Application of Convolutional Neural Networks to Particle Image Velocimetry

In the last decade, deep learning architectures have overwhelmingly outperformed the state-of-the-art in many traditional computer vision tasks as image classification and object detection, so that it becomes interesting to test them on new tasks such as regression. Particle image velocimetry (PIV) is an experimental fluid dynamics technique that involves image analysis. PIV processing represents a complex regression problem. Therefore, it is interesting to test the ability of artificial neural networks (ANNs) to perform PIV. The first feasibility study on ANN application to PIV was "Performing particle image velocimetry using artificial neural networks: a proof-of-concept" published in 2017 by Rabault et al. The present work aims to build and train a convolutional neural network (CNN) to perform standard PIV and to test the model on realistic PIV data. More dense particle images and higher displacement than those studied by Rabault et al. are considered. Both training and test phase are performed using Matlab 2019b with Deep Learning Toolbox. Results are promising, although state-of-the-art performances are still not achieved. Moreover, performances can be improved by increasing image resolution and by computing the mean field from both several particle images and predictions of different trained cases related to the same network architecture, despite higher computational cost.