En
期刊文章
文章精选
当期目录
过刊浏览
高被引
高下载
特邀文章
专刊合集
待刊文章
Volume 113
您当前的位置：首页 > 期刊文章 > 过刊浏览 > Volumes 108-119 (2025) > Volume 113
3D particle localization using self-supervised learning method for Lagrangian particle tracking
Haoqin Huang a b, Tianyi Cai a b, Dapeng Zhang a b, Wu Zhou a b *
a School of Energy and Power Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China
b Shanghai Key Laboratory of Multiphase Flow and Heat Transfer for Power Engineering, Shanghai, 200093, China
10.1016/j.partic.2026.03.018
Volume 113, June 2026, Pages 218-235
Received 25 December 2025, Revised 26 February 2026, Accepted 6 March 2026, Available online 26 March 2026, Version of Record 8 April 2026.
E-mail: zhouwu@usst.edu.cn
Highlights

• A self-supervised learning method for 3D particle localization using MLP and iterative computational scheme.

• Relationship between particle defocus features and depth information was incorporated as a physical constraint.

• The proposed approach outperforms state-of-the-art IPR method in presence of a large number of defocus effect.


Abstract

To achieve high spatial resolution in 3D Lagrangian Particle Tracking (LPT), a self-supervised learning-based 3D particle localization method is proposed, which employs a Multilayer Perceptron (MLP) framework within an iterative optimization scheme. This neural network takes the estimated depth coordinate of each particle as input and is trained in a self-supervised manner by projecting particles onto multiple camera views, using the distance to the nearest matched projection point as the loss function. Depending on the input strategy, two implementations are introduced: the Sequential and Ensemble Particle Localization Technique (SpLTS and EpLTS, respectively), both based on self-supervised learning. Evaluated on synthetic particle fields, the method demonstrates significant improvements over the state-of-the-art Iterative Particle Reconstruction (IPR) method under defocus condition, since the proposed method can leverage particle defocus feature as a physical constraint to refine depth estimates. Particularly when the depth of field is less than half the reconstruction volume (fδ < 0.5), both SpLTS and EpLTS achieve approximately 10–15% higher localization accuracy than IPR, while maintaining robust performance up to high seeding densities of 0.14 ppp. Furthermore, under backlight illumination, neither the defocus effect within the intersection zone nor the particles outside it substantially affects SpLTS, which demonstrates superior performance compared to EpLTS. However, EpLTS exhibits higher computational efficiency than SpLTS, although both remain slower than IPR, indicating substantial potential for further optimization.

Graphical abstract
Keywords
Lagrangian particle tracking; 3D localization; Self-supervised learning; Multilayer perceptron
文章导读