ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
The similarity measurement d(R, J) depends of the intensity, the 
scale, and the rotation of the texture. Since this dependency is 
undesirable in many applications, three matching procedures are 
proposed: the Intensity-invariant-, the Scale-invariant-, and the 
Rotation-invariant matching. 
3.3.1  Intensity-Invariant Matching 
If one is only interested in the textural features, fpc has to be 
ignored for the computation of the similarity measure. 
w(0) 20 (3) 
3.3.2 Scale-invariant Matching 
The querying image R is zoomed with N different zoom factors, 
leading to 7 different scaled versions of the querying image. 
d (R,J,n) = distance (TD, (k),TD, (k)) (4) 
The similarity measure d(R, J) is then the minimum of the 7 
obtained distances. 
d (R,J) 2 minimum(d (R,J,n);n 2 1,2,.., N) (5) 
3.3.3  Rotation-Invariant Matching 
Here the feature vector TDg of the reference image R is shifted 
in the angular direction by à = 30°: 
d (R, J, mà) = distance (7D, .. (k), TD, (k)) (6) 
The distance used for the rotation invariant descriptor is then 
calculated as: 
d(R,J) = minimum{d (R, J, m6); m = 1,2,..6, 6 = 30°} (7) 
4. EXPERIMENTS 
An investigation of the HTD for the extraction of vegetation 
like bushes or trees is presented in this section. The 
qualification of the HTD for the coarse segmentation of aerial 
imagery was shown in (Manjunath et al., 2000), (Newsam et al., 
2002). 
As mentioned above we are mainly interested in the 
differentiation between roofs and trees. In our experiment we 
have selected quadratic regions from an aerial image, which 
show roofs or trees. These data will be used for the further 
investigation. The performance of the Intensity-Invariant 
Matching (refer 3.3.1) was tested. The Scale- and Rotation 
Invariant Matching Methods were not investigated. Scale- 
Invariant Matching is not so interesting in the given context, 
because the scale is usually known for aerial images. The 
Rotation-Invariant Matching procedure is not applicable for the 
distinction between roofs and trees, because tree textures are 
more or less isotrop and roof not. 
  
Figure 2: Example textures used for the investigation 
A closer investigation of the 7D is presented in the second part, 
focusing on object specific properties. These properties can be 
used in a more object specific approach for the extraction of 
trees in an urban environment. 
4.1 Description of the Used Image Data 
The quadratic image tiles, which we have used for a first test of 
the qualification of the HTD are taken from the green channel 
of an aerial CIR image with a GSD (Ground Sampling 
Distance) of 10 cm. The regions with trees and different kinds 
of roofs were selected by hand. A subset of 64*64 pixel is taken 
from the image, and enlarged to a size of 128*128 pixel using 
bilinear interpolation. The size of the subset was selected such, 
that the tiles cover a homogeneous textured region. The size of 
128*128 pixel for a tile was also used in the performance tests 
of the MPEP-7 consortium. The resulting image tiles show tree- 
or roof textures with a simulated GSD of 5 cm. A part of these 
tiles are depicted in Figure 2, 16 examples with typical tree 
texture, and 14 examples with roof textures. Two different types 
of roofs can be clearly distinguished, one type with a preferred 
direction of the texture, and another type without that property. 
The tiles in Figure 2 are ordered by the similarity to the tile in 
the upper left corner, refer section 4.2 for details. 
  
  
  
  
  
  
  
  
  
1 1 13 19 23 
2 8 14 20 26 
3 9 15 21 27 
4 10 16 22 28 
5 11 17 23 20 
6 12 18 24 30 
Table 2: ID's of the textures depicted in Figure 2 
4.2 Intensity Invariant Matching 
The test of the Intensity Invariant Matching method should give 
us a first idea of the performance of the algorithm. The feature 
A - 353