International Archives of Photogrammetry and Remote Sensing.Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
66 
2. The integration scale h, needed for determining the 
texture properties at the highest resolution. 
3. The number m of the used pyramid levels. 
4. The differentiation scale s 2 , needed for determining the 
integrated squared gradient of the texture features. 
5. The integration scale f 2) needed for determining the 
integrated squared gradient of the texture features. 
5.2 Choice of scale 
One of the main and up to now unsolved problem is the 
choice of suitable scales for texture edge extraction. This 
is a general problem and to our knowledge has not been 
solved satisfactorily. The reason is that textures may appear 
at very different scales. Though these scales may be identi 
fied by some automatic means it is not clear whether these 
scales correspond to textures or macro structures which 
the analyzing module wants to resolve. E. g. in the case 
of the tile row 2 column 4 in Fig. 8 two interpretations are 
possible: Either the tile is part of a textured region, where 
the texture is quite regular and has long wave lengths, or 
the tile may be interpreted as a composition of some few 
homogeneous regions separated by lines. 
That means, the user or the calling routine has to decide 
which levels of the Laplacian box should be used to obtain 
the required results. This strongly depends on the applica 
tion. This may be a severe problem in case of textures of 
very different scale. 
In our work this step was done interactively. 
6 RESULTS 
This section demonstrates the properties of the new texture 
edge extraction scheme. 
In all cases we compare the result of the texture edge ex 
traction with the grey level edge extraction from FEX. 
The basic problem is the proper choice of the scale param 
eters. 
For the first investigations presented below, we used the 
Sobel operators as differentiation kernel for the determina 
tion of the texture parameters, thus fixed s\ = I/a/2- 
6.1 Results from synthetic data 
First, we present our results of the texture edge extraction 
for synthetic data. The scales were selected such that the 
feature extraction could rely on the intensity differences be 
tween the tiles. Obviously, this was quite successful in this 
case. The texture edges, however, are a bit cleaner. 
6.2 Results from natural scenes 
In this section, we show the potential of our algorithm for 
texture edge extraction of natural scenes. The results are 
shown in Fig. 9. 
In order to show the difference between grey level and tex 
ture edge extraction, and not to obtain intensity edges alone, 
the integration scale t is chosen larger in the texture edge 
Figure 8: Results of texture edge extraction: Upper row) 
collage of Brodatz-textures; lower row left) using Intensity 
alone (s = 5.0, t = 5.0); right) using the Laplacian box 
(si — 0.7, t\ = 6.0, rn = 3, S2 = 5.0, t 2 — 5.0/ 
extraction scheme. This is reasonable, as we want to group 
several furrows into one field. Obviously, this reasoning 
leads to quite satisfying results. Due to our special imple 
mentation, not all texture edges at the image borders are 
captured. 
6.3 Results from non-textured images 
To achieve an improved method for image segmentation 
we have to ensure, that our technique provides good re 
sults not only for textured images, but also for non-textured 
images. Therefore, we applied our approach to some non- 
textured images too. The results are shown in Fig. 10. 
The result is satisfactory. The spurious texture edges in the 
background can be explained, as no thresholding is per 
formed, in contrast to the procedure for grey level edge ex 
traction. 
7 SUMMARY AND CONCLUSION 
This paper presented a filter based approach for texture 
edge extraction using the scale characteristics of the local 
autocovariance function. 
The approach was implemented and tested on synthetic 
and natural scenes and shows some promising results. One 
of the main problems is the choice of suitable scales for tex 
ture edge extraction. In our experiments this step was done 
interactively. 
To perform a qualitative evaluation, we have to compare 
our algorithm with other approaches, as shown (Shao and 
Forstner, 1994). Also more detailed aspects need to be 
analyzed, e. g. the effect of the higher levels of the Lapla 
cian box on texture edge extraction and the ability to extract 
edges at boundaries between textured and non-textured re 
gions.