Novel thresholding method and convolutional neural network for fiber volume content determination from 3D μCT images

In order to determine fiber volume contents (FVC) of low contrast CT images of carbon fiber reinforced polyamide 6, a novel thresholding method and a convolutional neural network are implemented with absolute deviations from experimental values of 2.7% and, respectively, 1.46% on average. The first method is a sample thickness based adjustment of the Otsu threshold, the so-called “average or above (AOA) thresholding”, and the second is a mixed convolutional neural network (CNN) that directly takes 3D scans and the experimentally determined FVC values as input. However, the methods are limited to the specific material combination, process-dependent microstructure and scan quality but could be further developed for different material types.