粒子图像测速(PIV)作为一种非接触的、全局的间接流体力学测量技术,能够从图像中获取流体的速度场,从而揭示流体的运动规律。随着深度学习技术的发展,用深度学习技术来进行粒子图像测速具有很重要的研究价值和广泛的应用前景。基于光流神经网络,提出了一种改进型轻量级卷积神经网络,在提高粒子图像测速精度的同时,减小了模型的参数量,提高了测试速度。首先,将目前能够获取最优刚体估计的光流神经网络架构进行了改进,采用人工合成的粒子图像数据集进行有监督训练。然后,将训练得到的网络模型与当前最先进的用于粒子图像测速的深度学习模型进行对比评估。实验结果表明,本文提出的基于轻量化深度学习模型的粒子图像测速模型在不损失精度的同时,模型参数量减小了9.5%,测试速度提高了8.9%。

Particle image velocimetry (PIV), as a non-contact, global indirect hydrodynamics measurement technique, can capture the velocity field of a fluid from an image to reveal the motion of the fluid. The development of deep learning technology and its use for PIV have significant research value and a potentially wide range of applications. In this paper, the authors propose an improved lightweight convolutional neural network based on the optical flow neural network. The proposed method improves the accuracy of particle image velocity measurement while reducing the parameter quantity of the model and improving the test speed. First, this work improves the optical flow neural network architecture with superior rigid body estimation performance, and uses an artificial particle image dataset for supervised training. The trained network model is then compared with a state-of-the-art PIV deep learning model. Experimental results indicate that the PIV based on the lightweight deep learning model proposed in this paper can reduce the number of model parameters by 9.5% and improve the test speed by 8.9% without losing accuracy.