Adopting traditional image algorithms and deep learning to build the finite model of a 2.5D composite based on X-Ray computed tomography

Abstract

A novel methodology combining traditional image algorithms with deep learning is proposed to accurately classify each pixel of the XCT image of 2.5D woven fabrics with fewer user involvement. For images with symmetrical microstructures, we first extracted the weft and matrix edges separately and then performed curve fitting to obtain the warp edges. The regions enclosed by the warp and weft edges were weft regions, and the areas between two warps were warp regions. Then, threshold segmentation was adopted to achieve pixel classification. For an image with asymmetrical microstructures, a fully convolutional neural network consisting of one encoder and two decoder networks was trained using the symmetry image. Finally, two finite element models of the 2.5D composite were established to predict the linear elastic modulus, one containing all the geometries and the other containing only the symmetrical geometry. The results show that the former prediction fit the experimental results better.