The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
414
contrast closed-boundary regions of an appropriate size, ponds
and lakes, buildings, forests, car parks or shadows. The general
criterion of closed boundary regions is prevalent. The regions
are often represented by their centers of gravity, which are
invariant with respect to rotation, scaling, and skewing and
stable under random noise and gray level variation. Region
features are detected by means of segmentation methods (Pal,
N., Pal, S., 1993). The accuracy of the segmentation can
significantly influence the resulting registration. The
researchers took much efforts to develop segmentation
technique robust to various changes in shooting conditions
(Zitova, B., Flusser, J., 2003).
Figure 1. Rural area
Figure 2. Urban area
Combining merits from two groups of algorithms, developed
for lines and regions separately one can try to construct the
improved technique for extracting contours of desired quality.
In this paper two control pairs of images are used (see Figure 1,
Figure 2) to demonstrate the main steps and the results of the
samples recognition technique.
2.2 Extraction of contours
For edges extraction Sobel filter and Canny detector was used
as basic technique. Then morphological algorithms was
followed for pruning the tales from contour lines. The results of
independent contour extraction are shown in the Figures 3 and 4.
This technique produced a large number of edges, but the result
is still too noisy and spotty.
Figure 3. Rural area, edges
Figure 4. Urban area, edges
2.3 Image segmentation
Image segmentation was performed by watershed algorithm,
which is known to produce many small areas at the preliminary
stage (Shafarenko, L., Petrov, M., Kittler, J., 1997). Then
closest areas joined according to some criteria. For images
under consideration, fusion was performed according to mean
gradient difference along the borders. Second rule for fusion
was characteristic object size, which amounts 200 pixels for
images under consideration. Pre-processing of images with
truncated median filter produces the resulted segmentation of
better quality.
Figure 5. Rural area, regions
Figure 6. Urban area, regions