Table 6 is taken from Mikhail et. al. (1983) and shows the effect of daz compression onto 
the precision of object location for 24 artifical cross targets introduced into digitized ima- 
gery. The precision of the location is estimated from the results of a least squares approach. 
Table 6 Empirical precision of target location vs. data compression rate, standard deviations 
in pels (adapted from Mikhail 1983) 
  
8 bit/pel | 2 bit/pel 11 bit/pel [0.5 bit/pe 
noiseless 0.028 0.054 0.105 "880,28 
with noise 0.038 0.058 0.115 0.53 
  
  
  
  
  
  
  
  
The precision is determined for the case where additional noise is introduced, having 1/4 of the 
standard deviation of the signal background and for the noiseless case. The original data being 
quantized to 8 bit/pel have been compressed to 2, 1 and 0.5 bit/pel using the cosine transform 
compression method. 
The results clearly show the decrease in precision due to information reduction. The additional 
noise seems to have an influence on the presision in the extrem cases. Data compression down to 
‘0.5 bit/pel seems to cause problems, as the identification process of the procedure (using the 
Fourier descriptors) is not able to work when one leg of a cross target is missing. The identifi- 
cation using the moments of the patch however worked also in these cases. Nevertheless, even when 
data are compressed to 1 bit/pel the precision of target location is still 1/10 of a pixel. 
Both investigations give clear indication that data compression decreases the quality of corre- 
lation. Further research has to clarify whether there are methods to economically store templates 
without loosing too much information needed for high precision image correlation. 
The previous sections discussed quality measures which have a direct impact on the economy of 
the whole procedure of image correlation. Of course there are other aspects which strongly influ- 
ence the performance, e.g. hierarchical structures (cf. Sharp 1965, Pearson et. al. 1977, Hobrough 
1978, Tanaka 1978, Wong 1978, Wong and Hall 1978, Makarovic 1980, Baker 1983) or the use of the 
epipolar geometry (cf. Helava and Chapelle 1972, Keating 1975, Kreiling 1976, Wrobel 1977, Benard 
1982), which could not be treated in this contribution, but on which the relations discussed above 
might have an influence. 
4, Final remarks 
Digital image correlation is the basic prerequisite for all tasks in digital photogrammetry and 
remote sensing where the geometric aspect is dominant. Digital correlation is the bottle neck 
for an economical implementation of systems for height measurements or differential rectification, 
but also aerial triangulation, deformation analysis or other high precision applications. The 
accuracy potential inherent to the images may be used to advantage if the correlation algorithms 
are able to react on extremely varying situations in an optimal way. This optimum has to be de- 
fined by the user who may then apply the results collected in this paper. 
Though there are still quite some problems to solve tempate matching obviously is a simple 
procedure compared to the algorithms which are necessary for automatic mapping especially in 
large scales. Here the oversimplification of the mathematical model becomes apparent, as it is 
only a 2D-model working on the basis of more or less undisturbed images of a smooth surface. Thus 
digital correlation in its classical form will fail in all cases where the surface tobe reconstruc-