[Paper Review] Particle reconstruction of volumetric particle image velocimetry with strategy of machine learning

Qi Gao, Shaowu Pan+3 more|arXiv (Cornell University)|Sep 16, 2019

Electrical and Bioimpedance Tomography30 references9 citations

TL;DR

This paper proposes a deep learning-based particle reconstruction method for volumetric PIV that uses a 3D convolutional neural network (CNN) to refine coarse initial guesses from traditional algebraic reconstruction techniques (ART), significantly improving reconstruction quality, noise robustness, and speed—achieving at least a 10x speedup over iterative methods like SF-MART while reducing ghost particles and better preserving particle shape.

ABSTRACT

Three-dimensional particle reconstruction with limited two-dimensional projections is an under-determined inverse problem that the exact solution is often difficult to be obtained. In general, approximate solutions can be obtained by iterative optimization methods. In the current work, a practical particle reconstruction method based on a convolutional neural network (CNN) with geometry-informed features is proposed. The proposed technique can refine the particle reconstruction from a very coarse initial guess of particle distribution generated by any traditional algebraic reconstruction technique (ART) based methods. Compared with available ART-based algorithms, the novel technique makes significant improvements in terms of reconstruction quality, {robustness to noises}, and at least an order of magnitude faster in the offline stage.

Motivation & Objective

• To address the long-standing challenge of robust and efficient 3D particle reconstruction in volumetric PIV, particularly under high particle seeding densities.
• To overcome limitations of traditional ART-based methods such as high computational cost, noise sensitivity, and ghost particle artifacts.
• To develop a machine learning-based approach that enhances reconstruction quality and speed using geometry-informed features and a 3D CNN.
• To validate the method on synthetic data and demonstrate its superiority over existing ART-based algorithms in reconstruction accuracy and robustness.
• To lay the foundation for future application to real experimental data by integrating calibration and domain generalization in training.

Proposed method

• The method uses a multiplicative LOS (MLOS) estimation to generate a coarse initial guess of particle locations and intensities from 2D projections.
• A 3D convolutional neural network (CNN) is trained to refine this initial guess by learning geometry-informed features from synthetic particle fields.
• The CNN is trained end-to-end on synthetic data with varying particle seeding densities (0.05–0.3 ppp) and noise levels (0–0.3).
• The network architecture is designed to restore spherical particle shapes and suppress ghost particles, improving intensity distribution and spatial fidelity.
• The method is evaluated using a quality factor Q defined as the correlation coefficient between synthetic and reconstructed particle fields.
• Training is performed on GPU for 16 hours over 100 epochs; inference is highly parallelizable, enabling real-time performance.

Experimental results

Research questions

• RQ1 Can a deep learning model significantly improve the accuracy and robustness of 3D particle reconstruction in volumetric PIV compared to traditional ART-based methods?
• RQ2 How does the proposed AI-PR method perform in terms of reconstruction quality across varying particle seeding densities and noise levels?
• RQ3 To what extent does the CNN-based refinement reduce computational time compared to iterative ART methods like SF-MART?
• RQ4 Can the network effectively suppress ghost particles and restore true particle morphology, such as spherical shape, from coarse initial guesses?
• RQ5 How does the method generalize across different particle densities and noise conditions in synthetic data?

Key findings

• AI-PR achieves a quality factor Q of approximately 0.7 at a particle seeding density of 0.25 ppp, while SF-MART-10 drops below 0.6 under the same conditions.
• At a fixed density of 0.15 ppp, AI-PR maintains superior stability against noise, with Q decreasing more slowly than SF-MART across increasing noise levels (0–0.3).
• The total wall time for AI-PR is 524.5 seconds, representing a 10x speedup over SF-MART-10 (9881.5 seconds), despite including MLOS and CNN computation.
• The actual inference time for the CNN component is only about 12 seconds, highlighting the efficiency of GPU-parallelized inference versus iterative voxel-dependent updates in SF-MART.
• AI-PR produces more spherical particle shapes and fewer ghost particles than SF-MART, as visually confirmed in cross-sectional reconstructions.
• The method demonstrates high robustness to noise and maintains better intensity distribution, indicating improved fidelity in reconstructed 3D particle fields.