Segmentation of multiple features in Carbon Fiber Reinforced Polymers using a Convolutional Neural Network

This paper shows how a convolutional neuronal network can be used to segment multiple features (such as
matrix, fiber bundles and defects) in a single step from X-Ray computed tomography data acquired from carbon
fiber reinforced polymer (CFRP) specimens. The sample analyzed was 5 plies thick plain weave CFRP widely
used in automotive and aerospace application. The specimen was scanned using a GE phoenix X-ray Nanotom
XCT with an voltage of 60kV and a voxel size of (2.5µm)2
. To allow for the prediction of multiple classes, the
standard U-Net architecture was extended to use a softmax (one-hot encoding) as output layer. The trained
network delivers similar results as compared to current state-of-the art methods, with the additional advantage
of reducing the number of required human