计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (6): 1321-1328.DOI: 10.3778/j.issn.1673-9418.2112069

• 理论·算法 • 上一篇    下一篇

MSV-Net:面向科学模拟面体混合数据的超分重建方法

曹斯铭,王晓华,王弘堃,曹轶   

  1. 1. 中物院高性能数值模拟软件中心,北京 100088
    2. 北京应用物理与计算数学研究所,北京 100088
  • 出版日期:2023-06-01 发布日期:2023-06-01

MSV-Net: Visual Super-Resolution Reconstruction for Scientific Simulated Data of Mixed Surface-Volume

CAO Siming, WANG Xiaohua, WANG Hongkun, CAO Yi   

  1. 1. CAEP Software Center for High Performance Numerical Simulation, Beijing 100088, China
    2. Institute of Applied Physics and Computational Mathematics, Beijing 100088, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 高保真度的可视分析通常依赖大规模科学模拟产生的耦合几何模型的高分辨率网格数据,这对数据存储和流畅交互均提出了巨大挑战。提出了MSV-Net,一个面向大规模科学模拟面体混合数据的超分辨率重建方法。该网络为端到端的深度神经网络,通过多层非线性变换实现从低分辨数据到高分辨数据的混合绘制映射的联合学习;该网络舍弃了全连接层,不仅可以减少网络参数,而且能够提升网络的灵活性与可复用性。此外,构建了面向大规模电磁模拟应用的面体混合数据集MSV-Dataset,用于模型训练和验证。该数据集由采用不透明几何模型绘制耦合半透明体绘制的混合绘制的图像构成。与多种传统方法和深度学习方法进行了对比,定量分析结果显示,MOS绝对评价指标达到了4.1,重建准确率仅次于真实图像;基于混合数据绘制1 500×1 500分辨率的图像,采用直接绘制需要66.28 s,而采用MSV-Net则仅需要4.14 s,交互性能提升了约15倍。

关键词: 科学数据可视化, 混合绘制, 大规模模拟, 超分辨率重建, 深度学习

Abstract: High fidelity visual analysis usually relies on high-resolution grid data of coupled geometric models generated by large-scale scientific simulation, which brings great challenges to data storage and smooth interaction. Therefore, this paper proposes a super-resolution reconstruction method for large-scale scientific simulated data of mixed surface-volume, which is MSV-Net. The network is an end-to-end deep neural network, which realizes the joint learning of hybrid rendering mapping from low resolution data to high resolution data through multi-layer nonlinear transformation. The network without the fully connected layer, can not only reduce the network parameters, but also improve the flexibility and reusability of the network. In addition, MSV-Dataset, a surface-volume mixed dataset for large-scale electromagnetic simulation application is constructed for model training and verification. This dataset consists of mixed rendered images using nontransparent geometric model rendering coupled with semitransparent volume rendering. The proposed method is compared with a variety of traditional and deep learning methods. The quantitative analysis results show that the MOS absolute evaluation index of this method reaches 4.1, and the reconstruction accuracy is second only to the real image; it takes 66.28 s to directly draw mixed data with 1500×1500 image resolution, but only 4.14 s with the proposed method. The interaction performance is improved about 15 times.

Key words: scientific data visualization, hybrid rendering, large-scale simulation, super-resolution reconstruction; , deep learning