ne XXXIX-B3, 2012
at the algorithm is termed
ix for a large number of
he RANSAC algorithm is
iaining data points, IDS
| they are regular samples
the difference in theory,
complementary.
flexible function in that it
rgy of a metal plate on
id surface stems from a
wo overlapping images.
ie transformed line and
ate discrepancies at m
ed as x =,
and
|, respectively. For an
ivolves both trend and
concept of a remove-and-
therstone, 2009).
x Is defined as
n)
m)
(6)
listance between points j
1e so-called fundamental
and takes on the form
r consisting of weight
- zi
K 6x and w,=K óy,
ions.
the m tie points, a vector
) . In association with the
or is employed for an
discrepancies, as follows
(7)
end function have to be
X, and y, - dy; , in order
and resampling for the
ter, the difference in
form and the thin-plate
d.
AGE TESTS
-590 nm) and Formosat
ain rural area (Figure |;
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
400 x 400 and 600 X 600 pixels, respectively) are chosen for a
test on image registration. SIFT could only provide faulty tie
points because of stark differences in the gray-level appearance,
especially of farmland.
SPOT-5 (green band)
COENA
© 2006 NSPO
Master Slave
Figure 1. Distribution of 119 edge-detected, cost-optimized and
LSM-matched conjugate tie-points, after removal of blunders by
IDS at a 5 percent significance level.
Figure 2. Registration between the dark SPOT-5 green and
bright Formosat-2 infrared images, in a regularly gridded
format; to the right, a local aquiculture pond area is enlarged.
Still, an alternative technique depends on three to four manual
seed points in an initial affine transform to align the
experimental images. One detects linear edges by the Canny
operator. Point correspondence is then established by the least
cost resulting from a function that is made dependent upon
between-point distance and between-edge angular error.
Potential homologous point pairs serve as input to LSM in
terms of Equations 2-3, where the least-squares IDS is
activated to safeguard against likely blunders. The final
matching result is displayed in Figure 4. The TPS algorithm
utilizes the available tie points to warp the Formosat-2 Level-
IA 8-m slave image onto the geocoded SPOT-5 10-m master
image. Figure 2 displays the warped result, employing alternate
square subimages for visual impression.
As far as the accuracy of image registration is concerned, 57
point pairs were chosen randomly for checking. With 4
peripheral corner points serving as basic control, there were in
total 62 tie points in controlling image warping. The computer
program was executed for a series of 30 trial runs, and the
averaged distance RMS error amounted to 0.50 pixels, as seen
in Figure 3. Accuracy difference between the affine transform
and TPS warping results that passed a significance test could be
partly due to an uneven point distribution.
-- Affine transfonm, 0.54-pixel mean
6.60 Toe Say
«TPS warping, (.50-pixel mean
/
Vn
¥
A f
fi d
f * À j A i À NA
pei
in eed
j
ud
À
4
Zz
ü30 + "+
Distance
RMS error {pixel}
4 8 12 16 OU 24 ^ 28
Times for random sampling
Figure 3. Comparison of registration accuracy between
affine- and TPS-based methodologies, during which there
were 57 different check points each time. The RMS values
are different, and their difference is significant, by a Fisher
test at a 10 percent level.
3.2 IKONOS and QuickBird images
An urban sector near a temple in Taipei and its meter-level
panchromatic 1A satellite images are illustrated in Figure 4.
The 400 X 400 pixels IKONOS master image was scanned with
1.0-m ground sampling distance on February 21, 2002. The 800
X 800 QuickBird 0.6-m resolution slave image was taken on
December 15, 2002. In total, there are 84 matched point pairs.
IKONOS
1.0 m resolution
33 * S
© 2002 DigitalGlobe
Master Slave
Figure 4. Satellite images from different viewing angles with
84 image tie-points determined by employing an algorithm
that involves the SIFT, RANSAC and LSM techniques.
The TPS-structured method is employed to co-register the
IKONOS and QuickBird images. Part of the registered
QuickBird image to the IKONOS image is displayed in Figure
5. In close inspection, roof edges are continuously linked across
the boundary. As usual, the effect on registration accuracy of
randomly distributed 44 control and 40 check points is
documented in Figure 6. Since all conjugate points lie on
rooftops, it appears that an affine relationship between the space
images can accommodate geometric distortion very well.
