ject space is 
ons resulting 
3rown model 
on gives the 
, as the input 
nverse trans- 
olving a two 
an iterative 
ziven regard- 
orted coordi- 
tion (see be- 
Q) 
ected by Ax 
ve distorted 
position. If 
ions can be 
the parame- 
€ large and 
n Fig. 7 the 
shown. The 
ed by calcu- 
ng with the 
id the itera- 
\s initial ap- 
(3) 
2p 
?)P, 
  
Q as oi Uri 
ox Oy Xj ET f. + Ox y X; 
os. o. Yid = Fi Fi Fs. y; 
x Oy ox y 
where i + 1 the solution in the j — th iteration 
  
Figure 7. Influence of distortion parameters to the ideal undis- 
torted image — distortions have been amplified 4.5 
times for better visualisation 
The mosaicking of the virtual image is done by resampling the 
grey value from the sub image with the minimum distance to 
the ground point. Alternatively, the weighted average of the 
grey values from all sub images can be used, using 1/distance as 
weight. Since the images were radiometrically balanced the dif- 
ferences in the seam lines were minimum; therefore the weight- 
ing could be avoided. The size of the 2 virtual images was 1562 
x 986 and 1374 x 1662. The difference in the sizes is due to the 
different number of sub-images and enclosed area. 
The created virtual image (Fig. 8) can be used for mapping and 
archaeological documentation without great loss in accuracy. It 
is significant though for the mosaicking and further more for 
the matching the images to be radiometrically balanced. The 
advantages are twofold: features can be extracted and mapped 
that are covered with dust or lie in shadowed or overexposed 
areas and also less radiometric differences exist in the image 
that is mosaicked. 
  
Figure 6. Virtual image created from first strip, using 3 sub- 
images. The sub-images prior to generation have 
been radiometrically equalized using Wallis filter 
(Baltsavias, 1991). 
S. DSM GENERATION 
5.1 General 
Methods for automatic DTM and DSM generation exist al- 
ready in various commercial digital photogrammetric sys- 
tems. Many of them can not handle close range imagery, since 
they are developed mostly for processing aerial imagery. 
Moreover they show to have problems when dealing with im- 
ages that have large tilts and bases between the image pairs 
are small. Therefore we used own developed algorithms in 
both virtual and normal case of images. In commercial system 
the virtual images may be used instead, free of distortions and 
large tilts and a DSM may be extracted. 
5.2 Preprocessing 
The quality of the images was poor and many features lied in 
shadows, also some GCP’s. To optimise the images for subse- 
quent processing, filtering has been applied to reduce noise, 
while preserving even fine detail such as one-pixel wide lines, 
corners and line end-points. The filter employs a fuzzy method 
and require as input an estimate of the noise, which may be 
known or estimated by the method mentioned in Baltsavias et 
al. (2001). As an estimation of noise the standard deviation has 
been computed and the average noise level of the images was 
1.2 grey values, after the filtering the noise level reduced 50%. 
In the next step, Wallis filter has been applied to enhance the 
content of the image and reveal small structures (Fig. 7). It is 
important prior to Wallis filtering to reduce the noise level oth- 
erwise the existing noise will be enhanced. 
5.3 Extraction of features 
The Canny operator was used to extract edge features in all im- 
ages, both original and virtual ones. A dense matching can be 
achieved and the result can be accurately interpolated into a 
grid. Canny operator was chosen among others because it per- 
forms better than others in terms of processing time and the re- 
sults are of reasonable quality. 
  
Figure 7. Details of image before (left) and after (right) pre- 
processing. Features are enhanced even in areas with 
low contrast 
-443-