LAUSR

Abstract Scanning electron microscope (SEM) is proven effective to analyze the morphology of carbon fibers (CFs) presenting in CFRC. However, the qualitative observation has limited contribution to the improvement of CF distribution as well as the properties of CFRC. In this work, a fully convolutional network (FCN) was developed to segment CFs from SEM images for quantitative CF distribution characterization. Three processes involved in the establishment of the FCN and its application for the CF distribution evaluation, which were: (a) generating a database including 560 CFRC SEM images in different scales; (b) designing, training, and testing an encoder-decoder network and other layers for the FCN; and (c) evaluating the CF distribution and analyzing the relationship between the CF distribution and the CFRC properties using segmentation results. The results showed that the FCN provided reasonable segmentation results for CF clusters with the 0.94F-Measure, 0.92 recall, and 0.96 precision, respectively. The FCN had stable segmentation results under different SEM magnifications. The FCN-based method was proven effective to segment CF clusters in real time, which met the demand for continuous SEM observation. The continuous observation results indicated that the mechanical and electric properties of CFRC were improved by the improvement of the CF distribution.