92 Prakt. Met. Sonderband 38 (2006)
3. SEGMENTATION
Using only the combined feature image of the LOM and SEM images, a classification Aft
process can not be directly achieved. An intermediate segmentation step is necessary, in or
order to extract more information about the phases. sec
Aim of the segmentation is the combination of neighbouring pixels, with similar properties, wh
into segments. It is desired, that the derived segments correspond to different phases, so im:
pixels of different phases, have to be merged into different segments. Otherwise the the
phases can not be separated from each other during the following classification process. Fo
Various features of the calculated segments can be extracted from the segmented image
and this information mainly supports the subsequent classification process.
To reduce calculation time a pre-filtration with the Mean Shift Algorithm [4] is applied
before segmentation, where small noise is smoothed and sharp edges rest in the image.
For the segmentation process, the Water Shed Algorithm [5] is used. It is an edge based
segmentation method which uses the gradient image of combined image. For each local
minimum within the gradient image, a segment is generated. Although a noise reducing Th
prefiltering was applied to the combined image, the Water Shed Algorithm leads to an over fea
segmentation. That's why a subsequent merging step, where similar adjacent segments Th
are combined, is necessary. ma
The criteria for a subsequent merging process are: dis
- Difference in grey value of neighbouring segments cla
Shared contour length/contour length cal
Average value in the gradient image of the combined picture on the common contour of Th
neighbouring segments Th
fe:
Figure 5: Result of segmentation after pre-filtration, segmentation and merging
In Figure 5 is the result of segmentation shown. The representation of ferrite and
martensite in single segments was successful. The pixels of different phases in the image
of the multiphase steel are merged in different segments. Various features from segments Th
are extracted and can be used for classification. mil
Ar
en
Cl: