Kinematic training of convolutional neural networks for particle image velocimetry

Abstract Convolutional neural networks (CNNs) offer an alternative to the image cross-correlation methods used in particle velocimetry (PIV) reconstruct fluid velocity field from experimental recording. Despite flexibility of CNNs, accuracy and robustness standard processing remains unsurpassed for general PIV data. As CNNs are non-linear typically entail up millions trainable parameters, they require large carefully designed training datasets avoid over-fitting obtain results that accurate a wide range flow conditions length scales. Most consist PIV-like data generated displacement fields resulting numerical simulations, which, addition being computationally expensive, may be able inform network only about relatively few classes problems. To overcome this issue improve reconstructed by we propose train with synthetic random fields. The underlying idea is dataset simply needs teach kinematic relationship between position velocity. These inexpensive allow much richer variability terms scales varying generation parameters. By state-of-the-art CNN, investigate test cases, such as sinusoidal wind-tunnel turbulent-boundary-layer cylinder-wake experiment. We demonstrate can drastically CNN allows outperform conventional methods, more robust respect noise providing have considerably higher spatial resolution (at pixel level).