Geometrical and deep learning approaches for instance segmentation of CFRP fiber bundles in textile composites

Abstract

Segmenting micro-Computed Tomography (mu CT) images of textile composites is a necessary step before modeling the material at the mesoscale. However, the accurate segmentation of fiber bundles (or tows) remains a challenge in carbon fiber reinforced textile composites. Segmentation approaches based on local fiber orientation perform well in recognizing individual tows only under ideal conditions, namely when the local fiber orientation bordering two tows' interface is different, or when the touching area is small relative to the thickness of a tow. Unfortunately, in many textile composite laminates used in the industry, these ideal conditions are not found. Such materials often consist of multiple plies, where each fiber is aligned in one of the two orthogonal directions, and where the touching area between similar-orientation tows is often much larger than the tow thickness. Therefore, we propose two new methodologies for splitting tow instances. One is based on the geometrical analysis of the material structure using conventional image analysis; the other is based on the deep learning prediction of ideal inputs for segmentation based on the watershed transform. The deep learning-based method is trained using randomly generated synthetic images of a woven composite material, which avoids an expensive human annotation step.