Prakt. Met. Sonderband 38 (2006) 93
4. CLASSIFICATION
ssification After the segmentation step, a decision must be made, if a segment belongs to the ferritic
6SSary, IN or martensitic phase. In order to achieve a classification, several features out of the
. segments have to be derived. The segments are regarded as masks for an image, in
roperties, which mean and variance values are calculated. The features are derived, using the LOM
hases, SO image, SEM image, combined image (chapter 2) and the gradient and Laplace images of
rwise the these images. So in total 9 images are used.
} Dbrocess. Following features of segments are determined from the images:
ted image - Average grey value of points in a segment
Average grey value of neighbouring segments
s applied Variance of grey value of points in a segment
image. Average grey value of contour points in a segment
ge based Variance of grey value of contour points of a segment
2ach local
reducing The above mentioned features are calculated in all 9 images, as consequence 5 * 9 = 45
to an over features are extracted in total.
segments The classification rules are based on groundtruth — data, determined by experts. In this
manner about 850 segments in 40 pictures of dual phase steel were classified as ferrite or
martensite. The analysis of these groundtruth — data shows approximately a Gauss -
distribution for the used features. Therefore a Bayesian classificator is used for the
classification step. The distribution of features for the two classes ferrite and martensite
can be approximated by these groundtruth — data.
contour of The Bayesian classificator is based on the statistical criteria of conditional probability [6].
The principle is presented in Figure 6. If the two classes show big differences for one
feature, this feature can be particularly used for the separation of ferrite and martensite.
Class 2
Frequency
Feature
rging Decision function
errite and Figure 6: Bayesian Classificator
the image
segments The 45 nearly Gauss — distributed features may be correlated with each other. To
minimise the correlation and extract the interesting information a Principle Components
Analysis (PCA) [6] is applied. Thus the error rate of the classification is reduced. At the
end 41 different non correlated and nearly Gauss distributed features are used for
classification.