Yunham Dong 
  
3 RESULTS 
To evaluate the performance of the proposed filtering algorithm, we have to follow the criteria for a speckle filter. As 
indicated above, the most important ones include mean preservation, reduction of standard deviation, edge preservation 
and texture preservation (restoration). 
3.1 Mean Preservation 
Because this is an adaptive mean filter, it guarantees the preservation of the mean. 
3.2 Standard Deviation Reduction (Smoothing Uniform Areas) 
As indicated, most existing speckle filtering algorithms can effectively reduce the speckle level, and this algorithm is no 
exception. Meanwhile, suppression of speckle is somehow at the cost of losing local details. For example, using a larger 
moving window always reduces more standard deviation but also smears local detail. A comparison of smoothing 
uniform areas should be associated with the comparison of preservation of edges. To avoid such complex comparison, 
we assume that this algorithm be used in conjunction with other speckle filtering algorithms. In this way, it only needs 
to examine if edges are sharpened. The following subsection looks closely at the edge enhancement. 
3.3 Edge Sharpening and Enhancement 
Shown in Figure 4 is a 512 x 512 AirSAR C-band HH image of Alligator River of Northern Territory, Australia, 
acquired by the NASA/JPL AirSAR system in 1996. The image is filtered using various filters including the Lee filter, 
Kuan filter, Frost filter, mean filter and the proposed filer. Results are compared using numerical analysis for, in 
particular, two areas, each consisting of a 40 x 40 pixel block, one representing a uniform area and one an edge crossing 
area as shown in Figure 4. 
  
Figure 4. A 512 x 512 pixel NASA/JPL AirSAR CHH image. Indicated are two 40 x 40 pixel blocks representing a 
uniform area and an edge crossing area, respectively. 
To evaluate the performance of a filter, one can analyze data in the frequency domain. In order to view frequency 
component changes, each block of data is treated as one-dimensional data with 1600 pixels consisting of consecutive 
lines of pixels in the block. The DFFT is then used for data analysis. Figure 5 shows differences of frequency 
components between the original data and the filtered data for the uniform area. It can be seen in the figure that 
behaviors of the four filters, i.e., the mean filter, Lee filter, Frost filter and the proposed filter, are very similar for 
uniform areas, as expected. The difference in frequency components to be positive indicates the frequency components 
have been reduced, thus the data in spatial domain have been smoothed. It is worth noting that each figure is 
symmetrical to the center, and the frequency components spread from high at the center to low at two ends. The Kuan 
filter is nearly identical to the Lee filter, and its result is therefore not shown in the figure. 
  
92 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 
Perf 
The 
in e 
muc 
sme: 
the c 
Figu 
perf 
Figu 
perf: 
Detz 
Cros: 
filte1 
data 
and 
34 
Text 
infor 
of tl 
base 
mati 
It h: 
mea: 
cons 
algo 
imag 
for e 
rem 
is th