Deep learning segmentation of fibre, matrix, and void microstructure in carbon fibre and flax fibre polymer composites : phase segmentation in carbon fibre and flax fibre polymer composites using lightweight deep neural networks

Abstract

This thesis develops and evaluates a fully automated deep learning workflow for segmenting voids and material phases in X-ray computed tomography of carbon fibre and flax fibre polymer composites. For the carbon fibre system, fibre, matrix, and void masks were generated through a reproducible Python-based pipeline combining slice-wise intensity normalisation, two-stage void detection, morphological refinement, and Otsu-based fibre–matrix separation, followed by targeted manual verification in Dragonfly. For the flax fibre system, segmentation masks were produced and refined manually and unified into a consistent three-phase labelling scheme (matrix, fibre, void). These reference datasets were then used to train five lightweight convolutional neural network architectures (UNet++, UNet3+, Attention UNet, DeepLabV3+ Transformer, and LR-ASPP Transformer) under identical conditions, using paired 256 × 256 patches, controlled label-safe augmentation, and a fixed slice-level train–validation split. A higher-load carbon fibre state (140 N) was withheld entirely from training and used exclusively to evaluate the models on previously unseen microstructural damage. The results demonstrate that compact encoder–decoder networks can accurately localise voids and robustly separate fibre and matrix phases across both composite systems, including under low contrast and evolving damage, while maintaining computational efficiency suitable for routine XCT workflows.