Parametric study and lightweight design of implicit neural representation for tomographic background-oriented Schlieren
Tong Jia, Jiawei Li, Xuhui Meng, Dae Geun Park, Yuan Xiong
Beihang University Huazhong University of Science and Technology Korea Institute of Industrial Technology
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Multi-view tomographic background-oriented Schlieren (TBOS) has become an important optical technique for measuring 3D density-varying flow fields. Reconstructing the 3D density field from TBOS measurements is a typical inverse problem, which becomes severely ill-posed with a limited number of views. Recent studies have shown that implicit neural representation (INR) exhibits significant advantages over voxel-based reconstruction algorithms for TBOS under such conditions. However, systematic parametric studies of INR remain scarce and the conservative sampling strategies employed during INR training produce redundant sampling points and extra network overhead, reducing both efficiency and accuracy. Based on the neural deflection field (NeDF) developed in our previous work, a sequential parametric study of the network is carried out using synthetic BOS images, revealing that the network depth dominates the reconstruction accuracy and that the positional encoding lowers the network scale required to reach the attainable accuracy rather than raising its upper limit. On this basis, to improve network training efficiency and reconstruction quality, a lightweight design with an adaptive sampling strategy is developed, which uses a single network to progressively concentrate the sampling points toward the key flow regions as training proceeds. Finally, a turbulent helium jet experiment is conducted, in which the proposed lightweight NeDF achieves the best cross-view generalization among all the methods compared. This study provides an effective solution for coping with this ill-posed reconstruction challenge and achieving high-accuracy, high-efficiency TBOS flow diagnostics.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
物理Model Reduction and Neural Networks
Optical measurement and interference techniques · Fluid Dynamics and Turbulent Flows