ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
3.3 Multi-sensor Image Registration
Fig.3 (a)-(b) shows a pair of SPOT and SAR images. 20 pairs
of control points are manually selected to register the two
images, and its result is shown in Fig.3 (c). The result by our
approach is shown in Fig.3 (d). It is found that it is very
difficult for us to select control points in this pair of images
probably due to their large differences in radiometric. From
Table I, we also know that differences of the transformation
parameters from the two methods are large. The RSME
between them is more than 3 pixels and the max D between
them is even more than 7 pixels. In order to test the two
methods, we compare two pairs of 64x64 sublimages which
are cut at the adjoins of two registered images. The two pairs of
sub-images are shown in Fig.3 (e)-(h) scaled 4 times to be
clear. From the first pair of images in Fig.3 (e)-(f), we find that
there should be a road crossing through the center of images
from the left-bottom corner to the right-top corner, but there is
a distinct jump in the image Fig.3 (e) which is cut from the
image registered by manual methods. In the second pair of
images in Fig.3 (g)-(h), there is a place connecting land with a
river. Although it is not very precise to compare the results
from the two methods owing to changing water level of the
river, we find that the link manner in Fig.3 (h) by our approach
is more reasonable than in Fig.3 (g) by manual methods.
Therefore, our approach is better than the manual methods.
Table I
Comparison Of The Global Optimal Registration Results With The Manual Registration Results Or With Truth
Test Data Methods a, a a, by b, b, Energy RMSE max D
Optimal 0.00000 0.00000 1.00000 511.0000 | -1.00000 0.00000 51.78687 0.00000 | 0.00000
Synthetic Initial 0.77563 -0.01030 | 0.99921 515.3251 -0.99921 -0.01030 34.13717 3.53388 | 6.51936
Images Manual -2.15291 3.01e-05 1.00452 512.8174 -0.99897 | -0.00406 36.89599 1.70201 3.17100
Our method | 0.00099 6.18e-06 0.99998 511.0105 -1.00001 | -7.39e-06 | 51.04321 0.00834 | 0.01258
Multi- Initial 8.04031 0.96688 -0.23845 | -120.3902 0.23845 0.96688 23.95547
Temporal Manual 6.66587 0.97010 -0.24062 | -119.9549 | 0.23776 0.96976 25.63776 0.81365 1.69900
Images Our method | 5.95431 0.97262 -0.24227 | -121.1075 0.24040 0.97059 26.64661
Multi- Initial 115.9111 0.92227 -0.39794 | -86.73397 0.39794 0.92227 49.43873
Sensor Manual 112.591 0.92473 -0.39309 | -83.58278 0.37902 0.02507 49.54566 3.82170 | 7.31851
Images Our method 111.605 0.91877 -0.39289 | -86.18712 | 0.39607 0.92361 49.98123
4. CONCLUSIONS
In this paper, we propose a global optimal image registration
method. In our method, we develop a new strategy in which a
global mapping function is estimated by a few local control
points, but acquires the mapping function in the whole image
range. Therefore, the registration accuracy of our method is
much higher than that of conventional methods. In our method,
at first, we define an energy function that is directly related to
parameters of the mapping function, and thus an estimation of
the mapping function is translated into an energy optimization.
On defining the energy function, we do not use similarity
measures that are sensitive to radiometric distortion, but exploit
the average edge strength that can describe structural features
and shapes of scene. Therefore, our approach is not only
applicable for registering images acquired from different
sensors, but also for images acquired on different dates in
which there may be big radiometric differences between the
images because of variations in solar illumination, atmosphere
scattering, and atmosphere absorption. Second, we present a
hybrid scheme combining a SM and GAs sequentially to
optimize the energy function: firstly, a statistical method is
used to acquire a set of rough initial guesses for each parameter
in the whole images, and then GAs are exploited to search
further precise guesses of parameters from many sub-images.
Finally a SM is employed to gain the global optimal parameters.
One advantage of the hybrid scheme is that it is not easily
entrapped in local optima, and converges fast.
In our method, we avoid exploiting advanced feature extraction
and feature matching techniques. Thus, our approach
successfully avoids the two inherent difficulties faced by
existing methods. Therefore, our algorithm is robust and
automatic.
The experimental results from our method have been compared
with the ones by manual registration methods, and it is
demonstrated that our method is very efficient and effective.
Meanwhile, the energy function derived in this paper can be
also regarded as an assessment criterion for the image
registration.
S. REFERENCES
[1].Fonseca, L and Manjunath,B.,1996. Registration techniques
for multisensor remotely sensed imagery. Photogram.
Engineering & Remote Sensing, 62(9), pp.1049-1056.
[2]Dai,X, and Khorram,S,1998. The Effects of Image
Misregistration on the Accuracy of Remotely Sensed Change
Detection. [EEE Trans. Geosci. Remote Sensing, 36(5),
pp.1566-1577.
[3]Rignot,E, 1991. Automated multisensor registration:
requirements and techniques. Photogramm. Engineering &
Remote Sensing,57(8), pp.1029-1038.
[4]Li,H, Manjunath,B and Mitra,S, 1995. A contour-based
approach to multisensor image registration. /EEE Trans. Image
Processing, 4(3), pp.320-334.
[5] Renders,J and Flasse,S, 1996. Hybrid methods using
genetic algorithms to global optimization. /EEE Trans. System.
Man Cybern., 26(3), pp.243-258.
[6] Lagarias,J,1998. Convergence properties of the nelder-
mead simplex method in low dimension. SIAM Journal on
Optimization, 9(1), pp.112-158.
[7] Goldberg, D, 1989. Genetic algorithms in search,
optimization and machine learning”, MA: Addison-Wesley.
[8] Canny,J, 1986. A computational approach to edge detection.
IEEE Trans. Pattern Analy. Machine Intell., 8(6), pp.679-698.
6. ACKNOWLEDGMENT
The work is supported by the Post Doctoral Fund of China.
A - 399
