Investigation of particle reconstruction quality for three-dimensional light field PIV

Abstract

This study presents a comprehensive investigation into the reconstruction quality factor, a critical metric for assessing particle position reconstruction accuracy in light field particle image velocimetry (LF-PIV). Key factors influencing the reconstruction quality are analyzed, and a benchmark criterion for reconstruction quality is proposed to ensure high-accuracy three-dimensional flow measurement. Numerical reconstructions of random particle and 3D displacement fields are performed to optimize the tomographic and deep learning reconstruction approaches. Strategies for generating optimal datasets for deep learning models are presented. The findings indicate that the generation of ghost particles and the omission of true particles are the primary causes of low reconstruction quality. The latter has a more noticeable impact, particularly when ghost particle intensities are significantly lower than true particles. A reconstruction quality factor of above 0.7 is recommended for reliable, high-accuracy flow measurements. Learning-based methods outperform tomographic algorithms in particle reconstruction, achieving comparable reconstruction accuracy with a single light field camera (LFC) to that of tomographic methods using dual LFCs. To generate high-quality datasets for deep learning, an optimal angular separation of 0.01° between sampling rays, a seeding density range of 0∼1 particle per microlens, and variable particle peak intensities are suggested. Additionally, incorporating noise at 10 % of the image intensity standard deviation into training data significantly enhances model robustness.