基于深度卷积神经网络的大气湍流强度估算 下载: 1565次
Objective Atmospheric turbulence causes a random fluctuation in the refractive index. When a laser propagates in atmospheric turbulence, the light intensity fluctuation phenomenon during beam propagation occurs, seriously influencing laser propagation. Because different atmospheric turbulence intensities have different effects on laser propagation, it is significantly important to estimate the atmospheric turbulence intensity. In general, the refractive index structural constant Cn2 of the atmospheric turbulence is used to measure the turbulence intensity. The value of Cn2 is directly proportional to the impact of turbulence on laser propagation. Traditional estimation methods include instrument measurement and model estimation. The instrument measurement allows building an experimental platform to directly measure Cn2, in contrast, the model estimation allows obtaining Cn2 by measuring other atmospheric parameters and establishing a model. In recent years, deep learning has allowed achieving good results in the field of image processing, which can extract the feature information of an image layer by layer. This study proposes a method to estimate the refractive index structural constant Cn2 of atmospheric turbulence based on deep convolutional neural networks. The neural network model is built to extract the features of the light spot images under the influence of atmospheric turbulence and the turbulence information is obtained to estimate the turbulence intensity.
Methods A spot image under the turbulence influence contains the turbulence information. In deep learning, neural networks can extract the characteristic parameters of an image. Based on the above mentioned information, neural network models are built to estimate the turbulence intensity. According to the phase screen theory, the Gaussian beam spot images under the influences of different turbulences are simulated. The spot images are divided into a dataset and a test set. Three-thousand images are selected as the training set, and a neural network model is used to obtain the estimation models. Three-hundred images are used as a test set to analyze the estimated results. In addition, the influences of different network structures on the estimation results are analyzed, which provides a new way for estimating turbulence intensity.
Results and Discussions In this study, a traditional AlexNet network model and a VGG16 deep convolutional neural network model are established. VGG16 is optimized on the basis of the traditional convolutional neural network, which increases the layer numbers of the network, reduces the size of the convolution kernel, and has more advantages on feature information extraction of images. The light spot images at different moments under the same turbulence intensity are selected as the inputs of the neural network to verify the feasibility of the above mentioned method and obtain the corresponding estimation results. Moreover, the standard deviation is calculated, and the estimation results are analyzed. The results show that the method can well estimate the turbulence intensity, and the standard deviation increases with the turbulence intensity. To better analyze the results of the neural network model and measure the estimation results, four statistics, i.e., mean absolute error (EMAE), mean relative error (EMRE), root-mean-square variance (ERMSE), and correlation coefficient (Rxy), are selected. The spot images under the influences of different turbulence intensities are randomly selected as the inputs of the neural network model to obtain the corresponding output. The estimation results of the two neural network models are shown in Table 5. After 20 iterations, the estimation result of the VGG16 neural network model is relatively ideal, the correlation coefficient reaches 99%, and EMAE, EMRE, and ERMSE are controlled within 5%. After 500 iterations, EMAE, EMRE, and ERMSE are further reduced to 2%. By analyzing Table 5, it can be seen that both models can well estimate the turbulence intensity after 500 iterations, and the estimation effect of VGG16 is better than that of the AlexNet neural network model. When the number of iterations is the same, EMAE,EMRE, and ERMSE estimated by the VGG16 neural network model are less than half of those of the AlexNet neural network model. Compared with the traditional AlexNet neural network model, the VGG16 neural network model optimizes the network structure and improves the estimation effect to a certain extent.
Conclusion In this study, a method based on deep convolutional neural network model is proposed to estimate turbulence intensity. First, the laser spot images under the influence of turbulence can be simulated according to the classical phase screen theory. Then, the laser spot images under the influence of turbulence are taken as the inputs of the deep convolutional neural network model, and the convolutional layer of the deep convolutional neural network model is used to extract feature information of images layer by layer. After the training of a large number of datasets, the network model is obtained, and the turbulence intensity is estimated. Finally, the estimated effect is analyzed. Compared with the traditional AlexNet neural network model, the VGG16 model adopts a small convolution kernel, which can better retain the image properties, and has high advantages on image feature extraction and better estimation effect. Therefore, the neural network model can be further optimized to improve the estimation effect, which provides a new way to estimate turbulence intensity.
