive and sensitive for the
loop control parameters. Fixing the number of epochs, neurons,
partition size, the number of data points has also great influence
on the performance. It was observed, that the smaller sized
partitions resulted faster runs, than bigger partitions, which is
not unexpected looking at the above algorithm.
The function d(x) is the permanently decreasing neighborhood.
The result of the image analysis can be seen in Figure 6, which
shows the regular partitions, too. The structuring elements have
detected visually most of the road segments and their junctions,
only some small errors were occurred. These errors must be
eliminated by implementing constrains for the final neuron
graph structure during the processing, which is the topic of the
current research.
5. CONCLUSION
The newly developed self organizing neuron graph is a flexible
tool in analyzing different types of remotely sensed images. The
template matching problem was solved by a given fiducial mark
graph. Because the camera manufacturers have own figures for
fiducials, the SONG algorithm is capable to fit not only a single
neuron graph to the assumed image part containing the fiducial,
but a whole series. Measuring the fitness for all matching, the
highest fitness identifies furthermore the type of camera. In this
way, the SONG technique recognizes the camera itself. The .
SONG matching is fast, it can provide an alternative solution in
the automatic interior orientation.
The building and road detection belong to the object detection
approaches, which are the most interesting in the modern digital
photogrammetry. The building detection, if we have any
preliminary hypothesis about the building structure, is a
relational matching task. In this sense, the SONG method can
help to realize similar solutions, which are mostly related to the
relational matching. The shown example with the Pentagon
building can be generalized: if one can describe the structure of
a building in form of a graph, it can be found in an image using
the developed self-organizing technique; furthermore this
building can be "traced" in an aerial image strip or block. Only
the same preprocessing operators (e.g. classifiers) must be
executed, prior the SONG run.
The other presented building detection experiment proved that
this method can be used to detect the structure of an unknown
building by creating a hypothesis neuron graph and testing its
suitability. The given example has shown that this hypothesis
testing can be a way to improve the current version of the
algorithm. The test was an alternative solution for getting the
skeleton of an object, solved by the application of artificial
intelligence instead of the known classical skeleton operator.
The most interesting test was the road detection. In this situation
a black-and-white orthoimage ensured the necessary data points
by simple thresholding, which is a very fast image processing
technique. After the segmentation, the SONG algorithm has
found the rough road structure. This can be interpreted in two
Ways.
Once these road segments are objects to perform further
grouping methods to obtain real road network. In this way the
creation of a classical GIS-type topology, then a follow-up
topology analysis and restructuring can lead to the network.
The other possible application of the obtained results is their
interpretation as first approximation of the road network. Using
this philosophy, the connecting processing steps can be
buffering the graph edges. With the fusion of the independent
buffers, we will get a subsample of the image, where the
probability of the roads in the image is relatively high.
Thereafter the different checking algorithms can be executed,
which test these image subsamples on containing roads. If the
test is positive, the exact road segment and its geometric
position can be detected using more accurate methods. In this
checking the SONG technique can be involved in the complex
algorithm, i.e. with also the tuning phase.
The result image of the road detection contains some noisy
parts, especially in urban areas. The method should be used with
a preliminary masked input image, which doesn't have any
urban land covers. Other alternative would be to establish an
urban parameter set, which can have better performance under
built-in circumstances.
The last test points to a new possibility for using highly
parallelized algorithms in photogrammetric image analysis,
because after partitioning the input image, the same steps of the
algorithm must be evaluated for each image parts. If a multi
processor computing environment (e.g. dual Pentium PC, or
even computer cluster or grid) is available, the method can be
implemented and used.
The general assumption of the SONG technique is to own an
adequate initial neuron graph. A nice initialization could be the
use of an obsolete topographic map, where the novel method
would be responsible to check the old map (database) content to
the new image information. In this meaning the method suits
also to the map update procedures. :
The road detection test lasted in Matlab (interpreter type)
environment about two minutes on an Intel P4 1.7 GHz
machine, all the other are significantly faster thanks to the
compiler (MS Visual C++) realization. The image size was
555 x 827 pixel. This fact also underlines the performance
power of the method.
As the paper presented, the generalization of the original
Kohonen-type SOM can be extended by graphs. The newly
developed SONG method has been proved its capability in
different photogrammetric and remote sensing tasks. The
technique has shown how to cope with different types of tasks
using the same algorithmic background. The method is
important in the point of view of the artificial intelligence and
neural networks, because the general suitability and
applicability has also been proved.
ACKNOWLEDGEMENTS
The author expresses his thanks to the Alexander von Humboldt
Foundation, when the work was started; to the Institute of
Photogrammetry and Geolnformation, Hannover; to the
Hungarian Higher Education Research Program (FKFP) for
partly financing the research.
Internat
Barsi, /
Machin
Lecture
Verlag,
Barsi, /
Analyzi
Geoscie
Toulous
Barsi, A
Resolut
Resoluti
Barsi, A
graph ¢
Networ
Baumge
HE, 19
Extract
(II) Moi
Carpent
recogni
pp. 243
Doucett
Autome
using Kk
Stefanic
and Gl.
Springe
Doucett
Self-org
imagery
Vol. 55
Henley.
enginee
Hinz, S
H., 20
Baltsav
Objects
Lisse, p
Kohone
Transa
Kohone
Springe
Kohone
Malsbu
sensitiv
100
Reinhai
cop350