Prakt. Met. Sonderband 52 (2018) 311 
aff SCal | Lo = Ai . . . © CY = 
8 ae Wh Slicing the images is one obvious strategy to maintain the possibility for the U-Net to make 
“M@ precise predictions based on size and shape. In = 
dion fy As in [8], we also tried scaling the whole input image to 568x568 (now we use padding so 
is Shot that the output segmentation map has the same size as the input image) and feed it 
Shag directly into the Small U-Net resulting in only 5 million parameters. Then we trained the 
fog “small U-Net with scaled images” for 65000 iterations. Afterwards, we upsampled the 
Bey output of the network to the resolution of the input image. This results in a mloU = 0.78. 
Melly The resulting segmentation map of this architecture is coarser than with the slicing 
Mage oe method. On the other hand, the training time decreases to only one day. 
08 Wil not 
Ohboring tes 
SBF : 
be image 5. Results 
Fig. 5 gives an overview of the model predictions in comparison to the ground truth 
images. The dashed lines visualize the acceptable deviation between a human 
metallography expert and the ground truth. For both the small U-Nets with sliced and 
scaled images, 96% of the test images are within the acceptable range. We call this value 
“deviation”. 
| 
N eg. a dar Le Ground truth phase fraction. = Ground truth phase fraction ~ 
an artficaly | | . 
el Fig. 5: Small U-Net with sliced input images (left) and scaled input images (right) 
order pattem a 
1y unlely The outliers visible in both plots are due to inconsistencies of the ground truth data set. 
q fight pies Like a child that was educated not to cross a red traffic light, U-Net shows us, where we 
did. In other words due to the human influence our ground truth data set contains images 
with errors in segmentation. oo | 
For a “correctly” segmented ground truth image the example in Fig. 6 shows how precise 
the Small U-Net with sliced input images can predict the segmentation. 
Table 1 gives an overview of the discussed models. 
: system Init imaae “round Truth “radiction 
rameters 0 
re was then 
firming the 
map. Due t 
15. The use 
Jr hardwaz ur : 
y bs Fig. 6: Precision of prediction by Small U-Net sliced in comparison to ground truth