1 引言
当激光在大气中传输时,由大气湍流引起的折射率随机起伏[1]破坏了激光光束的相干性,导致激光发生光强起伏、光束漂移、光束扩展和到达角起伏等现象[2-3],使得激光光束质量下降。因此,了解大气湍流的性质具有重要意义。通常采用折射率结构常数
仪器测量和模式估算是获取
近些年,机器学习算法被广泛应用于气象等领域。Wang等[12]使用人工神经网络,将温度、相对湿度等气象参数作为神经网络的输入,对Mauna Loa附近海面的
本文提出了一种利用大气湍流影响下的高斯光束光斑图像估算
2 大气湍流对激光传播的影响
2.1 大气湍流模拟
通过数值仿真生成随机相位屏以模拟湍流环境,进而研究大气湍流影响下的激光传输[14]。根据波前相位表达式的不同,生成随机相位屏的方法可以分为两类[15-16],即功率谱反演法和Zernike多项式法。相比于Zernike多项式法,功率谱反演法具有可操作性强、结构简单、计算速度快等优点,因此本文采用功率谱反演法生成随机相位屏,并结合多相位屏模型模拟激光在大气湍流中的传播。如
式中:空域内x=mΔx,y=nΔy,Δx、Δy为取样间隔,m、n为整数;波数域内Kx=m'ΔKx,Ky=n'ΔKy,ΔKx、ΔKy为取样间隔,m'、n'为整数;C为常数;R(Kx,Ky)为零均值单位方差的高斯随机数;Fφ(Kx,Ky)是由大气折射率起伏引起的相位畸变的近似功率谱密度。在各向同性的湍流环境下,通常采用Kolmogorov模型作为折射率功率谱:
式中:
式中: k为波数;z为高斯光束传播距离;Δz为相邻两个相位屏之间的距离;Kr=
2.2 湍流影响下的高斯光束
为了模拟高斯光束在大气湍流中的传输,采用基模高斯光束并假设光束沿z轴方向传播,高斯光束的光场表达式为
式中:E0为常数;r2=x2+y2,其中
式中:E(r,zj)为第j个相位屏处的光场;zj+1、zj分别为第j+1个和第j个相位屏的位置;Δzj+1=zj+1-zj;f
表 1. 仿真参数
Table 1. Simulation parameters
|
图 2. 标准高斯光束及不同强度大气湍流影响下的高斯光束光斑图像。(a)标准高斯光束; (b) =1.0083×10-16 m-2/3;(c)
Fig. 2. Spot images of standard Gaussian beam and Gaussian beams under influence of atmospheric turbulence with different intensities. (a) Standard Gaussian beam; (b) =1.0083×10-16 m-2/3; (c)
3 基于深度卷积神经网络的
估算
3.1 VGG16模型
根据上述分析,大气湍流对高斯光束的影响相当于附加了一个相位扰动项,文献[
18]利用深度卷积神经网络对大气湍流影响下的高斯光束光斑图像进行了相位提取。在此基础上,本文提出了一种基于深度卷积神经网络的
深度卷积神经网络在卷积神经网络的基础上增加了神经网络的深度,在图像处理方面更有优势。本文采用的VGG16模型是由牛津大学VGG团队提出的,该模型在2014年ImageNet竞赛定位任务中获得第一名,在分类任务中获得第二名。如
3.2 VGG16估算结果及分析
本文随机选取1.0×10-16~1.0×10-13 m-2/3范围内的
式中:Δi=yi-xi为第i个实际值yi与对应估算值xi之间的误差;
分析
图 4. VGG16模型下lg 估算值与实际值的散点图、频率分布直方图和累积概率分布图。(a1)(a2)(a3)迭代1次; (b1)(b2)(b3)迭代10次;(c1)(c2)(c3)迭代20次;(d1)(d2)(d3)迭代500次
Fig. 4. Scatter plots, frequency distribution histograms, and cumulative probability distribution diagrams of estimation value and actual value of lg by VGG16 model. (a1) (a2) (a3) Number of iterations is 1;(b1) (b2) (b3) number of iterations is 10;(c1) (c2) (c3) number of iterations is 20; (d
表 2. 不同迭代次数下VGG16模型的四个统计量
Table 2. Four statistics of VGG16 model under different numbers of iterations
|
为了进一步验证该方法的可行性,降低随机性对估算效果的影响,将同一湍流强度下不同时刻的高斯光束光斑图像作为测试集,对VGG16神经网络的估算效果进行分析。当
表 3. 不同湍流强度下估算结果的标准差
Table 3. Standard deviation of estimation results under different turbulence intensities
|
3.3 VGG16模型估算效果与AlexNet模型估算效果的对比
为了验证基于深度卷积神经网络的方法在估算大气湍流强度方面的优势,本文用传统的AlexNet神经网络模型对
AlexNet模型的估算结果如
图 5. AlexNet模型下lg 估算值与实际值的散点图、频率分布直方图和累积概率分布图。(a1) (a2) (a3)迭代1次;(b1) (b2) (b3)迭代10次;(c1)(c2) (c3)迭代20次;(d1)(d2) (d3)迭代500次
Fig. 5. Scatter plots,frequency distribution histograms, and cumulative probability distribution diagrams of estimation value and actual value of lg by AlexNet model. (a1) (a2) (a3) Number of iterations is 1; (b1)(b2)(b3) number of iterations is 10; (c1)(c2)(c3) number of iterations is 20; (d1
表 4. 不同迭代次数下AlexNet模型的四个统计量
Table 4. Four statistics of AlexNet model under different numbers of iterations
|
表 5. 不同迭代次数下两种模型的四个统计量
Table 5. Four statistics of two models under different numbers of iterations
|
综合以上分析,可以看出,基于深度卷积神经网络估算大气湍流强度的方法是可行的。相比于传统的AlexNet网络模型,深度卷积神经网络在湍流强度估算上更具优势。随着迭代次数的增加,EMAE、EMRE和ERMSE逐渐减小,相关性越来越高,各项误差也都能控制在较小的数值范围内。
4 结论
提出了一种基于深度卷积神经网络估算
[1] 刘旻雯, 李迎春. OFDM-OAM光信号在大气湍流中的传输[J]. 光学学报, 2019, 39(7): 0706002.
