A lightweight optical flow model for particle image velocimetry

Abstract

Particle Image Velocimetry (PIV) is a reliable non-contact velocimetry technique in fluid visualization, in which the method incorporating deep neural networks outperforms the mainstream conventional algorithms in terms of reconstruction accuracy and computation time. However, some optical flow models suffer from large parameter quantities and high complexity. This paper proposes a lightweight network, which combines blueprint convolution and attention mechanism based on the improved LiteFlowNet3 with supervised learning. The network uses blueprint convolution instead of ordinary convolution to compress the number of model parameters. An attention mechanism is introduced to enhance the extraction of particle image features and improve the accuracy of flow inference. It aims to reduce the number of model parameters and computational complexity while maintaining model accuracy. The experimental results show that the improved lightweight model reduces the average endpoint error by 37.71 % compared to the pre-improved model. Finally, a 2D PIV experimental system was constructed to capture particle image pairs. By applying this model to reconstruct the velocity field, the experimental results proved its feasibility and adaptability in processing real-world images.