Automatic yarn path extraction of large 3D interlock woven fabrics with confidence estimation

Modeling realistic textile composite structures remains a challenging task due to their complex geometry. In this paper, a novel method for reconstructing yarn paths based on micro-computed tomography is proposed. A deep learning approach is employed to convert mu- CT scan into an appropriate distance map, which is used for extracting yarn paths with a tracking algorithm. An ablation study is performed to understand the hyperparameters that matter the most. This study includes variation of the target images, selection of spatial dimension of the U-Net (2D, 2.5D and 3D), dataset sampling strategies and loss terms weighting. Additionally, a robust method for estimating the quality of the predictions without the need for annotation is introduced. The accuracy of the reconstruction method is demonstrated through the analysis 15 test mu- CT images, with 5 devoted to the optimal post-processing evaluation and 10 for assessing the final test results.