3-D Particle Reconstruction for Flow Motion Estimation via Multiscale Pyramid Network

Tomographic particle image velocimetry (Tomo-PIV) is a typical volumetric measurement technique for complex flows. To compute the instantaneous velocity field, the voxel intensity distributions are reconstructed by iterative tomographic reconstruction algorithms (e.g., multiplicative algebraic reconstruction technique, MART), followed by the application of 3D cross-correlation on a regular Eulerian grid. Classical reconstruction algorithms tend to generate ghost particles and elongated particles at high seeding densities and are computationally time-consuming. This paper proposes a neural network-enhanced multiplicative line of sight (MLOS) method to improve the accuracy and efficiency of particle distribution reconstruction. We use the MLOS algorithm to quickly obtain an initial intensity distribution and then refine particle positions and shapes using a 3D multi-scale pyramid network. To enhance useful information in the particle field and filter out irrelevant features, we incorporate an attention mechanism before each downsampling step and employ a feature pyramid with irregular convolutions to capture multi-scale information. We introduce a quality factor in the loss function to further improve the particle reconstruction quality. Experimental evaluations on synthetic data show that the proposed framework effectively eliminates most ghost particles and recovers elongated particles compared to existing MART-based reconstruction algorithms. In addition, the reconstruction speed is at least an order of magnitude faster at higher particle densities.