基于深度学习的厄米-高斯光束模式分解
Modal Decomposition of Hermite-Gaussian Beams Through Deep Learning
彭姝 1郭旭联 1马天宝 1刘奎 2郜江瑞2
作者信息
- 1. 山西大学光电研究所量子光学与光量子器件国家重点实验室,山西 太原 030006
- 2. 山西大学光电研究所量子光学与光量子器件国家重点实验室,山西 太原 030006;山西大学极端光学协同创新中心,山西 太原 030006
- 折叠
摘要
厄米-高斯(HG)光束模式分解在量子通信和超分辨成像等领域有重要应用,基于深度学习的传统卷积神经网络可以基于单幅光强图实现光场的模式分解且系统构成较为简单.研究了基于新型网络架构Swin Transformer的HG光强图的模式分解和重构.实验上使用HG00、HG10、HG01、HG20、HG11和HG02共6个模(权重和相位是随机值)的叠加光场作为输出光场,通过采集、预处理输出光场强度图和训练网络,完成对输入光场的各阶HG高阶模的强度图像的权重和相位的估计.实验结果表明,当涉及6个本征模时,该方案的权重预测误差为0.05,相位误差为0.15.使用预测出的权重和相位对光强图进行重建,发现输入光强图和预测光强图的相似度很高.这项工作对大容量量子通信、空间量子测量以及超分辨成像具有一定参考意义.
Abstract
Hermite-Gaussian(HG)beam mode decomposition has important applications in quantum communications and super-resolution imaging.The traditional convolutional neural network based on deep learning can realize mode decomposition of light field intensity and phase based on a single intensity image of HG beams,and the system is simple.In this study,the mode decomposition and reconstruction of a Hermite-Gaussian intensity diagram based on the new network architecture of Swin Transformer are investigated.In an experiment,the superimposed light fields of six modes(where the weights and phases are random values)of HG00,HG10,HG01,HG20,HG11,and HG02 are used as the output light fields.Experimental results show that when six eigenmodes are involved,the weight prediction error of the scheme is 0.05 and the phase is 0.15.In reconstructing the intensity image using the predicted weights and phases,we find that the input intensity image is very similar to the predicted intensity image.This work is of certain significance for large-capacity quantum communications,spatial quantum measurements,and super-resolution imaging.
关键词
厄米-高斯光束/模态分解/深度学习/非相干光束Key words
Hermite-Gaussian beams/model decomposition/deep learning/incoherent beams引用本文复制引用
出版年
2024
NETL
NSTL
万方数据