Voxel-based Deep Learning for Image Super-resolution of Areal Density Maps of Carbon-nanotube Sheets

Miralon sheets are an advanced carbon-based product that provides outstanding physical properties which enable strong, lightweight, and environmentally resistant solutions for aerospace and industrial applications. For quality control, beta transmission equipment is used to generate areal density maps to ensure the uniformity of the Miralon sheets. However, the original areal density maps obtained from the beta transmission equipment is insufficient to identify variation and defects at the required level. As areal density maps can be interpreted as images, current work in image super-resolution deep convolutional neural network (CNN) promises to reconstruct high-resolution areal density maps with fine texture details. However, for this promise to be realized two challenges need to be overcome. First, generating areal density maps is expensive, and only limited training data is available. Second, current 2D super-resolution neural networks cannot provide sufficient accuracy in the vertical axis (thickness) direction. Therefore, we propose a 3-layer voxel representation to adapt current deep CNNs to the particular application of Miralon sheets. In particular, our results demonstrate that appropriate data augmentation is essential to achieving accurate results. Using our proposed techniques, a high-resolution areal density map, which is rich in 3D information, can be obtained by leveraging 2D image super-resolution CNNs.