Application of deep learning and inline holography to estimate the droplet size distribution

TL;DR

This study evaluates deep learning architectures for droplet size estimation from inline holography images, finding U-Net and R2 U-Net most effective, and demonstrates their application in analyzing experimental data with improved distribution modeling.

Contribution

It compares five machine learning architectures for droplet segmentation, identifies the most proficient ones, and validates their use in real-world holographic data analysis.

Findings

U-Net and R2 U-Net outperform other architectures in segmentation accuracy.

Gamma distribution models droplet size distribution more accurately than log-normal.

Deep learning methods effectively estimate droplet sizes in experimental holography images.

Abstract

We examine five machine learning-based architectures to estimate the droplet size distributions obtained using digital inline holography. The architectures, namely, U-Net, R2 U-Net, Attention U-Net, V-Net, and Residual U-Net are trained using synthetic holographic images. Our assessment focuses on evaluating the training, validation, and prediction performance of these architectures. We found that U-Net and R2 U-Net to be the most proficient, displaying consistent performance trends and achieving the highest Intersection Over Union (IOU) scores compared to the other three architectures. We employ additional training using experimental holographic images for the two top-performing architectures to validate their efficacy further. Subsequently, they are employed to segment an experimental dataset illustrating the bag breakup phenomenon, facilitating the extraction of size distribution. The extracted size distribution from U-Net and R2 U-Net segmentation is then compared with the analytical model proposed by \cite{jackiw2022prediction} by employing the gamma and log-normal distributions. Our findings indicate that the gamma distribution provides a more accurate prediction of the multi-modal size distribution than the log-normal distribution owing to its long exponential tail. The present study offers valuable insights into the effectiveness of machine learning architectures in estimating particle/droplet sizes, highlighting their practical application in real-world experimental scenarios.