SEMANTIC SEGMENTATION OF μCT IMAGES OF 3D WOVEN FABRIC USING DEEP LEARNING

In this work, we have used deep convolutional neural networks (DCNN) for segmenting μCT images of a 3D fabric with orthogonal architecture with a focus on improving the segmentation of the binder yarn. A set of raw 2D slices were extracted from the gray-scale volume of the fabric. Each pixel in these slices was then annotated as voids/pores and weft/warp/binder yarns. A DCNN was implemented in MATLAB® and trained using the set of raw and annotated images. The trained network was then tested against the segmented ground truth images, and the segmentation accuracy and network performance metrics were thoroughly analyzed. The results confirm that the weight balancing improved the accuracy of binder yarn segmentation, however, at the expense of losing accuracy on the remaining classes.