Semantic Segmentation of Carbon-Fibre Defects from Multi-Channel Photometric Imaging

Carbon-fibre reinforced materials are increasingly adopted in high-performance industrial applications, where surface defect inspection plays a critical role in ensuring product quality and structural integrity. This thesis addresses the problem of automated defect inspection in carbon-fibre textiles using multi-channel photometric imaging data acquired from the PROFACTOR Fscan industrial sensing system, which produces sensor-derived representations encoding physical properties of the material surface rather than conventional RGB imagery. A central methodological contribution of the thesis is the analysis of data leakage induced by grouped acquisition protocols, showing how standard random splitting strategies can substantially inflate performance estimates. To address this issue, leakage-safe evaluation protocols are introduced in order to obtain more reliable measures of model generalization. The work further investigates sensor-channel combinations, normalization strategies, architecture selection, backbone capacity, imbalance-aware loss formulations, and augmentation techniques, including an adaptation of CutMix for industrial defect segmentation. Experiments based on SegFormer architectures are conducted across different evaluation regimes, providing a structured analysis of the principal factors affecting segmentation performance under realistic industrial constraints.