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 (2.5 mu m)(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 interaction steps. It is also shown how the change of the voxel size impacts the prediction of the model.