Deep neural network-based characterization and positioning of particles in three-dimensional particle fields in digital holography with phase imaging
Accurate, simultaneous determination of particle positions and physical characteristics is a fundamental requirement in digital holography for three-dimensional (3D) particle-field measurements. Recent approaches based on deep neural networks (DNNs) using two-dimensional (2D) U-Net architectures have demonstrated the estimation of lateral and axial positions and radii of particles directly from hologram patterns. However, these frameworks have limited axial-position accuracy and do not consider the estimation of refractive indices because they rely on models based on scalar diffraction theory. In this study, we propose a DNN-based method that simultaneously determines the lateral position, axial position, radius, and refractive index of a spherical particle in a 3D particle field recorded by an in-line hologram. The key idea is to use a multislice 2D phase stack along the optical axis, instead of hologram patterns, as training data. The phase stack is obtained using the transport-of-intensity equation from holograms computed via electromagnetic Mie scattering theory, yielding stable, accurate phase values at arbitrary axial positions. To effectively exploit the axial information contained in the phase stack while preserving the full lateral resolution, we use a hybrid architecture consisting of 3D convolution and 2D U-Net. Numerical results demonstrate that the proposed approach considerably improves the estimation accuracy of all particle parameters compared with a conventional hologram-based method.