n-1 M-1 -2ni (ot +) 
F(u,v) = NEN X ^  s(n,mje (5) 
n=0 m=0 
with u and v the discrete frequencies in x and y-direction and n,m the image coordi- 
nates of pixel s(n,m) (KOLOUCH et al. 1981). 
Fig.4 shows the original image with its logarithmic power spectrum in fig.5. 
  
  
Fig. 4: Original scene Fig. 5: Power spectrum and filter design 
There are no characteristic maxima on the v-axis which point to a narrow-band scan- 
line noise. 
One can find a relative uniform distribution of the noise over the whole frequency 
axis. Therefore a seperation of noise and signal is very difficult. The same effect 
is shown by the one-dimensional power spectrum of an image column in fig.6. Instead 
of an intensity decrease with higher frequencies there is a plateau-like behaviour. 
Frequency domain filter design and realization 
For the estimation of those parts of the power spectrum which have to be filtered 
channel 9 (A = 810 * 50 um) with nearly no noise, is transformed into frequency do- 
main. Fig.7 shows the v-frequency axis of thermal channel 11 (A = 10 * 2 um) and 
channel 9. 
For detection of scan-line noise in the thermal channel formula (6) is used. 
|F41(0,v)|. 7 |Fg(0,v)| 1 (6) 
F,, and F, are the normalized spectra of channel 11 and 9 of the scanner imagery 
rédpectivdly. The lower part of fig.7 contents the log. ratio of both. There is a 
strong increase of intensity with increasing frequency. Formula (6) is nearly ful- 
filled for all frequencies Fr ~ 0.2 Fy with Fy as Nyquist-frequency according to the 
sampling theorem. 
Using this analysis a narrow-band two-dimensional filter-transfer-function H(u,v) 
around the v-axis is designed with smoothed corners by a sinusoidale mask to avoid 
oscillation while filtering (SFB 149, 1981). Fig.5 shows the filter-position in the 
138 
ERE SACRA ULE re a a 5 50 rm d n 
mu a amar ES - ^ 
P 3 = Ex E 
freque 
ing th 
The hi 
relati 
anothe 
Fig. 6 
Fig. 7. 
For th 
sider ; 
by the 
(n, 
This c. 
compar 
It can 
higher 
Using 
number