€————— z
such sub-image pair, four patches were found, all
correctly; in the second sub-image pair, eight patches
were found with six being correct and the others being a
neighbouring area of similar colour. The failures were
due to the areas identified by the program as a patch
not really being what we would regard as features.
They were areas of vegetation of numerically uniform
grayscale only slightly differing in colour from
neighbouring areas, where the patch edge is
non-existent to the eye, and imprecisely defined within
the computer.
We tested a second SPOT image pair captured six
weeks apart, using a 500x500 section that contained no
water bodies. Thirty patches were found to match by
our algorithm, but of these, only seventeen were correct.
Again, a large number of the patches were in areas that
we would not identify as a feature, areas where
gray-level gradually changed, giving patches with
visually indistinct edges. Although better results would
have been expected from the shorter elapsed time
between capture of the two images, the lack of distinct
uniform areas other than fields with slightly varying
colour caused the large number of errors.
Our present definition of what makes a patch is clearly
inadequate. We have tried to improve it by requiring
that the uniform area be surrounded by an edge, and
using one of the many edge operators to locate the rapid
change in gray-value. This has so far proved
unworkable, because numerically defined edges are
seldom continuous, and linking the segments possibly
associated with a patch is not easy.
More work needs to be done in this area. An avenue
worth exploring is to require some proportion of the
patch boundary pixels also to be edge pixels, by some
definition. Points of high interest as indicated by the
Moravec or Fórstner operators should be abundant on
clearly visible patches. This should eliminate the
within-field patches which are the main source of error
at this stage of development.
3. From patches to points
The proof of the viability of a patch-based method for
automating registration of a pair of images will be the
success with which matching points can be found in the
images, from the patches. Although we could consider
the centre or centroid of the patch as such a point, any
vagueness in the location of the patch boundary will
translate into uncertainty in the position of the centre.
A better approach will be to look for significant points
along the boundary of the patch and systematically
match them.
In the testing we have done, we have selected all
boundary points as worth searching for. We select a
point on the boundary of a patch in the left image, and
search for its match in the right image, using a search
window. This window is centred on a point in a roughly
corresponding place, related to the centroids of the
patches. We had to choose between using a
least-squares iterative method (Ackermann, 1984;
Gruen, 1985) and a correlation method to find the
matching point (eg, Barnard and Thompson, 1980). We
chose to use a correlation method, because we felt that
the alternative least squares method might be too
susceptible to failure in the initial stages when the
geometry of the matches had not been determined with
much precision, even though it would probably be faster
and may give better accuracy.
Under our chosen method, the correlation between the
selected point in the left image and all points in the
related search area are calculated. The point in the
search area with the highest correlation with the
selected point is chosen, provided that this correlation
is greater than some specified threshold. The match is
subsequently refined to achieve sub-pixel accuracy.
We need to select the size of the search window, and the
size of the area around the point upon which the
correlation calculation is to be based. Although sizes of
correlation areas ranging from 3x3 to 27x27 have been
used by other workers, Shirai (1986) showed that large
windows were generally suitable for obtaining global
range information, but gave smooth changes in
correlation with a broad minimum around the
corresponding point, and consequent imprecision in the
match. Small windows gave a sharper minimum at the
corresponding point, but were more sensitive to noise.
We conducted our own tests, with the results being
checked carefully by hand to find the success in each
case:
Correlation Number of Number
area corresponding points correct
found
3x3 393 90
5x5 362 171
7x7 360 242
9x9 342 195
As can be seen the best results were achieved with the
7x7 correlation area, which we adopted.
The size of the search window will directly influence
calculation time and success. We tested two sized
areas, 11x11 and 23x23. We found the smaller area to
be quite adequate, with no fewer correctly matching
points, and with fewer ambiguities where more than
one point had a high correlation.
3.1 Sub-pixel accuracy
The correlation method as just described should find a
match with an accuracy of one pixel. In practice, the
actual point will more likely be somewhere around this
pixel, since the centre of each one of the corresponding
pixels usually is not the image projection of the same
point in the true object. To get to the point, a window of
size 3x3 centred on the corresponding point is chosen.
Then, a new coordinate value is calculated as a
weighted average of the pixel’s coordinates in the
window, using as weight the correlation value for each
pixel in the window:
= SS. Ui Xi
Xnew 8
Zio Wi
104
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
wh
coc
cor
she
Te
an
Int
1a
As
pai
SP
IM:
pre
the
rat
pai
hac
our
red
WO!
Ove
to ]
In:
we
acc
cor
pat
the
oft
Thi
pri
clos
wel
cor
cok
The
poi
pix