Deep Hierarchical Super-Resolution for Scientific Data Reduction and Visualization
1
Scientific Achievement
Super resolution methods using neural networks have been proposed as a method to generate high-resolution scientific data given low-resolution volumes, but the methods do not consider the sparsity of scientific data and using that to reduce data sizes further. We develop a hierarchical octree-base super resolution method using neural networks for super resolution of multiscale scientific data.
Significance and Impact
We leverage adaptive hierarchical decompositions to reduce computation and memory overhead of scientific data and then use hierarchical super resolution to reconstruct data, improving accuracy and reducing computation over uniform super resolution at the same data reduction levels.
Technical Approach
2D/3D scientific data are generated for training a hierarchy of neural networks.
After training, data may be generated in an octree format, or we can take full resolution data and create an SR-octree from it.
Our hierarchical super resolution network and algorithm upscale the octree data to a uniform high-resolution for rendering and analysis.
S. W. Wurster, H. Guo, H. -W. Shen, T. Peterka and J. Xu, ”Deep Hierarchical Super Resolution for Scientific Data,” in IEEE Transactions on Visualization and Computer Graphics, 2022. Early access.