ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
vectors TD; are calculated for every tile (refer 4.1), the first 
image is marked as reference image R. The weighting 
parameters w(k) and a(k) in Equitation (2) are computed 
from the feature vectors 7D; as mean value and standard 
deviation. The weight of a sub band, used for the computation 
of the similarity measure d, increases with its energy and 
decreases with its noise. 
w(k)= V XD (4) (8) 
izl-n 
a(k)= V, Ep ()- (9) © 
The distances d(R,J) between the image R and all images J, with 
j=2,3, ..30, are plotted in Figure 3, refer Figure 2 together with 
Table 2 for a qualitative inspection. The distances d(R,J) are 
scaled with the largest distance value d(R,30), which occurred 
in the test data. 
Figure 2 indicates the tiles J, with index j=2,3,...,30 ordered by 
the distances to the reference image R, with index /. The visual 
inspection of this data shows that, with some exceptions, the 
tree textures are well separated from the roof textures using the 
intensity invariant matching method. Thus the HTD seems to be 
qualified for the differentiation between buildings and trees. 
The numerical values for the similarity are represented in Figure 
3, the largest distance d(R,30)=4.5 is used to scale the distances 
d(R,J). As the textural energy in image 30 is very poor, this 
should give an idea of the range of distance values, which may 
occur in praxis. In our example a threshold value, placed in the 
centre of the range, i.e. dryresnorp=0.5, will lead to a retrieval 
of the first ten tiles, a success rate of 63%. A threshold value 
drurestorn=0.65, would lead to a success rate of 82%. An order 
of magnitude comparable with the results of the performance 
test using Brodatz Textures (Brodatz, 1966) in (Man Ro et al., 
2001). Nevertheless, the computation of dr7jsgsgorp from the 
values in Figure 3 is not obvious. 
  
29/29 
27 
24 25 26 
19 20 21 22 23 
16 17 18 
14 15 
13 
il i2 
10 
054 8 
jill 
Figure 3: Scaled Distances for the Intensity Invariant Matching 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
A further investigation should include, besides an enlarged data 
set, tree-textures which are clearly different, for example 
textures from conifers. Nevertheless, the first results are well 
promising, taking into account that there was no room for a fine 
A - 354 
tuning of the parameters of the filter bank, and the threshold. 
The intensity invariant matching method seems to be suited for 
the texture based separation of trees and roofs. 
4.3 Discussion of Typical Object Properties 
Once the feature vector is computed, a closer look onto possible 
object specific properties, which are reflected in the feature 
vector, is obvious. 
The texture of a tree does not have a major orientation. Thus the 
energies for all directions of one sub-band should have similar 
values. This observation is always valid for deciduous trees, for 
conifers only if they are quite close to the center of the image. 
Against textures of trees, roof textures have one or two main 
directions. This property should lead to one or two peaks in 
every sub band, and the angular index should be the same for 
these peaks. 
The TD mean values of the tree and building textures confirm 
this assumption, refer Figure 4. In the three middle sub bands 
the energy values are relatively homogeneous for trees, and for 
buildings peaks reflect their main orientation. 
IT ZI 
7 Buildings (Meanvalues) || 
749 Trees (Meanvalues) — | 
    
Figure 4: Mean values of energies for buildings and trees 
The standard deviation s gz ,) for the sub bands with constant 
Q -values can be used to measure this property. 
A, = # Em (6r+m), 
with :r = [01234]. mn =L0;1.2.3;4.51 
(10) 
Sur - Xn, cm (orem) (11) 
The mean value s; of the standard deviations s; gz , can be used 
to differentiate between trees and buildings. From the example 
data set, which is depicted in Figure 2 one gets srazz-0.1 and 
SBUILDING 0,18. 
5. SUMMARY AND OUTLOOK 
In this paper we have investigated the performance of the 
MPEG-7 Homogeneous Texture Descriptor. The main focus of 
this paper is on the investigation and discussion of the HTD’s 
qualification for the detection and possibly reconstruction of 
trees from high resolution imagery. The results are well 
promising, and it seems that an integration of the HTD in our 
system for the extraction of trees in urban environments will