measure of the errors in parallax measurements without gross errors,
and reliability, as the share of parallaxes without gross errors. Howe-
ver the precisions, as measured above, are of no practical value if the
largest part of the gross errors could not be identified and excluded.
In the report "Digital Matching of Simulated SPOT-images (Rosenholm,
1985a) it was found that gross errors could be identified to a large
extent. 80% of the gross errors were detected only by maximising the
number of iterations and the size of the estimated error. Methods for
gross error detection suggested in that report were, maximum value of
translation, maximum number of iterations, maximum value of estimated
errors and mimimum convergence rate. These criterias were tested in
this investigation. We will start by discussing the standard deviation
of the parallax as exclusion criteria. Matchings with an estimated
error larger than a certain threshold value, matchings which have
reached a maximum number of iterations or which were interrupted be-
cause of a too large translation after the first iteration were exclu-
ded. This gross error detection worked sufficiently well for the two
data sets "Rock" and "Low", which are of good radiometric quality, lea-
ving 3% and 18% of the gross errors unrejected when the strictest thre-
shold (3 um) was used. In the other data sets we never really reached
an acceptable accuracy, around half of the unrejected measurements con-
tained gross errors. The radiometrical quality of the image pair has to
be fairly good in order to make an acceptable result from the matching
possible. The automatic matching methods used do not really seem to be
useful on bad images, even with gross error detection. Only the data
sets "Rock" and "Low" seems to have the needed radiometric qualities.
Also another criterium for exclusion was investigated. Instead of using
the estimated error of the parallax, a minimum of the median of the
ratios between the last and the previous translation was the exclusion
criteria. The convergence rate, measured in this way, does not seem to
be quite as efficient as the estimated error of the parallax for gross
error detection. A combination of all these criteria was also investi-
gated, but could not be found to be significantly more effective than
the use of the estimated error for gross error detection. However, if
the convergency rate is measured we will have the opportunity of inter-
rupting an erroneous matching earlier, which could reduce the total
computational effort.
SOME OTHER INVESTIGATED IMPROVEMENTS
When the relative orientation is known, the matching could be done in
one dimension. In the investigation both the precision and the re-
liability of the measured x-parallax were improved when the matching
was done in one dimension instead of two.
In the report "Digital Matching of Simulated SPOT-images" (Rosenholm,
1985a) negative consequences because of the large influence from the
bright areas on the result were experienced. This is caused by the
additive constant noise model. If the noise is proportional to the
signal, a logarithmic grey level transformation will transform it to a
constant noise model, and the large influence from the bright areas
Will be reduced. The differences between this grey level functions,
both with regard to precision and reliability, were small and insigni-
ficant, probably due to the high redundance in the systems.
- 584 -