Improved particle peak detection in images using convolutional neural networks

Abstract

An important step in the application of Lagrangian particle tracking (LPT) techniques is the detection of particle peaks on the measurement images, the positions of which are then used for triangulation. Depending on the quality of the image, the shape and size of the particles, and the amount of particle overlap, it can be difficult to find all particle peaks in a measurement image. LPT techniques do not necessarily need a perfect solution to this problem as they can either utilize time-resolved temporal information to minimize their reliance on peak detection [1] or utilize an iterative approach [2] to recover particle positions. However, there is still value in improving the quality of peak detection as this is the first step in the LPT processing pipeline and improved detection and position accuracy has the potential to affect all following steps. This is especially true for two- or multi-pulse applications which rely much more on a good reconstruction of each individual snapshot than the time resolved applications. We introduce a supervised machine learning based approach utilizing convolutional neural networks (CNN) to improve particle peak detection on LPT measurement images.