In-line holographic droplet imaging: Accelerated classification with convolutional neural networks and quantitative experimental validation

Abstract

Accurate measurements of cloud particle size, shape, and concentration are essential for microphysical cloud research. Holographic imaging is ideal for three-dimensional analyses of particle size, shape, and spatial distribution in large sample volumes, but its post-processing often leads to operator-dependent results and introduces uncertainties in detection efficiency. Here we present CloudTarget, a set of chrome photomasks with a customized pattern of opaque circles, serving as a verification tool to quantify detection efficiency and evaluate size and position errors. CloudTarget provides a ground truth for optimizing hologram processing parameters, including detection, sizing, and classification thresholds, and it facilitates evaluations of size- and position-dependent detection efficiency and uncertainties. Additionally, we present a Convolutional Neural Network (CNN) for object classification that achieves high accuracy with moderate training data. In a holography setup featuring a 5120 × 5120 pixel imaging sensor, a 3 μm effective pixel size, and 355 nm illumination, the CNN achieves over 90 % recall and precision for particles larger than 7 μm in a 10 cm × 1.3 cm × 1.3 cm detection volume. The average focus position error remains below 150 μm (1.5 times the reconstruction resolution) for particles <10 cm from the image plane, with in-plane random position detection errors below 5 pixels (mean <2 pixels). By combining inverse techniques with CloudTarget, the sizing error standard deviation is reduced to about 2 μm. Overall, classification performance improves significantly, and a 100-fold increase in classification speed is achieved.