This
lica
any
loes
vith
ally
not
not
ard
Axis
and
t is
the
int,
ttle
her
uch
bly
the
the
"his
ich
the
"his
nes
nes
but
Lise
EM
ing
tly
and
le.
| of
1S
Correlation may be applied following a regular grid or along those points found to be
proper for correlation by an interest operator, e.g. the MORAVEC-algorithm (6).
The selected points may be checked by a human operator and then measured by
means of least squares correlation. While fracture surfaces tend to be very rough,
the high accuracies and correlation coefficients given in (1) were not obtained, but,
compared to cross correlation, the least squares correlation technique yielded a
distinct improvement in accuracy of parallax measurement.
The decision about the procedure (grid or interest operator) depends on the
topography and the texture of the object surface as well as on the quality of the
image data. With the SEM imagery this investigation is based on, the use of a grid
proved to be of advantage. The MORAVEC-algorithm either yielded not enough
points to ensure accurate interpolation or it turned out to be too costly if the
corresponding point quantity was demanded.
Because the resampled images are free of y-parallaxes, they can be processed by
automatic one-dimensional correlation. First, approximate values for the homologue
points are calculated by using the normalized cross correlation. With SEM data
correlation coefficients frequently are less than 0.7, which is in general not
considered a good matching result. But in most cases corresponding points are
certainly found (fig. 4).
SS OS
SERS
CS
em SS
ORO
T S
CSSS
EDS
ses
NES
(SS
Fig. 4: 3D representation of cross correlation coefficient (left)
and auto correlation coefficient (right)
In a second step a least squares algorithm is applied to determine the x-parallaxes of
corresponding points with increased accuracy. Up to now an affine transformation
has been used in this algorithm, but regarding the complex geometry of fracture
surfaces the application of higher transformations seems to be justified.
135