Prakt. Met. Sonderband 52 (2018) 307
“Onna 4 . . . .
5) gp Fully Automated Quantitative Microstructural Image Analysis
:jyy by a Fully Convolutional Neural Network
Siegen di
hy Werden Frederik Elischberger***, Dr. Hendrik Kramer*, Julian von Lautz*, Kerstin Spindler*
* MTU Aero Engines AG, Munich, Germany
* Department of Information Systems, Westfalische Wilhelms-Universitéat, Munster,
Germany
19 der Orign,
| he Zeige Abstract
Orphologie
We use Deep Learning techniques utilizing Fully Convolutional Neural Networks to
ung unter. segment metallographic images of the intermetallic material titanium aluminide in a
Stmmien be supervised fashion. We implement different semantic segmentation network architectures,
Restauster. based on a modified U-Net with Batch Normalization and train the Neural Nets on 1917
'gsanderung already labeled images. To maximize efficiency we show useful pre-processing techniques
Sen der the. in the context of quantitative analysis of images generated from light optical microscopy.
‚Berechnung
1. Introduction
Material properties of metals, ceramics and intermetallics are predominantly determined by
rusty their microstructure and therefore the quantitative analysis is commonly used as a quality
ge measurement tool.
fistorical As
1,2015
angth Steels
] watershed
te Par manual algorithm,
0, 2007 and size, shape,
ns, V. Maier and manua
Isothemaly threshold segmentation
ovets, “Fun
the Volume
lic Measure:
2
of Complex
"Mater. 50 fully automatic: FCN
0 Fig. 1: The semiautomatic workflow uses grey value thresholds as well as size and shape
2 fro parameters to cut away all microstructure constituents except the desired ones.
! pls The FCN shall generate the same output as the semiautomatic workflow.
nof EBD The analysis discussed in this paper is currently performed in a semiautomatic fashion by
| trained metallography experts. Conventional image manipulation techniques are used on
Lf. Pgs, images generated from light optical microscopy (compare Fig. 1): First, the two phases
