Estimation of image SNR 
The crucial point of all investigation is to find a good SNR estimation because normally 
intensity and distribution of noise is unknown. So simplifying assumptions must be made 
to derive SNR values from noisy image signals. 
A posteriori SNR estimation can be done using the values of maximum correlation or 
grey level differences (FOÓRSTNER, 1882: TRINDER. 1882). But doing so, we cannot 
avoid incorrect correlation. Therefore we need an a priori noise estimation. 7 
We assume that all superposed disturbance effects can be regarded as white noise n, 
i.e. having a constant power spectrum 
PLlfl = Ng (8) 
Pn is the noise power spectrum, Ng the constant intensity. A typical power spectrum 
Pg of image and noise is shown in fig.10. 
Intensity 
À ] 
\ 
N 
JONNY ? 
N P (f) 
SS ~ 
7 
PSS 
  
  
  
  
> 
rdi 
  
  
Y 
Fig.10: Power spectrum af image and noise 
The variances c^, for noise can be estimated Dy 
Go cà can be estimated in a representative image part according to (8). This cj is con- 
sidered to be valid for all correlation windows. Image and noise can be regarded as in- 
dependent stochastic variables. Hence the variance ca of the undisturbed signal g can 
be calculated as difference of cg , the variance of the disturbed signal g'. and of. 
ca = (7 2 - 2 =” 2; - 2 
eq eg On ed Ng 
So simple variance computation in a correlation window yields the corresponding SNR: 
rtm 
lg? , = NZ 
= — d 0 ; 
SNR = | — (9) 
N 2 
A 
: J 
Equation (9) is used to estimate the a priori probability P to avoid wrong correlation.