308 Prakt. Met. Sonderband 52 (2018) 
gamma (black) and beta (white) are identified by subjective grey-value thresholds and thus 
separated from the background (grey). The same phase can coexist in globular and 
lamellar microstructural shapes. However for the desired microstructural analysis only the 
globular shapes are of interest. But as they have the same grey value like the lamellar 
shapes they cannot be segmented by simple thresholding. Furthermore it is possible that 
there is no border between globular and lamellar shapes, so they cannot be distinguished 
by shape parameters without cutting. Microstructure constituents, that stick together are 
intended to be separated by a watershed algorithm, which does not work with the needed 
accuracy. Size and shape filters shall exclude microstructures, that do not fulfill certain 
size or shape requirements. Due to the complex microstructure this workflow is very 
imprecise, thus it needs afterwards resegmenting parts of the output manually. 
This analysis is a time consuming and repetitive task that due to human error leads to 
subjective differences between experts. Thus, this task is perfectly suited for 
improvements through automation, resulting in faster segmentation times and lower 
deviations between repeated measurements. 
Neural Networks offer the possibility to learn the underlying rules for image segmentation 
without any other input than a set of corresponding images with and without segmentation. 
2. U-Net Architecture 
it has been shown that Fully Convolutional Networks (FCN) are able to segment 
microstructures in an image with high accuracy [1]. In the following the details of the used 
software architecture are explained: 
A so called U-Net architecture [2] is implemented using the open source software library Prep 
TensorFlow [3], where only the number of classes of the output segmentation map had to 
be adjusted (Fig. 2). The architecture contains no fully connected layers and there is no The sex 
padding for a convolution (for explanation see [4]) which is why the resolution of the Ba 
feature map is smaller than the input resolution. Compare 
It is an asymmetrical architecture with roughly 33 million parameters in total in which the oreo 
encoder part has the same number of convolutional layers as the decoder part. done DV 
An advantage of the U-Net in comparison to other FCN architectures is the high resolution 
of the input image tile which was necessary to ensure that certain microstructures in the 
input image are not destroyed. Because the work of loffe and Szegedy [5] was published 
at roughly the same time as the work of Ronneberger et al. [2] batch normalization was not 
considered in their U-Net architecture. We added batch normalization layers after each 
convolution layer leading to a convergence of the loss function in about half the epochs as 
standard U-Net. 
The given dataset contains 2398 image pairs, which we parted into 1917 only for training 
and 481 only for testing. The training and testing of the U-Net was done using the 
framework TensorFlow on a NVidia Quadro K2200 GPU with 4 GB memory. The use 
We chose a modified U-Net with batch normalization after each convolutional layer trained Small gr 
for 326000 iterations because it achieved the highest mean Intersection over Union (mloU ar: 
= 0.83). The mean loU is a measure over all classes that indicates the agreement (area of 50 
intersection) between the predicted and the ground truth segmentation map in relation to Histoora 
the area of union of both maps. ’ 
Besides 
fem 
ex 
83 4 n