Efficient processing of μCT images using deep learning tools for generating digital material twins of woven fabrics

The greatest challenge in creating digital material twins from mu CT images is the lack of a robust and versatile tool for segmenting the mu CT images and post-processing the segmented volumes into a FE mesh. Here, we have used deep convolutional neural networks (DCNN) for segmenting mu CT images of a multi-layer plain-woven fabric. First, a set of raw 2D image slices extracted from the gray-scale volume of a single-layer fabric was used to train a DCNN using manually annotated images. The trained DCNN was then tested using some unseen manually segmented images, resulting in more than 96% global accuracy. Moreover, the trained DCNN was also used to segment unseen images from a multilayer stack of the fabric with good accuracy. A novel procedure based on the watershed segmentation technique was also successfully developed to separate individual yarns from connected yarn cross-sections during post-processing of segmented volumes. The work presented here provides a robust and efficient framework of segmenting CT scan images of woven fabrics for generating their digital material twins and FE mesh.

Automated summary

In this study, deep convolutional neural networks were utilized to segment mu CT images of a multi-layer plain-woven fabric with over 96% global accuracy. Additionally, a novel procedure based on the watershed segmentation technique was successfully developed for post-processing of segmented volumes.