International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
achieve a good estimate of both. The CMOS image of such a pop MR TY 
target is shown in Figure 4. an 
T 
  
  
  
  
Figure 8. Fitting a rectangle on Wiener profile 
  
Now, by fitting a proper 1-D Gaussian to the PSF profile, the 
Figure 4. Digital CMOS image of an ideal target oscillations of PSF will be reduced. The total PSF of the system 
will be achieved by expanding the 1-D Gaussian function to the 
The second step is to select a patch of one of the individual 2-D Gaussian function: 
  
         
  
targets (rectangles), that includes two vertical edges (Figure 5). RR QUE gare 
: : $ 
— 25 i 03 il 
Batch = is 03? ji 25 |l 
sf | | A 02 | | 
= 
         
      
  
E 
2 
| Profile 
| Selection 
         
        
  
  
| 
Pall, 
LAT ANA! 
AAW 
0 Bo 
dle), fitted 
Nn ES es wT 46. 907 76 E rs 
Figure 9. PSF profi ), Gaussian function (mid 
Gaussian function to PSF profile (right). 
OT 
    
Figure 5. One of the edge targets (left) and corresponding 
picked up patch 2.4 The Wiener Deconvolution Filter 
The third step is to reduce the noise in the above selected image By using the calculated PSF and noise to signal power ratio (K), 
patch by applying a low-pass filter, for example an averaging the Wiener deconvolution filter can be calculated by using 
filter with a 5x5 window (Figure 6). Equation 4. The size of this filter should be expanded to the size 
of the image patch, which is to be restored (Stockam, et. al, 
  
  
  
  
  
  
AA 1975). 
+10 ; 
on [3 2.5 Experiments and Results 
a J 1d 
@ | | : : 
m f^ To evaluate the proposed restoration process, the estimated PSF 
2e — is used for radiometric correction of an outdoor image (Figure 
Figure 6. Image (left) and signal (the data without noise) 10). Visual comparison of the obtained result with the input 
profiles (right). image demonstrates the capability of the applied restoration 
strategy. 
By subtracting the result of averaging filter from its 
corresponding image patch, the noise profile is achieved, as 
shown in Figure 7: 
  
  
  
SOE Noe pee EN 
| Subtracting NIE: V A 
Signal 38 | | 
| Averaging 3 \ | 
EIE IAS eh 80 
  
  
  
Figure 7. Left: Signal profile (result of lowpass filter), Right: 
Noise profile. 
  
Now, a Wiener filter should be applied to the image patch to a 
create a profile, which just contains the system blur effect. By 
fitting a rectangle (Figure 8), which indicates the ideal signal 
profile, on the profile of Wiener filter, the PSF of the CMOS 
system can be calculated in frequency domain by the 
expression: 
H LISD) 
Where H(orr,is OTF (Fourier transform of the 1-D PSF), 
(5) 
    
FT(SP)is Fourier transform of the signal profile and FES Sol cra 
FT(RP) is Fourier transform of the rectangle profile. Figure 10. Initial image (above) and the restored image 
generated by Wiener deconvolution (below). 
    
  
   
   
  
   
  
  
   
    
    
   
   
   
    
   
    
    
    
  
   
  
    
  
   
    
   
  
  
   
Jb S PRA 
—