312 Prakt. Met. Sonderband 52 (2018)
Table 1: Model comparison gildana
noe!
Large U-Net Small U-Net Small U-Net (scaled) 0
Iterations 326000(~1 week) 300000(~5 days) 65000(~1 day) =
#Parameters > 33 million > 15 million > 5 million
mioU 0.83399 0.81243 0.78289
Deviation 94.4% 96.9% 96.7%
Evaluationtime ~~ 23s/img ~ 9sfimg ~ 0.4s/img
. . (urzfass
6. Summary and Discussion
Deep Learning is used to segment metallographic images fully automatically in very good pe
accordance to ground truth images segmented in a semiautomatic workflow. To achieve a Sr m
uniform image input we investigated suitable pre-processing techniques. Categorical Tan
encoding of the jpeg compressed image by thresholding and contrasting by Contrast we
Limited Adaptive Histogram Equalization shows good performance. We demonstrate a
tiling strategy to partition input images larger than the U-Net input pixel-size, that contains
neighbor and border information. We evaluated three Fully Convolutional Network (FCN) jEinlen
architectures based on U-Net. The “Large U-Net” results in the highest mloU but has not
the highest conformance to the acceptable range. For the used data set it turned out, that vee
the trainable depth of the FCN can be reduced which leads to an increase in speed without Avon
losing quality. For an even faster FCN, the used overlap-tile strategy could also be kn einher
replaced by scaling down without significant change in the determined phase fractions. pine
The solved image analysis task is relative complex. Thus we are confident, that the “U- x entsta
Net” can also be used for other completely different image analysis tasks in the future. rr
set und
References iE
of
[1] Azimi, S.M. et al.: “Advanced Steel Microstructure Classification by Deep Learning de Beurte
Methods”, arXiv preprint arXiv:1706.06480, 2017 Bee
[2] Ronneberger, O. et al.: “U-Net: Convolutional Networks for Biomedical Image Eine einh
Segmentation”, International Conference on Medical Image Computing and mung. In‘
Computer-Assisted Intervention, p. 234-241, Springer, 2015 de Hon
[3] Abadi, M. et al.: “TensorFlow — an open source machine learning framework for schaun
everyone”, https://www.tensorflow.org/, 2018 [Online: accessed 07-May-2018] ce
[4] Dumoulin, V., Visin, F.: “A guide to convolution arithmetic for deep learning”, arXiv Krooe
preprint arXiv:1603.07285, 2018 de Form
[5] loffe, S., Szegedy, C.: “Batch Normalization: Accelerating Deep Network Training by Simmung
Reducing Internal Covariate Shift”. International Conference on Machine Learning, p. demonstr
448-456, 2015 fendung
[6] Wallace, G.K.: “The JPEG still picture compression standard”, Communications of the Og veg;
ACM, Volume 34 Issue 4, p. 30-44, 1991 Bory;
[7] Zuiderveld, K.: “Contrast limited adaptive histogram equalization”, Graphics Gems, p. Die Obje!
474-485, 1994 anderer
[8] Chen, L-C. et al.: “Encoder-Decoder with Atrous Separable Convolution for Semantic Bereichs
Image Segmentation”, arXiv preprint arXiv: 1802.02611, 2018 ntarang