U
&
> E AS o-82:0 o> E
(a) Mannesmann-Rexroth (b) Classification 1
N „DD
x m
" A ® en } .. ® 9 ® mn
A E A E a
(c) Classification 2 (d) Classification 3 23 A
Figure 1: Comparison of manual classifications. The area of the circles is proportional to the ven das
number of images classified as represented by the centre of the circle. a
3 1
by four different experts. The comparative classifications 1, 2, and 3 were performed one .
at each of the following institutions: Fachhochschule Darmstadt, IfG Diisseldorf, and ITWM Quality
Kaiserslautern. algorithr
2 Sketch of the algorithm EE
2.1 Geometric classification. As mentioned above, the lamellae in lamellar grey cast iron re
can not be isolated properly. In particular this holds for D- and E-graphite. The interpretation distance
of the graphite as particles causes any segmentation of the images to be unstable. See Figure An
2 for an example where — depending on the light conditions when taking the picture — the images
number of particles grows from 684 to 3516. range |
In order to avoid unstable object isolation, the algorithm uses morphological operations {hen {he
(see e.g.[Soi98] for background information) applied to the complete background or the union wo
of all graphite lamellae. Successive erosions and dilations with line segments in four directions Te
(the coordinate directions and the two diagonals) as structuring elements yield a geometric
classification of each pixel. Roughly speaking, the graphite pixels are put into bins according :
to the size and arrangement of the lamellae and the “holes” in their neighborhood.
2.2 Learning. In the next step the algorithm is taught by an expert, how to assign the
bins to the graphite types A, D, and E. To this end a representative sample is classified
manually. This classification is fed to an expectation-maximisation-algorithm (EM-algorithm)
184