Prakt. Met. Sonderband 52 (2018) 309
Olds and fy
Sober and
x W the ne os
ame U nec
POSSile hy
A SünQuishes a
gether 5
the Need
fui Ceri
ow is ey
Tor leads b
Suited fy
Sand lower
mentation Input image tile -ansposed convolution Training:
*Qmentation » ADAM optimizer (ß, = 0.9, B, = 0.999, £ = 1e)
Convolution + (Batch utpat segmentation map constant learning rate fe
Normiization) + Rell constant weight decay tes
Max-pooling opied and cropped feature Training and Evaluation:
map » dropout probability 0.8
N er Fig. 2: U-Net architecture as presented by Ronneberger et al. [2]
3 07 the used
Weim 3, Preprocessing
' map had to
tere IS no The segmented images in the initial dataset are JPEG compressed [6] as shown in the
luton of the left image in Fig. 3. Thus at the segmented borders the color values differ slightly
compared to their intended classes. But each pixel must be unambiguously assigned to
In which the one of three classes (blue for gamma, yellow for beta, white for background). This was
done by thresholding the RGB-values.
ah resolution
cures in the
as published
tion was no
5 after each
© epochs as
N Fig. 3: JPEG compressed ground truth segmentation map on the left and normalized
j for ert ground truth segmentation map on the right.
& using the
| The used dataset contains images of low contrast. These images are characterized by a
layer rainet small grey value difference between neighboring pixels. This makes it very hard to
Union (mlod automatically distinguish dark and bright patches from grey areas. For these images we
went {area used a technique for adapting and enhancing the contrast called Contrast Limited Adaptive
n relation Histogram Equalization (CLAHE) [7].
Besides the grey value also the size and shape of the microstructural constituent
determine its class. Therefore all images have to have the same scale, where scale in this
context means meters per pixels. For the initial data set there were in total 44 different
resolutions from which 2560 x 1920 was chosen as base resolution. Images with a