The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
335
5. RESULTS AND SUMMARY
Table 1 and Table 2 specify the respective results of classifying
sub-images as fragments being either "favourable" or
"unfavourable" from the point of view of their subsequent
matching, as well as results of matching sub-image
representations on two successive aerial images. Results
presented in Table 1 prove that the use of neural networks for
the preliminary selection of images is advisable on condition
that proper image representations have been applied. It is
possible to additionally increase the recognition percentage
through the use of the rejection technique, which consists in
testing the recognition reliability, and through the proper
selection of the threshold value for that reliability. Results
presented in Table 2 (the third column) were obtained upon
having assumed the boundary value of shift, below which
matching can be considered to be acceptable. That value,
amounting to 30 pixels, was assumed based on results of
matching whole (non-processed) sub-images by means of area
based matching method (see last line in Table 2). That was just
the maximum error of matching by means of that method. In
turn, the last but one line in Table 2 shows small possibilities of
direct use of the representation, which is based on a 2D
histogram of gradients. The best results were obtained as
regards representations in the form of signatures generated by
the ICM network (ca. 91% matching).
Description of
representation
Data set
description
Result
Publications
and
comments
outputs of neurons
of Kohonen's
SOM layer
36 pairs of
images
(the G
component)
10/36
(Czechowicz
et al., 2007a)
100-element
signatures,
6 signatures per 1
sub-image
44 pairs of
images
(in greyscale)
40/44
(Mikrut
et al., 2008)
19-element profile
of gradient
directions
36 pairs of
images
(the G
component)
2/36
(Czechowicz
et al., 2007a)
whole sub-image
- area based
matching
as above
36/36
(Czechowicz
et al., 2007a)
Table 1. Results of recognition of 2 image classes
Description of
representation
Type and
structure of
neural
network
Recog
nition
of test
set
Publications
and comments
2D histogram of
gradients:
profile of
average gradient
direction values
with the
aggregation of
20°
Kohonen's
SOM
19-28(7x4]
89.5%
(Czechowicz
et al., 2007)
rejections
were applied -
responses
from 6
neurons were
taken into
account
as above
SOM + BP
19-(7x4]-5-2
84.5%
(Czechowicz
et al., 2007a)
as above, but
with the
BP
73-10-2
84.8%
(Czechowicz
et al., 2007a)
aggregation of 5°
log-Hough space
- vector of
projections upon
the radius axis,
aggregated at
every 2
BP
25-2-2
25-28-2
72.5%
69.2%
(Piekarski et
al., 2007)
50-element
signatures
generated by the
ICM network
BP
50-6-2
73.0%
(Mikrut, 2007)
- at -20% of
rejections the
recognition
increased up
to -80%
Table 2. Results of image matching experiments
In the case of lower, acceptable distances of deviations
amounting to 10 and 5 pixels, the percentage values of correct
matching amount to ca. 63% and 38%, respectively.
The results obtained in the course of research works confirm the
original assumptions, though it needs to be stressed that the
issue is a complex one, and the research subject far more
extensive than originally assumed. The matched aerial images
are marked by the changeability of objects, textures and
illumination. It was mainly those features which, combined
with too poor initial processing of images, resulted in the
obtaining of not fully satisfying effects. Summing up, it is
necessary to state that the application of neural networks
succeeded in the case when a preliminary selection of sub
images had been performed.
The problem of matching requires probably another approach to
be followed, perhaps the one based on detection and description
of objects. The authors believe it is reasonable to continue
research with the use of algorithms that have been modified in
that way, but with a similar methodology.
REFERENCES
Arbib M.A. (ed.), 1995. The handbook of brain theory and
neural networks. Cambridge, The Mit Press.
Barsi A., 2000. The impact of data compression and
neighborhood information on the classification accuracy of
artificial neural networks, International Archives of
Photogrammetry and Remote Sensing. Amsterdam., Holland,
Vol. XXXIII, Part B3.
Bishop. C.M., 1996, Neural Networks for Pattern Recognition.
Oxford, Clarendon Press.
Czechowicz A., Mikrut Z., 2007. The use of Kohonen network
to the selection of images for automatic matching processes on
aerial images. AGH UST Automatics, 11, pp. 329 - 343 (in
Polish).
Czechowicz A., Mikrut Z., 2007a. Selection of sub-images for
aerial photographs matching purposes based on gradient
distribution and neural networks. Archives of Photogrammetry,
Cartography and Remote Sensingi, Vol. 17, Dobczyce (in
Polish).
Duin R.P.W., 2000. PRTools Version 3.0 - a Matlab Toolbox
for Pattern Recognition. Delft University of Technology,
http://www.ph.tn.tudelft.nl/prtools (accessed 1. Sept. 2007)