Carsten Garnica 
  
1.2 The smoothing problem 
In every digital image there is a certain amount of white noise. If feature extraction algorithms like interest point 
operators or line extractors are applied or if an image segmentation will be done, many of the features to be found, like 
small edge elements, for example, are only expressing the existing image noise. To avoid these annoying effects, which 
only consume computation time and affect the real image features, firstly a filtering process has to be applied to the 
original digital image, in order to get rid of the image noise. 
Then the problem arises, that the smoothing of the image within the homogeneous areas has to be achieved without 
blurring those gray value changes which are containing the information needed. Especially for following feature 
extraction steps it is essential to preserve the edges and corners, with respect to their geometrical shape and magnitude 
of the gradient. 
2 EXISTING SMOOTHING ALGORITHMS 
2.1 Lowpass filters 
The most commonly used smoothing algorithms are the n*n-boxfilter and the Gaussian Kernel Smoothing (Binomial 
Filter). Here, a square-sized convolution kernel is applied to each pixel. With the parameter "boxsize" (n) respectively 
"sigma", the degree of smoothing can be controlled. 
This smoothing strategy results in very nicely smoothed homogeneous areas, however possessing the disadvantage to 
blur gray value edges. This leads to three problems: 
e The edge region will be extended due to averaging of edge and non edge pixel. So the edge may come closer to 
other edges with the effect that they are no longer detectable as separate edges. 
e Corners are rounded off and small crinkles are blurred. This means that image structures are geometrically 
damaged, what might be fatal for the following feature extraction. 
e The magnitude of the gradients is degraded. In the worst case, the edge cannot be detected at all, because the 
magnitude falls beyond some given threshold. 
2.20 Edge-preserving algorithms 
Edge-preserving smoothing filters are much more suitable for feature extraction. Some examples of this filter class are: 
e Median Filter 
e Symmetrical Nearest Neighbour Filter (SNN) 
e Maximum Homogeneity Neighbour Filter (MHN) 
e (Conditional Averaging Filter 
These non linear algorithms are calculating the filtered gray value in dependence of the content of a defined 
neighbourhood. From the list of the neighbourhood pixels, only these are taken for the averaging, which have similar 
gray values compared to the pixel in consideration. Each edge-preserving filter has its own specific algorithm, but they 
all have in common, that the effect of this smoothing strategy is to preserve edges. Unfortunately, these smoothing 
filters have the characteristic not to smooth satisfyingly, because small gray value fluctuations existing in the really 
homogeneous areas are emphasized and not reduced. In addition, the Symmetrical Nearest Neighbour Filter is unable to 
produce reliable results in case of small areas. 
3 DEVELOPMENT OF THE NEW APPROACH 
3.1 Requirements for an algorithm with superior results 
In order to overcome the problems mentioned above an extended algorithmic concept has been designed which has the 
following features: 
e high degree of smoothing in the homogeneous areas 
e preserving of edges, even those ones being represented by small gray value changes 
e conservation of very small homogeneous image regions. 
The algorithmic extension is based on the edge-preserving Maximum Homogeneity Neighbour Filter, because it meets 
the last two demands. The concept of this filter is enhanced, taking the strategy of segmentation techniques like region 
growing into consideration, providing a higher degree of smoothing in the homogeneous areas. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 321