Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 228

A 1 2 3 4 B
A 1 2 3 4 B
A 1 2 3 4 B
A 1 2 3 4 B

A 4 3 2 1 B
A 4 3 2 1 B
A 4 3 2 1 B
A 4 3 2 1 B

(a) Image (b) Portion of (a) with the

geodesic distance from
pixel of the stripe to basin
A

(c) Portion of (a) with the
geodesic distance from
pixel of the stripe to basin
B

(d) Result

Fig. 13. An image, the geodesic distances from pixel of the stripe to the different basins and
the segmentation result using the Vincent-Soille algorithm

computation, since its cost is O(N 2 ), where N is the number of pixels in the image.
The original version of Roerdink´s and Meijster´s algorithm is not quite correct.
Therefore, Andreä and Haufe give a corrected version in [12].
As we already pointed out, watershed segmentation performs better when using the
gradient image of the input image. During our study we tested different edge detec-
tors. The Prewitt (with Chessboard Distance) and the Sobel (with Euclidian Distance)
(described e.g. in [13]) edge detectors produced the best results coupled with the use
of the new implementation of algorithms [10] with the Vincent-Soille Watershed
Transformation. But neither of them was clearly better than the other. All of the fol-
lowing tests were performed with the Prewitt detector, because we obtained the best
results with it when working with our test images.
Furthermore, independently of the selected approach, Watershed Transformation
tends to highly oversegment due to many regional minima that can possibly be
interpreted as noisy minima. To solve this problem different approaches exist, like
intensive preprocessing (e.g. [14]), marker controlled watershed (e.g. [15]) or region-
merging (e.g. [16], [17]). For preprocessing in order to reduce the number of local
minima, smoothing algorithms or extended edge detectors eliminating unnecessary
edges, sharpen edges or produce edges with less gaps are adopted. Often a combina-
tion of different preprocessing methods is used. A problem of the majority of the
preprocessings methods is the dependence on the result of the particular kind of im-
ages. By the marker-controlled watershed a set of regions called markers are used in
place of the set of minima. These regions are often manually determined by the user.
Therefore this Watershed Transform approach is often qualified for interactive use
and less for automation.
To solve the problem of oversegmentation, Frucci [17] combines an iterative com-
putation of the Watershed Transform with processes called digging and flooding.
Flooding merges adjacent basins. This is achieved by letting water increase the level,
so that it can overflow from one basin into an adjacent one if the level of water is
higher than the lowest height along the watershed line separating the two basins. Also
digging merges adjacent regions. In this case, to merge a basin A, regarded as non-
significant, with a basin B, a canal is dug in the watershed line separating A and B to
allow water to flow from B into A. The effect of merging is that the number of local
minima found at each iteration diminishes. Flooding and digging are iterated until
only significant basins are left.