areas 
>-error 
metric 
mage, 
pixels 
ulated 
is the 
t from 
found 
etween 
d area 
ice the 
to the 
re four 
he im- 
s. The 
els dif- 
distant 
20 dis- 
alue of 
= 017 
andard 
an im- 
nce for 
s don't 
ast one 
of the 
several 
iore the 
orithms 
sting of 
 follow- 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
TYPE | CONT || EPOS | GEOM | R-LEE [ E-FRO | FRO | KUAN | LEE | E-LEE | G-MAP | MEAN 
POINT | 100 0.29 4.60 | 3.54 673 | 676 | 330 323] 3% 3.74 | 16.02 
POINT | 80 0.26 4.62 | 6.80 757 | 15^] 378 | 381 | 4.26 4.96 | 10.48 
POINT | 60 4.44 4.63 | 5.86 5.31531] 392] 335] "425 4.48 6.11 
POINT | 40 2.81 3.06 | 2.85 268 | 2608] 298 " 280| 277 2.89 2.89 
POINT | MEAN 1.95 423 | 4.76 5.59 | 559] 345] 345] 363 4.02 8.88 
LINE 100 2.49 | 13.70 | 10.58 | 10.40 | 10.40 | 1144 | 11.46 | 12.80 | 12.95 | 150.00 
LINE 80 2.61 | 14.03 | 12.23 | 12.52 | 12.52 | 12.48 | 12.55 | 12.98 | 14.99 | 96.47 
LINE 60 21.75 | 14.81 ] 1748 | 20.11 | 20.11 {. 12.81. [1234 |=13.00.{ 1721] 5487 
LINE 40 23.87 | 14.72 | 22.84 | 15.96 | 15.96 | 11.46 | 11.35 | 12.66 | 16.80 | 25.01 
LINE MEAN || 12.68 | 14.31 | 15.78 | 14.75 | 14.75 | 12.05 | 12.05 | 12.86 | 1549 | 81.59 
AREA | 100 1.18 2.97 | 4.63 6.11 | 6.12 | 11.50 | 11.76 | 1148 | 1589 | 56.42 
AREA | 80 1.18 2.93 | 420 628 | 629 ^ "9954-1025 | 9.57 | 1465 | 3705 
AREA | 60 3.26 2.76 | 4.46 6.90 | 6.90.1..9.18 ]- 8.30-|.:... 7.74}. 12.39)... 21.45 
AREA | 40 4.43 2.32 0> 3,52 5.68 | 5.68 | 5.84 | 5.86 | 5.95 7.75 | 10.42 
AREA | MEAN 2.51 2.74 | 420 5.24 | 625 |. 8.87] 9.04] 665] 1267 | 313 
[MEAN | MEAN [| 571] 710] 825] 8.86 | 560] 812] 815 | 8.30] 1073] 40.60 | 
[RATED | MEAN [| 491] 601] 724] 820] 821] 831] 839] 847] 11.21 | 38.28 ] 
  
Table 2: RMS-error for different geometric primitives and contrast levels. 
5 RESULTS 
The basis of the comparison is the similar reduction of speckle 
variance in homogeneous areas for all filters. As mentioned in 
Section 4.1 this task is not easily solved, since a continuous 
parameter is not available to adjust the smoothing capability. 
The attempt to adjust all filters at a similar speckle reduction 
near R = 0.1 results in the values given in column two of 
Table 3. One should notice that especially the filters LEE, 
KUAN and G-MAP are adjusted to a quite less smoothing 
capability, thus geometric distortions may be less for those 
filters. 
  
  
  
FILTER | REDUCT | M-ORIG | M-SPEC 
MEAN | 0.0963 -0.936 0.000 
E-FRO | 0.1065 -0.706 0.230 
GEOM | 0.1096 2.468 3.404 
EPOS 0.1122 -0.874 0.062 
FRO 0.1126 -0.458 0.478 
E-LEE 0.1143 -0.438 0.498 
R-LEE 0.1165 -1.460 -0.524 
LEE 0.1309 -0.950 -0.014 
KUAN 0.1328 -0.440 0.496 
G-MAP | 0.1363 -0.578 0.358 
  
  
  
  
  
  
Table 3: Speckle reduction R and mean retention 
The capability of mean retention is also shown in Table 3. 
Column 3 contains the difference of the mean within the fil- 
tered image and the original greylevel. Since the MEAN filter 
also shows a difference which is obviously caused by the noise, 
we subtract the value of the MEAN filter in column 4. The 
best values are obtained for the LEE and the EPOS filter, but 
most filters cover an uncritical range of 0.5 greylevels. Only 
the GEOM filter changes the mean in the homogeneous area 
by more than 3 greylevels, what is regarded as insufficient. 
The most important criterion for the comparison, the distor- 
tions of the geometric primitives, is measured as RMS-error 
between the original, unspeckled and the filtered image. The 
139 
results for various measurements are shown in Table 2. The 
first column denotes the type of the geometric primitive. The 
second column denote the radiometric contrast within the 
test image, representing different signal to noise ratios. For 
each geometric primitive one line shows the MEAN calculated 
from all contrast values. The last two lines show the mean 
of all geometric primitives, in the last line a rated value is 
shown where areas are weighted by 1/2, points and lines by 
1/4 according to the significance of areas as considered in 
Section 2. The filter algorithms in Table 2 are sorted by the 
last line. The mean filter is also included in the table for a 
comparison. 
  
16 I T I T T T T 
  
  
  
  
@ 
o 
2 
o 
> 
o 
9, = e 
t = J 
O 
oc 
tr 
ui J 
o) 
z 
tr 
V 
v a 
\ m= 
Ye-------. Xe oe" 
Rated Mean -8— 
Areas -4-- à 
Lines -x--- 
Points -x--- 
1 1 1 1 1 À 
  
0 1 
EPOS GEOM R-LEE E-FRO FRO  KUAN LEE E-LEE G-MAP 
Figure 2: Distortions at geometric primitives (CONT = 100) 
To illustrate the contents of the table, two diagrams are plot- 
ted from the lines of the table. Figure 2 shows the RMS-error 
of different filters for the geometric primitives at a contrast of 
100 greylevels. It is significant that the EPOS filter perform 
much better than the other ones in all geometric disciplines 
for this high contrast. The RMS-error for different contrast 
values is shown in Figure 3 for areas. It is interesting that a 
reduction of the error with an increasing contrast is observed 
only with the EPOS filter. This reduction may be expected, 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996