Three-dimensional particle image velocimetry measurement through three-dimensional U-Net neural network

This paper proposes a light field (LF) three-dimensional (3D) particle image velocimetry (PIV) method based on a digital refocused algorithm and 3D U-Net neural network for 3D three-component (3D-3C) velocity measurement. A digital refocused algorithm is used to generate a stack of LF-refocused images of tracer particles for establishing the 3D U-Net. The 3D U-Net is then used for the 3D particle field reconstruction. Based on a pair of 3D particle fields, the 3D-3C velocity field is obtained through a 3D cross correlation algorithm. Numerical simulations and experiments are conducted to analyze the accuracy and efficiency of the proposed method. The simulation results show that the elongation along the depth direction and the efficiency of the 3D particle field reconstruction are improved by the 3D U-Net. The 3D U-Net also provides a better correlation coefficient. The experimental results show that the reconstruction time of the proposed method is ∼220 s which is 10 times faster than the LF tomographic PIV. This further demonstrates that the proposed method improves the reconstruction efficiency without affecting the accuracy of velocity measurement.