-
本文提出了一种基于卷积神经网络(convolutional neural network, CNN)与改进视觉变换器(vision transformer, VIT)的涡旋光束叠加态轨道角动量(orbital angular momentum, OAM)模式识别方法. 以海洋湍流畸变的三组拉盖尔-高斯光束模式叠加光场强度分布图为输入, 有机整合了CNN的局部特征提取优势与稀疏注意力机制驱动的VIT的全局快速分类能力, 实现端到端的波前畸变高效精准识别. 通过数值仿真模拟海洋湍流环境叠加态OAM模式, 利用功率谱反演法模拟海洋湍流, 以识别准确率和混淆矩阵作为OAM模式识别的评估指标. 实验结果表明, CNN-VIT模型在不同海洋湍流强度、波长、传输距离和模式间隔条件下均展现出优异的OAM模式识别准确率性能. 与现有的CNN和VIT相比, 本文模型在强海洋湍流条件下识别准确率分别提升了23.5%和9.65%. 这证明了CNN-VIT模型在涡旋光叠加态OAM模式识别的应用潜力.This work proposes a pattern recognition method for the superposition state orbital angular momentum (OAM) of vortex beams based on convolutional neural network (CNN) and improved vision transformer (VIT). Organically integrating the local feature extraction advantages of CNN with the global fast classification ability of VIT driven by sparse attention mechanism, using three sets of Laguerre-Gaussian (LG) beam patterns with superimposed light field intensity distribution maps of ocean turbulence distortion as input, efficient and accurate recognition of end-to-end wavefront distortion is realized. MATLAB numerical simulation is adopted to simulate the superposition state LG beam in ocean turbulent environment, power spectrum inversion method is used to simulate ocean turbulence, and recognition accuracy and confusion matrix are used as evaluation indicators for OAM pattern recognition. The experimental results show that the CNN-VIT model exhibits excellent performance in OAM pattern recognition accuracy under different ocean turbulence intensities, wavelengths, transmission distances, and mode intervals. Compared with existing CNN and VIT, the proposed model improves recognition accuracy by 23.5% and 9.65% respectively under strong ocean turbulence conditions, thus exhibiting strong generalization ability under unknown ocean turbulence strengths. This demonstrates the potential application of the CNN-VIT model in OAM pattern recognition of vortex light superposition states.
-
Keywords:
- vortex beam /
- ocean turbulence /
- convolutional neural network /
- mode recognition
-
图 1 3组OAM模式集中16种叠加LG光束强度分布图像 (a) Set1, Δl = 1; (b) Set2, Δl = 3; (c) Set3, Δl = 5
Fig. 1. The 16 superimposed LG beam intensity images in three OAM mode sets: (a) Set1, Δl = 1; (b) Set2, Δl = 3; (c) Set3, Δl = 5.
图 2 功率谱反演法模拟海洋湍流光强扰动
Fig. 2. Simulation of oceanic turbulence using power spectrum inversion method.
图 4 不同湍流情况下识别准确率随迭代次数的变化
Fig. 4. Variation of recognition accuracy with the number of iterations under different turbulence conditions.
图 5 不同距离(a)和不同波长(b)下识别准确率随迭代次数的变化
Fig. 5. Variation of recognition accuracy with iteration count at different distances (a) and different wavelengths (b).
图 6 不同模型识别准确率随传输距离(a)和湍流强度(b)的变化
Fig. 6. Variation of recognition accuracy of different models with transmission distance (a) and turbulence intensity (b).
图 8 不同湍流强度下识别准确率随不同模式间隔的变化
Fig. 8. Variation of recognition accuracy with different mode intervals under different turbulence intensities.
表 1 LG光束模式选择
Table 1. LG beam mode selection.
组别 拓扑荷数间隔 叠加OAM模式 Set1 $ \Delta l=1 $ $ \left\{{p}_{2}=0, 1, 2, 3;\right.\left.{l}_{2}=-2, -1, 0, 1\right\} $ Set2 $ \Delta l=3 $ $ \left\{{p}_{2}=0, 1, 2, 3;\right.\left.{l}_{2}=-5, -2, 1, 4\right\} $ Set3 $ \Delta l=5 $ $ \left\{{p}_{2}=0, 1, 2, 3;\right.\left.{l}_{2}=-8, -3, 2, 7\right\} $ 
下载: 导出CSV
表 2 硬件平台配置和模型参数
Table 2. Hardware platform configuration and model parameters.
