High Power Laser and Particle Beams, Volume. 33, Issue 8, 081004(2021)

Research progress in deep learning based WFSless adaptive optics system

Zhiguang Zhang... Huizhen Yang*, Jinlong Liu, Songheng Li, Hang Su, Yuxiang Luo and Xiewen Wei |Show fewer author(s)
Author Affiliations
  • School of Electrical Engineering, Jiangsu Ocean University, Lianyungang, 222005, China
  • show less
    Figures & Tables(29)
    Perceptron artificial neural network for phase retrieval[18]
    Modified Inception v3CNN model for predicting Zernike coefficients[20]
    Zernike coefficients predicting results of focused target[23]
    Zernike coefficients predicting results of overexposed target[23]
    WFSless system architecture[26]
    Architecture of CNN[26]
    Data flow of training and predictions[29]
    Residual wavefront RMS with and without compensation under different turbulence levels[29]
    Zernike coefficients prediction results by models trained with dataset of different turbulence levels[31]
    Trained neural network is optimized by TensorRT to build the inference engine for implementation[35]
    Wavefront Net (WFNet)[36]
    CNN architecture[38]
    Standard deviation of phase before and after phase aberration revision[38]
    An object irrelevant wavefront sensing scheme using LSTM neural network[41]
    Image restoration results based on wavefront error inferred by LSTM[41]
    Prediction results of the next 5 frames wavefront made by LSTM[44]
    Reinforcement Learning(RL) of WFSless AO[47]
    Intensity distribution of point target with wavefront error and that after restoration by deep RL[47]
    PSD of different vibration frequency[51]
    Residual phase[51]
    Principle of aberration correction in high resolution optical microscopes[59]
    Schematic diagram of MAL-WSAO-SS-OCT system[63]
    • Table 1. Accuracy of Zernike coefficients (RMS) with overexposure, defocus and scattering preprocessing[23]

      View table
      View in Article

      Table 1. Accuracy of Zernike coefficients (RMS) with overexposure, defocus and scattering preprocessing[23]

      Zernike coefficients
      in-focusoverexposuredefocusscatter
      point source0.142±0.0320.036±0.0130.040±0.0160.057±0.018
      extended source0.288±0.0240.214±0.0510.099±0.0640.195±0.064
    • Table 2. Dataset of three different turbulence levels[31]

      View table
      View in Article

      Table 2. Dataset of three different turbulence levels[31]

      dataset No.$ D/{r}_{0} $$ D/{r}_{0} $ interval data volume/ interval total data volume $D/{r}_{0}$ interval data volume/intervaltotal data volume
      training datasettest dataset
      1510015000101500
      21510015000101500
      31-151100150001101500
    • Table 3. Simulation results of wavefront restoration error under different turbulence levels (NPMS: Normalized Pixel Mean Square; RMS: Root Mean Square)[32]

      View table
      View in Article

      Table 3. Simulation results of wavefront restoration error under different turbulence levels (NPMS: Normalized Pixel Mean Square; RMS: Root Mean Square)[32]

      $ D/{r}_{0} $NPMSRMS/λ
      $ 20 $0.00670.1307
      $ 15 $0.00410.0909
      $ 10 $0.00290.0718
      $ 6 $0.00250.0703
    • Table 4. Wavefront restoration error and time consumption of experiments (NPMS: Normalized Pixel Mean Square; RMS: Root Mean Square)[32]

      View table
      View in Article

      Table 4. Wavefront restoration error and time consumption of experiments (NPMS: Normalized Pixel Mean Square; RMS: Root Mean Square)[32]

      $ D/{r}_{0} $NPMSRMS/λrunning time/ms
      $ 20 $0.00660.1304~12
    • Table 5. Comparison of inference time of PD-CNN with that of Xception[34]

      View table
      View in Article

      Table 5. Comparison of inference time of PD-CNN with that of Xception[34]

      networkfocal model/msdefocused model/msPD model/ms
      PD-CNN2.24952.29892.5591
      Xception10.46910.110810.469
    • Table 6. Comparison of inference time with and without optimization by TensorRT[34]

      View table
      View in Article

      Table 6. Comparison of inference time with and without optimization by TensorRT[34]

      modelbefore acceleration/msafter acceleration/msacceleration ratio
      focal model2.24950.46784.8091
      defocused model2.29890.44065.2178
      PD model2.55910.49095.2135
    • Table 7. WFSless AO simulation software

      View table
      View in Article

      Table 7. WFSless AO simulation software

      AO simulation tooldeep learning framework
      Soapy:Simulation ‘OptiqueAdaptative’ with Python    HCIPy:High Contrast Imaging for Python OOMAO:Object-Oriented MATLAB Adaptive Optics Toolbox    YAO:Yorick Adaptive Optics DASP: Durham Adaptive Optics Simulation Platform
      Soapy[21]PyTorch (www.pytorch.org)
      HCIPy[52]Keras (www.keras.io)
      OOMAO[54]TensorFlow (www.tensorflow.org)
      YAO[55]MATLAB + Deep Learning Toolbox
      DASP[56]Caffe (https://caffe.berkeleyvision.org/)
    Tools

    Get Citation

    Copy Citation Text

    Zhiguang Zhang, Huizhen Yang, Jinlong Liu, Songheng Li, Hang Su, Yuxiang Luo, Xiewen Wei. Research progress in deep learning based WFSless adaptive optics system[J]. High Power Laser and Particle Beams, 2021, 33(8): 081004

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Laser Atmosphere Propagation?Overview

    Received: Jul. 19, 2021

    Accepted: --

    Published Online: Sep. 3, 2021

    The Author Email: Yang Huizhen (yanghz526@126.com)

    DOI:10.11884/HPLPB202133.210295

    Topics