Height Prediction in Particle Images Using Deep Learning

Particle images are used to study flow fields by observing changes in tracer particle location over time. Typically, one camera is used to obtain 2D information. However, particles can be at different height. Knowing the particle height completes the particle movement vector in region observed. Without using a second camera, the current work provides a novel approach to extract height by approximating inverse Lommel function using a convolution neural network (CNN) to learn the relationship between particle shape and a dimensionless defocused distance as a regression problem. With the camera parameters, the true particle height can be deduced. The trained CNN predicts particle height with an R² as high as 0.983 on synthetically generated datasets. Knowing the particle height enables the user to observe complicated 3D motion in fluids. It also benefits the measurement of fluid properties, such as the diffusion coefficient since the displacement of particles is involved and is predicted accurately by considering all three dimensions. Another use of the technique is in the measurement of femtonewton-scale forces in the rapid electronic patterning. The utility of the measurement technique will be demonstrated by measuring the velocity field of a microfluidic electrothermal vortex.