基于深度残差神经网络的光场粒子图像测速粒子场重建方法  

Method of Particle Field Reconstruction in Light Field Particle Image Velocimetry Based on Deep Residual Neural Networks

作　　者：傅梦希 朱效宇 张良 许传龙[1] Fu Mengxi;Zhu Xiaoyu;Zhang Liang;Xu Chuanlong(National Engineering Research Center of Power Generation Control and Safety,School of Energy and Environment,Southeast University,Nanjing 210096,Jiangsu,China;Basic&Applied Research Center,Aero Engine Academy of China,Beijing 101304,China)

机构地区：[1]东南大学能源与环境学院大型发电装备安全运行与智能测控国家工程研究中心,江苏南京210096 [2]中国航空发动机研究院基础与应用研究中心,北京101304

出　　处：《光学学报》2024年第16期144-154,共11页Acta Optica Sinica

基　　金：国家自然科学基金(52306211);中国博士后科学基金(2023M730558);“慧眼行动”项目(08D1C750)。

摘　　要：提出一种基于粒子重构卷积神经网络(PRCNN)模型的快速、高分辨率粒子场重建方法。首先,对原始光场图像进行子孔径图像提取,构建粒子三维分布-光场子孔径图像数据集;然后,建立网络模型,采用自定义的损失函数进行训练获取预测模型。通过粒子场数值重建以及圆柱绕流三维流场测量实验对所提方法的重建准确性等进行评价。结果表明:相比于传统的联合代数重建技术(SART),PRCNN可以有效缓解重建拉伸效应,使粒子场重建质量因子提高了153.83%;基于GPU计算的PRCNN对单幅图像的重建时间仅为0.098 s,加速比达到3976.53;PRCNN与SART计算的速度场分布基本一致,在管道中心深度截面上,相对于平面粒子图像测速(PIV)测量结果,SART和PRCNN的平均相对误差分别为14.56%和12.92%。上述结果证明了PRCNN用于光场PIV技术中实现准确的三维流场测量的可行性。Objective Light field particle image velocimetry(PIV)is a single-camera three-dimensional flow field measurement method and has unique advantages under complex flow field measurement scenarios in narrow channels.The PIV technique consists of three parts:light field image acquisition,tracer particle spatial distribution reconstruction,and interrelated velocity field calculation.Among them,the reconstruction quality of the particle field will directly affect the accuracy and resolution of the velocity field measurement,which is the key link of light field PIV.Traditional particle field reconstruction methods for light field PIV,such as the joint algebraic reconstruction method,have low reconstruction efficiency,large computer memory requirement,and stretching effect of reconstructed particles.Optimization methods for traditional algorithms cannot completely solve the existing problems.Therefore,we introduce deep learning and propose a particle-reconstructed convolutional neural network(PRCNN)model based on deep residual neural networks to improve the quality and reconstruction efficiency of light field PIV particle field reconstruction.Methods Based on the geometric optics theory,we build an optical field imaging model and extract the optical field sub-aperture image containing information of multiple viewing angles from the original image of the optical field according to the optical field imaging characteristics.Meanwhile,the“three-dimensional spatial distribution of particles-optical field sub-aperture image”dataset is constructed by numerical simulations.A deep residual neural network model is built,and a weighted MAE coupled reconstruction quality factor loss function is customized for training,with specific task objectives and data distribution characteristics taken into account.Further,the reconstruction quality and accuracy of the prediction model are evaluated by adopting numerical reconstruction methods,and the reconstruction efficiency is compared with that of the traditional SART algorithm.Finally,t

关 键 词：测量 光场成像 粒子图像测速 三维粒子场 卷积神经网络 三维重建 

分 类 号：O435[机械工程—光学工程]