Peak-CNN: improved particle image localization using single-stage CNNs

An important step in the application of Lagrangian particle tracking (LPT) or in general for image-based single particle identification techniques is the detection of particle image locations on the measurement images and their sub-pixel accurate position estimation. In case of volumetric measurements, this constitutes the first step in the process of recovering 3D particle positions, which is usually performed by triangulation procedures. For two-component 2D measurements, the particle localization results directly serve as input to the tracking algorithm. Depending on the quality of the image, the shape and size of the particle images and the amount of particle image overlap, it can be difficult to find all, or even only the majority, of the projected particle locations in a measurement image. Advanced strategies for 3D particle position reconstruction, such as iterative particle reconstruction (IPR), are designed to work with incomplete 2D particle detection abilities but even they can greatly benefit from a more complete detection as ambiguities and position errors are reduced. We introduce a convolutional neural network (CNN) based particle image detection scheme that significantly outperforms current conventional approaches, both on synthetic and experimental data, and enables particle image localization with a vastly higher completeness even at high image densities.