配置 型号 参数设置 数值 操作系统 Windows10 迭代次数 100 CPU Intel40核E5
2670处理器批次值 16 GPU NVIDIA
RTX3090学习率 0.01 内存 64 G 标签平滑 Le-5 CDUA 11 学习率
衰减模型Cosine Annealing
LrUpdater编程语言 Python3.8 优化器类型 Adam
下载: 导出CSV
-
[1] Erhard M, Fickler R, Krenn M, Zeilinger A 2018 Light: Sci. Appl. 7 17146
[2] Tamburini F, Anzolin G, Umbriaco G, Bianchini A, Barbieri C 2006 Phys. Rev. Lett. 97 163903
Google Scholar
[3] Shen Y J, Wang X J, Xie Z W, Min C J, Fu X, Liu Q, Gong M L, Yuan X C 2019 Light: Sci. Appl. 8 90
Google Scholar
[4] Molina Terriza G, Torres J P, Torner L 2001 Phys. Rev. Lett. 88 013601
Google Scholar
[5] Wang J, Yang J Y, Fazal I M, Ahmed N, Yan Y, Huang H, Ren Y X, Yue Y, Dolinar S, Tur M, Willner A E 2012 Nat. Photonics 6 488
Google Scholar
[6] Ren Y X, Huang H, Xie G D, Ahmed N, Yan Y, Erkmen B I, Chandrasekaran N, Lavery M P J, Steinhoff N K, Tur M, Dolinar S, Neifeld M, Padgett M J, Boyd R W, Shapiro J H, Willner A E 2013 Opt. Lett. 38 4062
Google Scholar
[7] Zhu X L, Guo L, Zhu Q 2018 IEEE Photonics J. 10 135
[8] Fan W Q, Gao F L, Xue F C, Guo J J, Xiao Y, Gu Y J 2024 Appl. Opt. 63 982
Google Scholar
[9] Krizhevsky A, Sutskever I, Hinton G E 2017 Commun. ACM 60 84
Google Scholar
[10] 杨春勇, 闪开鸽 2020 中南民族大学学报(自然科学版) 39 390
Yang C Y, Shan K G 2020 J. South-Central Minzhu Univ. (Nat. Sci. Ed. ) 39 390
[11] Wang X Y, Wang Z Y, Cheng Z Y, Pu J X 2022 Chin. J. Quantum Electron. 39 955
[12] Wang Z L, Li X F, Cai Y J, Liu X L 2025 Opt. Express 33 12591
Google Scholar
[13] 郭焱, 吕恒, 丁春玲, 袁晨智, 金锐博 2025 物理学报 74 014203
Google Scholar
Guo Y, Lü H, Ding C L, Yuan C Z, Jin R B 2025 Acta Phys. Sin. 74 014203
Google Scholar
[14] 张成志, 曹阳, 涂巧玲, 彭小峰 2023 激光杂志 22 062235
Zhang C Z, Cao Y, Tu Q L, Peng X F 2023 Laser J. 22 062235
[15] Zhou X, Chen C Y, Yu H Y 2023 Laser Optoelectron. Prog. 60 2306003
[16] Wang J, Wang C B, Tan Z K, Lei S C, Wu P F 2024 Sci. Sin. Phys. Mech. Astron. 54 284211
Google Scholar
[17] Lu J N, Cao C Y, Zhu Z Q, Gu B 2020 Appl. Phys. Lett. 116 201105
Google Scholar
[18] 吴鹏飞, 王小蝶, 王姣, 谭振坤, 贾致远 2023 光学学报 43 1126001
Google Scholar
Wu P F, Wang X D, Wang J, Tan Z K, Jia Z Y 2023 Acta Opt. 43 1126001
Google Scholar
[19] 蔡冬梅, 王昆, 贾鹏, 王东, 刘建霞 2014 物理学报 63 104217
Google Scholar
Cai D M, Wang K, Jia P, Wang D, Liu J X 2014 Acta Phys. Sin. 63 104217
Google Scholar
[20] Nikishov V 2000 Int. J. Fluid Mech. Res. 27 82
Google Scholar
[21] Hu J, Shen L, Albanie S, Sun G, Wu E H 2020 IEEE Trans. Pattern Anal. Mach. Intell. 42 2011
Google Scholar
[22] Wang T T, Cao D H, Han J Z 2022 Front. Neurosci. 16 5
下载:
计量
- 文章访问数: 646
- PDF下载量: 11
- 被引次数: 0
