ligitized aerial
marks in the
ular attention
o the internal
tions) for how
anner for dig-
four different
n to recognize
ue asymmetric
1e coordinates
is possible to
itions where it
is scanned. A
to the orienta-
zed image
strically placed
fiducial mark
ent of whether
and from 90°
tion invariant.
criterion, while
in this case is
least four ori-
>, without any
on between the
be estimated,
and possibly a
mirror reversing. Therefore we locate an asymmetric feature
to derive how the image was placed in the scanner.
As fiducial marks are 2-D objects with well defined geometric
and radiometric characteristics and usually the scanned im-
ages have no more rotation than 10°, the cross correlation is
an appropriate matching strategy for locating fiducials. Fidu-
cial marks are usually synthetically faded into the image on
an unexposured, and therefore dark and homogeneous back-
ground. This model information has been integrated in the
localization process by applying a binarization the image us-
ing both criteria, the low intensity and the homogeneity. An
additional advantage of the binarization is the possibility of
using a very efficient binary correlation.
On different pyramid levels we use different representations
for the fiducials, i.e. different templates for the correlation.
Fig. 2 illustrates these different representations for a RMK-
TOP camera. On the highest levels we use the whole fiducial
mark including its surroundings (Fig. 2 left). On the lower
levels the fiducial figure (Fig. 2 middle), and only on the
lowest level where the final measurement is done, the fidu-
cial mark itself is used (Fig. 2 right). To exclude areas with
undefined radiometric characteristics like the contents of the
image or yet unsolved asymmetric features we use so-called
don't care regions. These regions, shown in grey in Fig. 2
left, are not taken into consideration when correlating.
Figure 2: Left: Pattern of the fiducial mark surrounding,
with grey don't care regions. Middle: Pattern of the fiducial
mark figure. Right: Pattern of the fiducial mark
BEE
e oos
A
3 OVERVIEW
The whole procedure of the AIO consists of following steps:
e Resampling of the templates.
According to the resolution of the image.
e Image pyramid derivation.
If not yet present.
e Robust localization of at least four orientation in-
variant fiducials.
In this main part of the hierarchical pattern recognition
process an outlier detection algorithm is implemented
to make the localization process more robust. It results
in very good approximate positions for the orientation
invariant fiducials of less than £5 pixels. Within this
procedure the system also determines whether the im-
age is positive or negative. The robust localization
process is described in chapter 4.
e Detection of the orientation of the image.
There are eight different possible orientations how to
scan an image, i.e. wrong reading or right reading with
four different multiple 90° rotations respectively. To
detect the correct orientation, one asymmetric feature
in the image is located.
747
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
e Fine measurement of all fiducial marks.
This is done using a grey level correlation with subpixel
estimation. The accuracy of the individual location is
about 1/10 of a pixel.
e Estimation of the transformation parameters.
In this step the transformation between the plate sys-
tem and the image system is estimated, and different
kinds of transformation types are possible.
e Self-Diagnosis.
This is done by analyzing the final result with respect to
precision and sensitivity. The self-diagnosis is described
in chapter 5.
4 ROBUST LOCALIZATION OF ORIENTATION
INVARIANT FIDUCIAL MARKS
The principle here is to locate at least four orientation in-
variant fiducial marks individually using a binary correlation.
As the interior orientation should be performed automatically
without any approximate values, we use a hierarchical search
strategy through the image pyramid from coarse to fine for
the location of the rotation invariant fiducial marks.
The procedure of the robust localization which is performed
hierarchically consists of the following four steps:
Definition of the search space
Binarization
Binary correlation
A5 0 hN #H
Consistency check
Steps 2 and 3 are replaced by a grey level correlation and
a positive/negative recognition, in the case that the infor-
mation on whether the image is positive or negative is not
available, or as long as this recognition task has not been
solved significantly.
Each of these steps is described in the following subsections.
4.1 Positive - Negative Recognition
The task here is to detect whether the image is positive or
negative, which can easily be solved by analyzing the grey
levels in the surrounding of the fiducial marks. The idea is
to use the definition of the fiducial mark surrounding which
is given by the templates. After a fiducial mark is located all
the corresponding pixels in the image which are black in the
template are used to calculate a mean grey level, from which
the information whether the image is positive or negative can
be derived.
This approach needs a localization technique which is inde-
pendent from the information whether the image is positive
or negative. Against previous assumptions the approach to
use a binary correlation and only the homogeneity as the cri-
terion for the binarization to first locate the fiducial marks
and then do the grey level analysis, has been proven to not
work reliably enough.
Therefore we now use a grey level correlation on the highest
pyramid level, which will result in a negative correlation co-
efficient in the case the image is negative and the template
positive. For all further steps the much more efficient binary
correlation is used, if the grey level analysis is significantly
solving the positive/negative problem.
