Reconstructing Unsteady Flow Data From Representative Streamlines via Diffusion and Deep-Learning-Based Denoising
Reconstructing Unsteady Flow Data From Representative Streamlines via Diffusion and Deep-Learning-Based Denoising
复制标题
DOI：
10.1109/mcg.2021.3089627
复制
发表时间：
2021-11-01
影响因子：
1.8
通讯作者：
Wang, Chaoli
中科院分区：
文献类型：
--
作者：
Gu, Pengfei;Han, Jun;Wang, Chaoli
我们提出了VFR-UFD，这是一种新的深度学习框架，可以为非定常流数据（UFD）执行矢量场重建（VFR）。给定整体流线（即，流线），我们首先通过扩散生成低质量的UFD。然后，VFR-UFD利用卷积神经网络来重建时空相干的高质量UFD。VFR-UFD的核心在于递归残差块，其在局部和全局上以不同尺度迭代地细化和去噪输入向量场。我们采取连续的时间步长作为输入来捕获时间相干性，并应用基于流线的优化来保持空间相干性。为了证明VFR-UFD的有效性，我们用几个矢量场数据集进行实验，报告定量和定性的结果，并将VFR-UFD与两种VFR方法和一种压缩算法进行比较。
We propose VFR-UFD, a new deep learning framework that performs vector field reconstruction (VFR) for unsteady flow data (UFD). Given integral flow lines (i.e., streamlines), we first generate low-quality UFD via diffusion. VFR-UFD then leverages a convolutional neural network to reconstruct spatiotemporally coherent, high-quality UFD. The core of VFR-UFD lies in recurrent residual blocks that iteratively refine and denoise the input vector fields at different scales, both locally and globally. We take consecutive time steps as input to capture temporal coherence and apply streamline-based optimization to preserve spatial coherence. To show the effectiveness of VFR-UFD, we experiment with several vector field data sets to report quantitative and qualitative results and compare VFR-UFD with two VFR methods and one compression algorithm.