[2] 张冬晓, 陈志斌, 肖程, 等. 基于修正大气谱的湍流相位屏高精度生成方法[J]. 光子学报, 2020, 49(6): 0601002.
[3] 王飞, 余佳益, 刘显龙, 等. 部分相干光束经过湍流大气传输研究进展[J]. 物理学报, 2018, 67(18): 184203.
[4] Wyngaard J C, Izumi Y, Collins S A. Behavior of the refractive-index-structure parameter near the ground[J]. Journal of the Optical Society of America, 1971, 61(12): 1646-1650.
[5] C A. La Rue J C, Champagne F H, et al. Effects of temperature and humidity fluctuations on the optical refractive index in the marine boundary layer[J]. Journal of the Optical Society of America, 1975, 65(12): 1502-1511.
[6] Davidson K L, Schacher G E, Fairall C W, et al. Verification of the bulk method for calculating overwater optical turbulence[J]. Applied Optics, 1981, 20(17): 2919-2924.
[9] 青春, 吴晓庆, 李学彬, 等. WRF模式估算丽江高美古大气光学湍流廓线[J]. 中国激光, 2015, 42(9): 0913001.
[10] 青春, 吴晓庆, 李学彬, 等. 典型地区高空大气光学湍流模拟研究[J]. 光学学报, 2016, 36(5): 0501001.
[11] Qing C, Wu X Q, Li X B, et al. Use of weather research and forecasting model outputs to obtain near-surface refractive index structure constant over the ocean[J]. Optics Express, 2016, 24(12): 13303-13315.
[12] Wang Y, Basu S. Using an artificial neural network approach to estimate surface-layer optical turbulence at Mauna Loa, Hawaii[J]. Optics Letters, 2016, 41(10): 2334-2337.
[13] Lü J. Research on estimating model of atmospheric optical turbulence[D]. Hefei: University of Science and Technology of China, 2017.
吕洁. 大气光学湍流预报模式研究[D]. 合肥:中国科学技术大学, 2017.
[14] 李玉杰, 朱文越, 饶瑞中. 非Kolmogorov大气湍流随机相位屏模拟[J]. 红外与激光工程, 2016, 45(12): 1211001.
[15] 马春波, 黄可, 敖珺. Kolmogorov与Non-Kolmogorov湍流相位屏的数值仿真方法研究[J]. 计算机与数字工程, 2019, 47(12): 2982-2988.
[16] 骆传凯, 卢芳, 苗志芳, 等. 径向阵列涡旋光束在大气中的传输与扩展[J]. 光学学报, 2019, 39(6): 0601004.
[17] 牛超君, 卢芳, 韩香娥. 相位屏法模拟高斯阵列光束海洋湍流传输特性[J]. 光学学报, 2018, 38(6): 0601004.
[18] 徐启伟, 王佩佩, 曾镇佳, 等. 基于深度卷积神经网络的大气湍流相位提取[J]. 物理学报, 2020, 69(1): 014209.
[19] 吕洁, 朱文越, 蔡俊, 等. 两种估算近海面大气光学湍流强度方法的比较[J]. 光学学报, 2017, 37(5): 0501001.
[20] 刘美菊, 曹永战, 朱树云, 等. 基于卷积神经网络的特征融合视频目标跟踪方法[J]. 激光与光电子学进展, 2020, 57(4): 041502.
Article Outline
马圣杰, 郝士琦, 赵青松, 王勇, 王磊. 基于深度卷积神经网络的大气湍流强度估算[J]. 中国激光, 2021, 48(4): 0401018. Shengjie Ma, Shiqi Hao, Qingsong Zhao, Yong Wang, Lei Wang. Atmospheric Turbulence Intensity Estimation Based on Deep Convolutional Neural Networks[J]. Chinese Journal of Lasers, 2021, 48(4): 0401018.