classification 
ifferent textures 
jance with respect to 
re deciduous trees are 
         
leciduous trees 
fferent type of textures 
nly implicit (e.g., fre- 
features, like minimal 
e used. 
ated in a very specific 
nplex structure which 
se of deciduous trees it 
specially if they stand 
t is possible to detect 
8 shows two extreme 
nage can be extracted 
n, 3D information, like 
10 procedure known to 
In figure 19 one tree 
seen. Looking at the 
ucture of the branches 
se trees is base on the 
in, 1986): 
> outer parts. 
hadows). 
1S some texture. 
r the influence of the 
reted as distinct bright 
e watershed algorithm 
now along the darkest 
part of the trees a local 
  
D 
  
  
  
  
  
  
TAT 
4 7 iJ 
|. 
| 
  
  
1 
m4 
  
  
  
  
  
polygons parallel lines 
homogenous areas extended lines final result 
Figure 15: Grouping process for road extraction 
  
  
bot spruce 3D-plot smoothed image watersheds local threshold 
Figure 19: Segmentation of fir trees by calculation of watersheds in a smoothed image 
    
single trees dense forest 
Figure 18: Different forms of deciduous trees 
6 BASIC SEGMENTATION ALGORITHMS 
A variety of algorithms are available as building blocks for the 
construction of a complete segmentation procedure. Depending 
on the object class to be extracted, one or more of them is needed. 
6.1 Pixel Classification 
This is the well known class of point operations, mainly applied 
to multichannel images in the field of remote sensing. In the case 
of aerial images the algorithms can be used with color or infrared. 
As we will see in section 6.3 synthetic channels can the generated 
with the help of texture filters. The problem of pixel classification 
is the lack of context. Some kind of context can be added by using 
resolution pyramids as additional channels or by post processing 
the classified pixels (closing, dilation, etc.). 
6.2 Primitives 
The most polular approach for the extracton of objects is to find 
edges. Different operators have been proposed. Besides simple 
filters like Sobel, Kirsch, or Prewitt more sophisticated have been 
developed: (Shen and Castan, 1992), (Canny, 1983), (Canny, 
1986), (Lanser and Eckstein, 1992). Edge detection assumes that 
an object consists of one or more constant areas. In general it 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
is not possible to extract closed contours, so the edges are used 
as primitives. In many cases they are approximated to polygons. 
Junctions and points of high curvature are derived as additional 
primitives. Using an edge detector, blobs, defined as areas of low 
gradient, can be extacted by a threshold operation with optional 
postprocessing e.g. opening. 
: + YA 
Pr {+ ot 
AL, 
         
    
gray image 
polygons 
Figure 20: Two types of primitives: blobs and polygons 
A more general approach is given by (Forstner, 1994). Here 
different types of primitives are extracted simultaneously: 
e¢ Homogeneous areas 
e Edges and lines 
e Points (boundary points of high curvature or junctions) 
These features are well defined in the case of artifical objects like 
buildings (see section 5.1). The approach is based on the average 
squared gradient defined by 
2 
IgG olo 15 deo (1) 
JyIx gy 
where G is symmetric Gaussian function with standard deviation 
c. Theextraction of primitives is composed of the following steps: 
1. Estimation of noise characteristics. 
2. Information preserving restoration using a Wiener filter (see 
section 3).