Multiscale Feature Convolutional Neural Network for Particle Image Velocimetry

Particle image velocimetry (PIV) is a crucial technique in experimental fluid dynamics for non-invasively measuring the velocity components of flow fields. Deep learning methods applied to PIV for velocity field estimation are primarily derived from computer vision. However, these methods often fail to account for the distinctive characteristics of fluid dynamics, such as the very small scales of objects and displacements relative to pixel resolution, high-density particle distributions, and complex and irregular motion patterns. As demonstrated in this work, these approaches encounter three main challenges: 1) failure to accurately estimate the motion for very small particles and displacements; 2) significant computational discrepancies with high-density particle distributions; and 3) considerable inaccuracies in the analysis of complex flow fields. To address these issues, we propose a novel convolutional neural network, named multiscale feature recurrent all-pairs field transforms (MSF-RAFT). Our approach involves: 1) developing a high-resolution multiscale feature extraction network; 2) integrating an innovative residual block to improve the network; and 3) creating a multiscale feature correction module that constructs a four-layer correlation pyramid. The proposed network has been validated on various datasets and experimental images. MSF-RAFT exhibits superior performance, particularly with high-density small particle images, achieving a significant reduction in estimation error and accurately capturing small vortex details. Comparative analyses with benchmark models demonstrate that MSF-RAFT achieves up to a 33% reduction in estimation error compared to previous best models on complex datasets and provides a flow field estimation that more accurately represents the fluid dynamics around the nacelle. Additionally, evaluations of model size, runtime, and memory consumption demonstrate that MSF-RAFT is highly efficient, making it well-suited to meet the industrial demands for high accuracy and rapid response.