Figure 8: Example of segmentation errors with different
neighbourhood systems (left) and the original image (right)
Results with airborne images
The methods and approaches presented were taken into account
for segmentation of airborne images. Images taken by a DMC
airborne camera (Intergraph GmbH) were radiometrically
normalised, the model parameters were normalised to a uniform
length after extraction and a neighbourhood system with 10
neighbours was used. The segmentation was carried out for
different texture window sizes and scaling.
Good results could be achieved with the used pictures when the
scaling was larger than 50% and the texture windows were
small (3-5 pixels). Nevertheless, the segmentation is affected if
there is poor image radiometry (see figure 9, middle picture).
Moreover, heterogeneous areas, e.g. urban areas with many
single trees, make segmentation difficult, though in these cases
a crude discrimination is possible.
The finer the texture and the higher the scaling the smaller the
texture window should be chosen. However, no general rules
can be defined from these results. The texture window size and
the scaling must be adjusted for every special texture.
The neighbourhood system, by contrast, has only little
influence on segmentation results. The reason is that only a few
parameters contribute to the indication of texture quality.
Additional parameters (a bigger neighbourhood system) do not
contain extra information and do not enhance segmentation
results. A NBS with 10 neighbours is usually adequate, and
therefore was used for further investigations.
5. CONCLUSION AND OUTLOOK
Figure 9. Three examples of segmentation results with aerial
photos. Blue is the extracted road area.
The paper considers road detection in panchromatic images for
traffic observation from airplane platforms. Because of the
limited image size structure based approaches cannot be applied.
Therefore, in this paper texture based algorithms are utilized.
The influence of image quality and pre-processing, image
scaling, size of the texture window and size of the
neighbourhood system, as well as the influence of parameter
normalization where investigated.
In this context it is interesting to realize that the MRF
characteristics are independent of illumination of the observed
area. According to this it is possible to minimize the influence
of cast shadow - a common problem in natural scenes.
Altogether, it could be shown that the described method is
suitable for the distinction of streets and surrounding areas. For
real-time applications the algorithm and the implementation
must be optimized.
REFERENCES
Besag, J., 1986. On the Statistical Analysis of Dirty Pictures,
Journal of the Royal Statistical Society, B26, pp. 25-36.
Brodatz, P., 1966. A Photographic Album of Artists and
Designers, Dover Publications
Descombes, X., Sigelle, M., and Preteux, F.,1999. Estimating
Gaussian Markov random field parameters in a nonstationary
framework: Application to remote sensing imaging. IEEE
Transactions on Image Processing, 8(4), pp. 490-503.
Gimel’farb, G., 1999. Image Textures and Gibbs Random
Fields. Kluwer Academic Publishers, Dordrecht/Boston/
London
Haralick R., Shanmugam K., Dinstein I., 1973. Textural
Features for Image Classification. IEEE SMC 3, pp. 610-621