Abstract
We present an approach for hierarchical super resolution (SR) using neural networks on an octree data representation. We train a hierarchy of neural networks, each capable of 2x upscaling in each spatial dimension between two levels of detail, and use these networks in tandem to facilitate large scale factor super resolution, scaling with the number of trained networks. We utilize these networks in a hierarchical super resolution algorithm that upscales multiresolution data to a uniform high resolution without introducing seam artifacts on octree node boundaries. We evaluate application of this algorithm in a data reduction framework by dynamically downscaling input data to an octree-based data structure to represent the multiresolution data before compressing for additional storage reduction. We demonstrate that our approach avoids seam artifacts common to multiresolution data formats, and show how neural network super resolution assisted data reduction can preserve global features better than compressors alone at the same compression ratios.
Abstract (translated)
URL
https://arxiv.org/abs/2107.00462
