7 Conclusions 
We have developed a new approach to region growing, based 
on an edge prediction stage and a merging stage. The 
method assumes (limited) a priori knowledge about the im- 
age noise. 
Over the split-and-merge method, our scheme has the 
advantage of not dislocating edges. Due to the inherent 
smoothing abilities of the predictor, for most natural im- 
ages, noise smoothing is not necessary. Images contami- 
nated by heavy noise, should be preprocessed; in Stage I 
by the extended Kuwahara filter or the Gaussian filter and 
in Stage II by the median filter. Further, in Stage II, not the 
mean but the median of the grey values should be taken to 
characterize the region grey value average. The rationale is 
that regions may be corrupted by all types of noise and tex- 
tures. By taking the median and applying a median filter 
a priori to the computation of the variances, small clusters 
of deviating grey values will much less affect the decision 
about merging. 
The size of the half plane predictor is not very critical. 
In Stage I it is much better to accept that an edge is present 
than accepting wrongly that no-edge is present. Edges that 
are uncorrectly traced are easily removed in Stage II. How- 
ever, a non-detected edge will never be discerned in a later 
stage. Consequently, o, the probability of rejecting wrongly 
a non-edge may be set rather large. However, there is a 
trade-off between choice of a and computation time, since 
a large o yields many regions resulting in a heavy compu- 
tational burden in Stage II. In Stage II wrong not-merging 
of regions should be avoided, so o should be rather small. 
Since our method is based on a notion about image 
noise, we arrive at a better insight into the setting of the 
thresholds, than is usually achieved with heuristical thresh- 
old settings, since noise is physically appealing. A priori 
knowledge about the noise is only necessary in Stage I, and 
there the noise estimation is not critical. To avoid that 
edges are missed, one should estimate the noise optimisticly, 
i.e. better too low than too high. 
801 
References 
[1] Ballard, D.H., Brown, C.M., 1982, Computer Vi- 
sion, Prentice-Hall Inc., Englewood Cliffs, New Jersey. 
[2] Horowitz, S.L., Pavlidis, T., 1976, Picture seg- 
mentation by a directed split-and-merge procedure, 
Proc. 2nd IJCPR, pp. 424-433. 
[3 
MÀ 
Jain, A.K., 1989, Fundamentals of digital image pro- 
cessing, Prentice Hall 
[4 
uu 
Janssen, L.L.F., Jaarsma, M.N, van der Linden, 
E.T.M., 1990, Integrating topographic data with Re- 
mote Sensing for land-cover classification, Photogram- 
metric Engineering & Remote Sensing, vol. 56, no. 11, 
pp. 1503-1506. 
[5 
— 
Lemmens, M.J.P.M., Verheij, K.F., 1988, Auto- 
matically GIS updating from classified satellite images 
using G.LS. knowledge, Int. Arch. of Photogrammetry 
and Remote Sensing, vol. 27, Part B4, pp. 198-206. 
[6 
beds 
Lemmens, M.J.P.M., 1990, Automized interface of 
digital spatial imagery and geographical information 
systems, Int. Arch. of Photogrammetry and Remote 
Sensing, Vol. 28, Part 3/2, pp. 374-385. 
[7] Lemmens, M.J.P.M., 1991, GIS, the data problem, 
Proceedings EGIS ’91, pp. 626-636. 
[8 
MÀ 
Pavlidis, T., 1977, Structural pattern recognition, 
Springer.