Pyramidal deep-learning network for dense velocity field reconstruction in particle image velocimetry

Particle image velocimetry (PIV) is a broadly used technique in fluid dynamics experiments. Traditional velocity derivation algorithms, e.g., the cross-correlation-based WIDIM (widow deformation iterative multi-grid method) and optical flow-based HS (Horn–Schunck), are still somewhat challenged by computational efficiency and model robustness, as well as their sensitivity to parameter settings such as the size of interrogation window. A few recent works have proposed different velocity calculation methods based on deep learning and present superiority on addressing the challenges, but the methods still need further exploration on the relevant accuracy and robustness. This paper reports an end-to-end convolutional neural network, namely PIV-PWCNet, to reconstruct the dense velocity field from particle image pairs. The main aim is to improve the accuracy and robustness of the velocimetry algorithms, meanwhile maintain a low computational cost. The PIV-PWCNet uses an available net, PWCNet, as the flow constructing backbone and makes special modification and enhancement to the network structure, flow estimator, refinement network, and loss function to allow the new model applicable for PIV measurements. The proposed PIV-PWCNet was tested on both synthesized and experimental particle images, showing higher accuracy than the cross-correlation-based WIDIM method, presenting superior robustness than the optical flow-based HS method, and outperforming the deep-learning-based PIV-NetS and PIV-DCNN models in accurately recovering flow details. In addition, the proposed PIV-PWCNet also has advantages such as high vector content and image processing efficiency.