Particle streak velocimetry using Ensemble Convolutional Neural Networks

This study reports an approach and presents its open-source implementation for quantitative analysis of experimental flows using streak images and Convolutional Neural Networks (CNN). The latter are applied to retrieve a length and an angle from streaks, which can be used to deduce kinetic energy and directionality (up to 180$^{\circ}$ ambiguity) of an imaged flow. We developed a quick method for generating essentially unlimited number of training and validation images, which enabled efficient training. Additionally, we show how to apply an ensemble of CNNs to derive a formal uncertainty on the estimated quantities. The approach is validated on the numerical simulation of a convenctive turbulent flow and applied to a longitutidal libration flow experiment.