基于波前整形与深度学习的高密度单目三维粒子图像测速技术
High-Density Monocular 3D Particle Image Velocimetry by Wavefront Shaping and Deep Learning
中文总结 AI 辅助
该研究提出 PET-PIV 方法，通过单薄相位掩模和深度学习，解决单目三维 PIV 紧凑性与密度的矛盾，实验显示其与真值相关性超0.97，速度提升一个数量级，拓展了三维测速的应用场景。
中文摘要 AI 辅助
三维粒子图像测速（3D PIV）通过成像流体中激光照明的示踪粒子来测量流速，广泛应用于 academia 和 industry。许多应用需要紧凑的光学可访问装置，理想情况下仅用单台相机，同时要求高 seeded 密度以实现准确测速，但这两个要求通常相互矛盾：单目方法在高密度下失效，高密度测量则通常依赖多相机断层摄影系统。本文提出基于点扩展函数工程训练的粒子图像测速技术（PET-PIV），这是一种紧凑的单目三维测速方法，通过 minimal 光学修改和基于深度学习的算法解决了长期存在的紧凑性-密度权衡问题。PET-PIV 仅需在常规 PIV 装置中插入一个薄相位掩模，显微镜中置于物镜下方，宏观成像中置于镜头光圈处，并对成像系统进行原位校准，易于实现且可扩展。计算上，PET-PIV 运行在两种互补模式：跟踪/拉格朗日模式，用于定位和连接单个粒子；更重要的是，适用于超高密度的场模式，可直接从二维图像序列重建三维速度场。在使用断层摄影 PIV 系统的实际实验验证中，PET-PIV 与 ground-truth 参考结果高度吻合，相关系数（CC）超过 0.97，计算速度提升一个数量级。值得注意的是，PET-PIV 实现了仅用单目三维宏观 PIV，达到了此前仅多相机装置才能实现的密度，显著拓展了三维测速在空间受限和光学受限环境中的适用性。
英文摘要
Three-dimensional (3D) Particle Image Velocimetry (PIV) measures flow velocities by imaging laser-illuminated tracer particles seeded in a fluid, and is widely used in both academia and industry. Many applications require a compact setup for optical accessibility, ideally with a single camera, while also demanding high seeding densities for accurate velocimetry. These requirements, however, are typically incompatible: monocular methods break down at high densities; high-density measurements instead generally rely on multi-camera tomographic systems. Here, we introduce Point-spread-function-Engineering Training-based PIV (PET-PIV), a compact monocular 3D velocimetry approach that resolves this long-standing compactness-density trade-off through a minimal optical modification and deep-learning-based algorithms. PET-PIV requires only the insertion of a single thin phase mask to an otherwise conventional PIV setup, beneath the objective in microscopy or at the lens iris in macro-scale imaging, and calibrates the resulting imaging system in situ, making the approach straightforward to implement and readily scalable. Computationally, PET-PIV operates in two complementary regimes: a tracking/Lagrangian mode that localizes and links individual particles, and, more importantly, a field mode for ultra-high densities which directly reconstructs 3D velocity fields from 2D image sequences. In realistic experimental validations using a tomographic PIV system, PET-PIV demonstrates strong agreement with ground-truth references, with correlation coefficients (CC) exceeding 0.97, alongside an order-of-magnitude improvement in computational speed. Notably, PET-PIV enables monocular 3D macro-scale PIV at densities previously achievable only with multi-camera setups, dramatically expanding the applicability of 3D velocimetry in space-constrained and optically restricted environments.
