eration
evolution
ramely the
ay that they
population
at negative
f the genes.
ulation, the
ing can be
diagram. À
0
bling
)ptimization
^" If one has
produced the final population by DE in the studied image, a
strong gene multiplication can be diagnosed as in Fig. 9. A
skilled human operator can extract the right candidates from the
resulting set, passing them into the last phase.
Final population
Figure 9. Final population from differential evolution
The last experiment is a comparison with the artificial images to
obtain recognition features of the technique. The synthetic
images were processed by the same DE settings and the result is
self-explainable (Fig. 10).
5. CONCLUSION
The proposed neural and genetic solution is based on pure
optical image information. There was no additional help in the
segmentation to extract the road pixels. Being able to execute
this step on a highly reliable level, the next step can obtain
better fitting performance. The original assumption was that the
widely used modern neural techniques, like SVM and SOM can
bring excellent result in road recognition. The first experiment
series aimed to check this hypothesis. The results were
surprising, because the hyperbox method could reach better
scores in this test. The additional image based information band
(like the NDVI) and the principal component analysis could
increase the accuracy. Object oriented segmentation or more
sophisticated road detection methodologies producing pixel-
type result can replace the presented methods.
The biggest novelty is the application of the genetic algorithms
in information retrieval from segmented images. The common
genetic algorithm has a quite long run until a stable state can be
reached, the new differential evolution technique can replace it
because of its higher speed. The shown GA and DE methods
have global optimization feature; the dependence from the
implemented fitness function is crucial.
Figure 10. Synthetic segmentation image processed by DE
algorithm (longest 5 of the 10 best genes)
More research must focus on the suitable local applicable
fitness function, which is able to compile the pixels into road
segments, but is fast and reliable at the same time. The genetic
solution is almost independent from the image size and
resolution, as well as from the number of road elements
(segments), when enough genes are handled in the population.
The experiment has proven that based on mutation this
algorithm can extract such linear image elements.
The used software environment was Mathworks Matlab, which
is an interpreter type environment. A great experience was with
the differential evolution technique that it was suitable at full
resolution to bring acceptable results. A future development by
e.g. OpenCV can dramatically increase the size of the image to
be processed.
6. REFERENCES
Beale, M.H., Hagan, M.T., Demuth, HB., 2012. Neural
Network Toolbox. User’s Guide, R2012a, Matlab, The
MathWorks Inc, Natick
Laky, S., 2012. Metaheuristic optimization in the geodesy, PhD
thesis, Budapest, p. 115
Rottensteiner, F., Baillard, C., Sohn, G., Gerke, M., 2011.
ISPRS Test Project on Urban Classification and 3D Building
Reconstruction, ISPRS - Commission III — Photogrammetric
Computer Vision and Image Analysis, Working Group III / 4 -
Complex Scene Analysis and 3D Reconstruction.
http://www.commission3.isprs.org/wg4/
7. ACKNOWLEDGEMENT
This work is connected to the scientific program of the
"Development of quality-oriented and harmonized R+D+I
strategy and functional model at BME" project. This project is
supported by the New Hungary Development Plan (Project ID:
TAMOP-4.2.1/B-09/1/KMR-2010-0002)