Volume upscaling with convolutional neural networks
Zhenglei Zhou, Yule Hou, Qirui Wang, Guangxiang Chen, Jiawei Lu, Yubo Tao, Hai Bo Lin · 2017
Volume upscaling generates high-resolution volumes from low-resolution volumes to make data exploration more effective. Traditional methods, such as the simple trilinear or cubic-spline interpolation, may blur boundaries of features and lead to jagged artifacts. Inspired by recent progress in image super-resolution with Convolutional Neural Networks (CNN), we propose a CNN-based volume upscaling method. Our CNN contains three hidden layers: block extraction and representation, non-linear mapping, and reconstruction. It directly learns an end-to-end mapping from low-resolution blocks to high-resolution volume. Compared to previous methods, our CNN can preserve better structures and details of features, and provide a better volume quality in both the visualization and evaluation metrics.