Petra Zimmermann
non-shadow regions, Hue is used to support separation of vegetation. The presented algorithm computes a low-level
automatic colour-based segmentation. It fuses colour and edge features (see Figure 12) in region-growing process to
derive homogeneous colour regions with accurate boundaries (Figure 15). Colour is the major segmentation feature of
this algorithm edge information (Figure 15) is as a secondary cue implemented because the boundaries of colour areas
may not be accurate due to noise, too low resolution etc. edge information is implemented in the process of
segmentation to win accuracy. This combination uses complementary information: colour based merging works well on
smoothed and quantized images while edge detection finds high frequency information in images (Figure 12). Input
data for this processing is an aerial image (Figure 11), optional an edgemap if available instead of computing it in this
step, and if available the blob data. Segmentation of only colour data will lead to misclassifications e.g. from roads and
roofs of similar materials. Figure 13 shows one example - the flat roof in the bottom part is due to its photometric
properties classified to be similar to the road on the left side. Through to the blob information the classification region is
reduced. The processing steps are:
e the image is quantized in an perceptually uniform colour space (CIE L*u*v*
or HSV) to a limited number of colours using clustering (k-means)
e the image is smoothed with a non-linear median filter, the edgemap is
extracted by Canny operator from every colour channel, strong and weak
edges are labelled
e iterative clustering is done, two neighboured regions with the smallest colour
distance are continuously merged until the difference is larger than a given
threshold
* only non-edge pixels are processed and labelled, edge pixels are not merged
into any region, regions clearly separated by long lines will be prevented to
be connected or merged
e recursively common boundaries of adjacent regions are removed
e after merging all non-edge pixels, edge pixels are assigned to the neighbour
region according to the colour similarity (threshold)
e refinement through multiple repetitions, small regions can be eliminated
e perimeter, bounding box, number of neighbour pixels and edges are
computed
segmented regions must be homogeneous, checked by texture comparison
Figure 13: edgemap
and segmentation
without blobs
Region boundaries are stored as an Polygon2D, with an attributes for colour range in the regions, number of merged
regions, number of edges with their attributes (4.3), perimeter, bounding box, number of neighbour pixels and their
texture attributes (see 4.5). The quotient of region pixels to perimeter gives a measure of compactness.
3.5 Texture
To guarantee the spatial homogeneity of each segmented region and to get a criterion to differentiate between e.g. trees
and buildings, or roads and low vegetation (Table 1), texture information is integrated and also segmented. We use the
well-known definitions of texture from Haralick at al. 1973 and derive 6 texture parameter: contrast, correlation,
direction, entropy, homogeneity and uniformity. Pure texture segmentation gives a only a coarse segmentation, so we
use texture segmentation only as auxiliary tool to check colour segmentation and get texture parameters for the
segmented regions.
Figure 14: shadow areas Figure 15: edgemap used for classification
1068 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
3.6 Sha
Shadow ca
has to be r
14), edge 1
by shadow
misclassifi
4 REA!
For buildir
attributes c
colour, ed;
detectors, t
and colour
colour and
A DSM 1
elevation a
shape, hei;
axes. Fort
we derive
dependent
store it
knowledge
characteris
buildings :
texture, : 1
shape and
of man-m
natural o
texture 3
implicitly
fuse these
done wit
elevation
thresholds
different «
edge info
refinement
regions wv
planes, oui
steps objec
As a result
with their :
give const
primitives
extended «
regions. B:
4.1 Pu
The size :
algorithms
Figure 16
Quasi-Bay
*Buildingl
e.g. "Segn
accepts a 1
preference
the user to
