Particle Image Velocimetry Estimation Method Based on Lightweight Feature Fusion Network

Particle Image Velocimetry (PIV) technology is an important experimental method for near-instantaneous, non-intrusive measurement of fluid velocity fields. In recent years, with the rapid development of deep learning, the use of convolutional neural networks can effectively improve the estimation accuracy of PIV. However, these methods often suffer from the issue of large network parameter sizes. To effectively reduce the parameter size of deep networks for PIV, a lightweight feature fusion network-based method for PIV is proposed in this paper. Firstly, a lightweight backbone network is designed based on depthwise separable convolutions and enhanced ShuffleNet modules to perform multi-scale feature encoding on particle velocity images. Secondly, a feature fusion and decoding module is designed to hierarchically fuse and decode the extracted fluid field features. Finally, the flow field velocity is inferred and predicted based on the fused fluid field features. The proposed method is quantitatively validated using publicly available datasets. Experimental results demonstrate that the estimation accuracy of the proposed method outperforms traditional PIV methods across various types of flow fields, achieving an average accuracy improvement of 48% compared to the H-S optical flow method. It also surpasses several deep learning-based methods, with an average accuracy improvement of 42% over the LiteFlowNet-en method. Furthermore, the proposed method's model parameter size is only 0.19M, which is 3.4% of LiteFlowNet-en's parameter size and 5.1% of PIV-RAFT-2P's parameter size, significantly enhancing the computational efficiency of PIV estimation.