jectral |
SPOT |
MAD
135
0.7
0.9
2.6
Li
1.5
0.7
0.5
0.34
ruction.
ed * and the
trograph image
ry 1000 - 2000
olar radio emis-
6 channels (the
grey level range
ie radiospectro-
cultural type of
sub-scene from
the TM1 band
o xk *
*
ral image Fig.3
located in the
X 0. t
al scene from
* o * *
uction results.
nal forgetting)
th the regres-
getting factor”
of our method
nonstrates iso-
is no need for
with complete
und better es-
nodel parame-
side, on the
multispectral
method | TM | SPOT
MAD | MAD
A 53 2.75
B 49 1.5
C 48 1.9
G 37 0.9
H 33 0.7
H 26 0.4
Table 2: Multispectral line reconstruction.
remaining examples the optimal side is oscillating so single
model method cannot reach performance of the double-model
method G even if repeated on both corresponding sides.
Table 2 shows multi-spectral line reconstruction results. The
regression model is again superior over the classical methods
applied separately on every missing spectral component line.
The multi-dimensional model has v? times more parameters
than the single-dimensional model and so it is more sensitive
to overparametrization resulting in degraded model perfor-
mance. Equation (20) is again used for the optimal model
structure selection, but in this case 1 z 87 so ki # k»
in (21).
8 CONCLUSION
The results of our test are encouraging. The proposed
methods were always the best ones in all our experiments.
The advantage of the regression-type method increases with
an increasing number of correlated spectral bands but even
on monospectral images (the radiospectrograph and SPOT
panchromatic examples) it is also the best one.
Applying the presented reconstruction method in the radio-
spectrograph image reconstruction problem, we obtained im-
ages without missing lines or pixels. These reconstructed
radiospectrograms were successfully used for the evaluation
of the observation of fast drift bursts during the solar activ-
ity. While the former practice was to discard such unusable
radiospectrograms with the unfortunate consequence of dis-
ruption of observation series.
We have not seen any other scheme for correcting image
defects ( off colour stripe, colour gaps, horseshoe effect ) ap-
plying the regression method. The advantage of the present
method | over our previously published regression method
G [Haindl, 1992] with constant exponential forgetting lies
in more precise parameter estimation and consequently the
model prediction - reconstruction quality improvement.
Our method can also be used if more than one of the
monospectral line components is missing. In this case the
reconstruction is done with the multi-dimensional model (H)
with dimensionality v equal to the number of missing spec-
tral line components. Alternatively, the single-dimensional
model can be applied repeatedly to all missing monospectral
lines. Single-dimensional model results are worse than the
multi-dimensional model ones but the model structure opti-
mization is easier. It is also possible to combine the multi-
dimensional model with the directional forgetting concept.
The algorithm can be used to remove scratches as well if it
is applied sequentially on linear parts of a scratch.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Finally if the method is used for isolated image pixels recon-
struction then the predictor and similarly the model probabil-
ity expression do not need any data approximation and the
regression method performs better than for line reconstruc-
tion and much better than any of the classical methods.
The proposed method is fully adaptive, numerically robust
and still with moderate computation complexity so it can be
used in an on-line image acquisition system.
REFERENCES
[Bernstein, 1984] Bernstein, R., Lotspiech, J.B., Myers,
H.J., Kolsky, H.G., Lees, R.D., Analysis and Processing
of Landsat-4 Sensor Data Using Advanced Image Process-
ing Techniques and Technologies. IEEE Trans on Geosci.,
GE-22(3), pp. 192-221.
[Broemeling, 1985] Broemeling,L.D. Bayesian Analysis of
Linear Models. Dekker, New York.
[Haindl, 1992] Haindl,M., Simberová, S., A Multispectral
Image Line Reconstruction Method. In: Theory & Applica-
tions of Image Analysis. P. Johansen, S. Olsen Eds., World
Scientific Publishing Co. Singapore.
[Haindl, 1996] Haindl,M., S$imberová, S., A high - resolution
radiospectrograph image reconstruction method. Astron-
omy and Astrophysics, Suppl.Ser., 115(1), pp. 189-193.
[Hàgglund, 1983] Hàgglund, T. The problem of forgetting
old data in recursive estimation. Proc. IFAC Workshop on
Adaptive Systems in Control and Signal Processing, San
Francisco.
[Kulhavy, 1993] Kulhavy, R. Zarrop, M.B. On a general con-
cept of forgetting. Int. J. of Control, 58(4), pp. 905-924.
