基于光流卷积网络的粒子图像测速自动掩模及速度场计算  

Automatic mask and velocity field calculation of particle image velocimetry based on optical flow convolution network

作　　者：郭春雨[1] 范毅伟 韩阳[1] 于长东 徐鹏 毕晓君 GUO Chun-yu;FAN Yi-wei;HAN Yang;YU Chang-dong;XU Peng;BI Xiao-jun(College of Shipbuilding Engineering,Harbin Engineering University,Harbin 150001,China;College of Information Engineering,Harbin Engineering University,Harbin 150001,China;College of Information Engineering,Minzu University of China,Beijing 100081,China)

机构地区：[1]哈尔滨工程大学船舶工程学院,哈尔滨150001 [2]哈尔滨工程大学信息通信学院,哈尔滨150001 [3]中央民族大学信息工程学院,北京100081

出　　处：《船舶力学》2024年第3期379-391,共13页Journal of Ship Mechanics

基　　金：中央高校基本科研业务费项目(3072020CFT0104)。

摘　　要：粒子图像测速(PIV)技术是一种定量的非接触式全局速度场测量技术。在船舶与海洋工程领域,PIV实验中拍摄的粒子图像常出现结构物遮挡或自由液面等干扰现象,需要对其进行掩模后计算液相区域速度场。因此,实现PIV图像中干扰区域自动掩模及液相区域速度场高精度计算具有重要的意义。本文基于光流卷积神经网络LiteFlowNet,设计了一种可实现自动掩模及速度场计算的深度学习模型Mask-PIV-LiteFlowNet,并使用基于物体入水PIV实验图像掩模数据集和PIV速度场计算数据集制作的数据集对其进行训练和测试。测试结果表明,该模型能够有效减少临近掩模边界区域的速度场计算错误并能够精细地提取流场小尺度流动信息,相比于当前先进的PIV深度学习模型PIV-LiteFlowNet-en,本文提出的模型在对带结构物的合成粒子图像进行流场计算时精度获得了至少14.5%的提升,计算速度上获得了5.7%的提升。最后,使用楔形体入水PIV图像对提出的模型进行了测试,验证了模型的泛化能力。Particle image velocimetry(PIV)technology is a non-contact global velocity field measurement technology.In the field of shipbuilding and ocean engineering,the particle images taken in the PIV experi⁃ment often contain interference such as structure occlusion and free liquid surface,which needs to be masked before the liquid phase velocity field is calculated.Therefore,it is of great significance to realize the automat⁃ic masking of the interference area in the PIV image and the high-precision calculation of the velocity field in the liquid phase area.In this paper,based on the optical flow convolutional neural network LiteFlowNet,a deep learning model Mask-PIV-LiteFlowNet that can realize automatic mask and velocity field calculation was designed.Furthermore,based on the PIV mask dataset of the object entering the water and on the PIV ve⁃locity field calculation data set,a data set was made to train and test.The test results show that the model can effectively reduce the calculation errors of the velocity field near the boundary of the mask and can extract small-scaled flow information of the flow field finely.Compared with the current advanced particle image ve⁃locimetry deep learning model,the calculation accuracy was improved by more than 20%,and the calculation speed was improved by 5.7%.Finally,the proposed model was tested with the actual images of the wedge�shaped body entering the water and the carp swimming PIV,verifying that the model has a strong generaliza⁃tion ability.

关 键 词：粒子图像测速 深度学习 自动掩模 速度场计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]