Deep learning-based spatial refinement method for robust high-resolution PIV analysis

Abstract

In this study, we propose a deep learning-based spatial refinement method to provide robust high-resolution velocity fields for particle image velocimetry (PIV) analysis. We modified the architecture of the convolutional neural network (CNN)-based optical flow model, FlowNet2, to receive the subdomain of particle image pair and provide sub-velocity fields to better train the network in small-scale flows. The modified deep learning model was trained with synthetic particle image datasets replicating the various PIV particle conditions. Then, we incorporated the modified CNN into the correlation-based spatial refinement approach to achieve robust multiscale flow measurement. The proposed method was quantitatively validated using synthetic particle image pairs of simple and complex flows. The proposed method successfully refined the velocity fields and estimated the large displacement with good accuracy, whereas the modified deep learning model was trained only for small displacement. The validation results demonstrated that the proposed method is robust under various displacement, particle density, and velocity gradient conditions. The evaluation on real PIV particle image pair in turbulent jet flow and flow over a hill showed that the proposed method captured turbulence structures with high-resolution velocity field and improved performance in multiscale flow measurement. The CNN-based approach has advantages over the classic correlation and global optical flow methods in obtaining velocity fields without averaging, smoothening, nor further parameter tuning processes.