a 
KP le SV 
- 
Q =. 0 N 
Carsten Garnica 
4 RESULTS 
4.1 Preserving of image structures 
The new algorithm preserves the geometric shape of all image structures as well as the original MHN Filter does. 
Figure 2 compares the effect of some smoothing filters on a gray value edge with a low signal/noise-ratio. The effect of 
the Gaussian Kernel Filter is a blurring of the edge, while the two other algorithms preserve the edge significantly. 
  
  
  
  
  
  
  
  
  
  
200 + = 
Original 
190 | A 
- Gauss Kernel 
180 } Y 4--- MHN 
+ 1 | 
- / ——R—— New Approach 
il ] 
/ —— . 
170 / ~ | 
4 rs ~ i 
Li 
= 
d —— 
0 15 20 
Figure 2. Effect of smoothing filters on a gray value edge 
The conserving effect for important image structures is to be found in Figure 3 too. Here images are displayed after 
different smoothing filters having been applied. It is obvious, that the geometric shape of all images structures is 
preserved in the cases (c) and (d). Comparing with the result of the Gaussian Kernel Smoothing (b) a blurring of the 
edges and an off rounding of the corners is avoided. 
  
    
     
  
a) original image 
  
| b) image filtered with 
Gaussian Kernel, 
sigma = 2.0 
  
  
  
c) image filtered with 
MHN 
  
  
  
d) image filtered with 
new extended MHN 
approach 
  
  
  
Figure 3. Resulting images after different smoothing filters applied 
4.2 Noise cleaning 
With respect to noise cleaning, the degree of smoothing of the new algorithm is considerably higher than for 
conventional edge-preserving algorithms. For Gaussian Kernel Smoothing, the degree of smoothing depends on the 
adjustment of the sigma threshold. Figure 4 shows a comparison of the calculated gradient magnitude that remains after 
the different filters have been applied. Obviously, the Gaussian Kernel smoothing (b) produces a high degree of 
smoothing, but the strength of the gradients representing the real edges has been degraded. The edge-preserving filters 
(c) and (d) conserve the edge gradients. It can be seen, that the new strategy (d) leaves a significantly smaller amount of 
spurious gradients in the homogeneous areas than the conventional MHN (c). 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 323