基于深度学习光流法的荧光油膜全局速度测量

Global velocity measurement of fluorescent oil film based on deep learning optical flow method

针对基于先验的传统光流法存在前提条件苛刻的问题,提出使用基于深度学习的光流法进行荧光油膜全局速度测量。采用数值仿真试验对基于先验的改进HS光流法和基于深度学习的FlowNet2光流法进行对比,结果显示:在不外加干扰时,改进HS光流法和FlowNet2光流法的平均端点误差分别为0.4587像素/s和0.3817像素/s;在亮度变化、噪声干扰或不同的演化时间下,FlowNet2光流法的平均端点误差均明显低于改进HS光流法,平均端点误差差值最大可达5.19像素/s;风洞试验进一步证明,FlowNet2光流法能够获得正确、清晰、定量的荧光油膜全局速度场,较改进HS光流法鲁棒性更高,对风洞工程应用具有一定的参考价值。

In order to solve the problems of the traditional optical flow method based on priori,such as harsh preconditions,an optical flow method based on deep learning was proposed to measure the global velocity of fluorescent oil film.The numerical simulation experiments were used to compare the improved HS optical flow method based on prior with FlowNet2 optical flow method based on deep learning.The results showed that the average endpoint errors of improved HS optical flow method and FlowNet2 optical flow method were 0.458 7 pixel/s and 0.381 7 pixel/s without external interference,respectively;the average endpoint error of FlowNet2 optical flow method was significantly lower than that of HS optical flow method under the conditions of brightness change,noise disturbance or different evolution times,and the maximum difference of average endpoint errors could reach 5.19 pixel/s.The experimental results in the wind tunnel further prove that the FlowNet2 optical flow method can obtain the correct,clear and quantitative global velocity fields of fluorescent oil film.With stronger robustness than the improved HS optical flow method,this method has certain reference value for wind tunnel engineering application.