Table 1
COMPARISON OF PRECISIONS OF MATCHING OBTAINED BY LEAST SQUARES
AND CROSS-CORRELATION METHODS
Feature
Spread function 25.0 um
Pixel size 12.5 um
Quantisation level
Cross 100 x 10 Um
8 bits/pixel.
Precisions in terms of pixel size x 1000
Method of Scale Variations Rotation (degrees)
matching
1.0 0.9 0.8 0 10 20 30
LSQR 38 47 42 38 31 47 180
Cross- 52 75 120 52 83 170 *
Correlation
* - no results
Table 2
COMPARISON OF MATCHING PRECISIONS OBTAINED FOR DIFFERENT SHAPED FEATURES
FOR THE LEAST SQUARES METHOD
Pixel size
12.5 um
Precisions in terms of pixel size x 1000
TYPE
OF
FEATURE
SCALE
CIRCLE
CROSS
ELLIPSE
* - no results
Table 1 reveals that the cross-correlation method, for images
in which there are no scale distortions or rotations, will result
in marginally worse results than the least squares method.
However, as the scale distortions and rotations increase, the
precisions obtained by cross-correlation rapidly deteriorate and
are a factor 3 to 4 larger than those obtained by the least
squares method, while matching fails for a rotation between
the two images of 309. This result confirms the widely held
views amongst photogrammetrists that the cross-correlation
method will be significantly affected by distortions between
the two images, and indeed may result in completely erroneous
matches.
In Table 2 are shown comparisons between 3 different
features subject to scale variations and rotations, and different
levels of blur. As previously demonstrated, this table
confirms that better results are generally obtained for a cross
than for a circle or ellipse, and also the small influence that
image quality has on the precision of least squares matching.
Foerstner (1982) has given a formula for the theoretical
precision of matching 2 digital images by the least squares
method as follows:
c?
aad rr
X © N.SNRZ c.
8
where o2 is the variance of estimating the shift
822
ROTATION
(degrees)
10
SPREAD FUNCTION
(um)
30 10 25
* *
44
36
50
38
32
48
31
68
parameter in the matching,
N is the number of pixels containing the
information about the feature,
SNR is the signal to noise ratio in the image,
o; is the variance of the signal,
o?, is the variance of the gradient of the image.
8
Adopting N=50 for a circle of 100um with pixel size of
12.5um, SNR=10, o2 -3600, and o?
g g' =900 for an image
quantized to 8 bits, Ox is calculated to be 0.04 pixel, as
indeed has been obtained for the feature in Figure 1.
Similarly, for a larger circular feature of 200um, the precision
will be a factor 2 less or 0.02 pixel which also agrees with that
shown in Figure 2. As stated above, a blurred image will be
larger than a sharp image of the same nominal size because of
the effects of blur, and therefore, the factor N of a blurred
image in equation 1 will be larger than for a sharp image,
while G,' will be slightly smaller. This appears to ex,
the better results obtained for blurred images than for shi. p
images. These results demonstrate an agreement between the
theoretical precisions and those obtained by simulation.