The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
The results of image segmentation are shown in the Figures 5
and 6. Although the areas obtained are very close to the outlines
of real objects in the source scene we sure, these results could
be still more distinct. One of the disadvantages of watershed
algorithm is that it finds region border biased due to
morphological dilatation used for regions growing.
2.4 Contours versus regions
Next step is contours refinement with taking into consideration
aggregated data from both areas with their borders and
independent contours. For further images processing let us use
three additional rules:
delete all closed contours which are fully contained inside
some region,
merge all regions inside closed contours into one region,
make more accurate borders of regions, taking the nearest
independent contours as precise position of the border.
Figure 7. Rural area, contours versus regions
Figure 8. Urban area, contours versus regions
From successive implementing of these rules the final set of
contours showing in the Figures 7 and 8 (together with regions),
and in the Figures 9 and 10 (together with source images) arises.
Figure 9. Rural area, borders versus source image
Analysis of Figures 7-10 shows that implementing of combined
technique for every image from the presented ones gives a
distinct set of closed contours, near to what we could expect
from human operator. Now one should try to match the
appropriate pairs of contours.
Figure 10. Urban area, borders versus source image
3. CONTOURS MATCHING
3.1 Contour presentation
As a result of extraction and refinement procedures in previous
section a set of contours in the form of simple chain codes for
every image was obtained. In further investigation, the means
of contours presentation plays an important role for success of
their digital comparison by any way. We have chosen for this
purpose rational splines (Zavjalov, Y., Kvasov, B.,
Miroshnichenko, B., 1980), which shows very good behaviour
in situations when shape of curve is not known a priori. A
preliminary step in spline approximation is to find any valuable
set of nodes by means of some regular procedure. A simple and
natural way to obtain the set of nodes required is to build up a
piecewise linear approximation for chain code. A widely known
algorithm described in Prett, W., 1978 was used for this purpose.
A set of polygon vertices obtained is taken as a set of spline
nodes.
On the basis of rational spline one then build up 1-D curve
representation, curvature versus arc length (Davis E.R., 2005,
Schenk, T., 1999). The form of curvature function for 2
contours from 2 images are presented in Figures 11-14. The
initial point of every curve is arbitrary and proper length shift is
estimated during feature matching.