Improved Machine Learning Approach for Wavefront Sensing

Hongyang Guo, Yangjie Xu, Qing Li, Shengping Du, Dong He, Qiang Wang, Yongmei HuangSensors 2019 · DOI

TL;DR

This paper introduces a novel deep learning approach, the Deconvolution VGG network (De-VGG), for wavefront sensing in adaptive optics systems. By directly mapping point spread function (PSF) images to wavefront phase maps, it achieves higher accuracy and significantly faster computation times compared to traditional methods and existing machine learning techniques, making it promising for real-time applications.

What Problem Does This Paper Address?

Adaptive optics (AO) systems are crucial for correcting wavefront aberrations in applications like improving human vision and astronomical observation. Wavefront sensing-less (WFS-less) techniques, which rely on intensity measurements from the image plane without a dedicated wavefront sensor, offer advantages like reduced optical complexity and lower energy requirements. However, existing WFS-less methods, particularly those using machine learning, often struggle with accuracy, computational speed, and the ability to handle complex aberrations. Traditional gradient-based optimization methods are time-consuming, and prior neural network approaches that predict Zernike coefficients can be limited by fitting errors and truncation of the Zernike series. This paper aims to address these limitations by proposing a more direct and efficient machine learning model for accurate and rapid wavefront phase map reconstruction.

Methods

The core of this paper is the proposed Deconvolution VGG network (De-VGG). This model is based on the Visual Geometry Group (VGG) architecture, a well-established convolutional neural network (CNN) known for its deep structure and smaller filters. The key innovation is the integration of deconvolution layers within the VGG framework.

Here's a breakdown of the approach:

  • Phase-Map-Output CNN: Unlike previous CNNs that predicted Zernike coefficients, the De-VGG directly outputs the wavefront phase map. This one-to-one correspondence between input PSFs and output phase maps is hypothesized to lead to higher learning and fitting accuracy.
  • Deconvolution Layers: These layers are crucial for reconstructing the unknown input (phase map) from the measured output (PSF). They effectively reverse the convolution process, allowing the network to learn the complex relationship between the distorted image and the underlying wavefront aberration.
  • Architecture: The De-VGG network consists of multiple convolutional layers, batch normalization layers (to improve training stability and prevent overfitting), ReLU activation functions, and importantly, several deconvolution layers. Fully connected layers are omitted to speed up computation.
  • Input: The network takes pairs of PSF images (one in-focus and one defocus) as input. These are normalized to the range [0, 1].
  • Output: The network directly outputs the reconstructed wavefront phase map.

Datasets

The authors generated a dataset for training and testing the De-VGG model.

  • Training Data: A large set of simulated point spread function (PSF) images was generated. These PSFs were created based on Fourier optics principles, incorporating various levels of atmospheric turbulence (simulated by different D/r₀ values, where D is the telescope diameter and r₀ is the atmospheric coherence length). Specifically, 20,000 pairs of in-focus and defocus PSF images were used for training under strong turbulence conditions (D/r₀ = 20).
  • Testing Data: The model was evaluated on separate sets of simulated PSF images for different turbulence levels (D/r₀ = 6, 10, 15, and 20). Additionally, experimental data was collected using a custom-built optical system, with 2000 pairs of PSFs for testing at D/r₀ = 20.

Key Insights and Ideas

  • Direct Phase Map Reconstruction: The shift from predicting Zernike coefficients to directly outputting phase maps is a significant conceptual leap. This bypasses potential errors introduced by fitting approximations and allows the network to learn a more direct, non-linear mapping.
  • Deconvolution Layers as Key Enablers: The inclusion of deconvolution layers is central to the model's ability to "undo" the blurring and distortion introduced by the optical system and atmospheric turbulence, effectively reconstructing the phase information.
  • Speed vs. Accuracy Trade-off: The paper demonstrates that by removing fully connected layers and focusing on deconvolution, the De-VGG achieves remarkable speed improvements (12 ms per calculation) without sacrificing accuracy, a common challenge in deep learning for real-time systems.
  • Robustness to Turbulence: The model shows good performance across a range of turbulence conditions, from medium to strong, indicating its practical applicability.

Key Results

The De-VGG network demonstrated strong performance in both simulations and experiments:

  • Accuracy:
    • In simulations, the root-mean-square (RMS) wavefront error was as low as 0.1307 λ for D/r₀ = 20, 0.0909 λ for D/r₀ = 15, and 0.0718 λ for D/r₀ = 10. These values are reported to be better than existing neural networks.
  • Experimental results for D/r₀ = 20 yielded an RMS error of 0.1304 λ, closely matching the simulation results and validating the model's effectiveness.
  • Speed: The De-VGG network achieves a calculation time of approximately 12 ms per image, which is significantly faster than traditional methods like the Stochastic Parallel Gradient Descent (SPGD) algorithm (which took 304 ms for a similar RMS error under weaker turbulence).
  • Comparison to SPGD: The paper highlights that SPGD struggles with convergence and accuracy as turbulence intensity increases, whereas the De-VGG maintains its performance.

Limitations and Future Work

  • Tip/Tilt Terms: The current De-VGG model does not explicitly consider tip or tilt terms, which are often handled by separate centroiding algorithms. The authors suggest that future work could incorporate these into the training process for a more comprehensive solution.
  • Dynamic Aberration Correction: The current work focuses on static aberrations. The authors propose future research to combine static aberration correction with dynamic aberration estimation, which is crucial for real-time applications where aberrations change rapidly.
  • Aberration Correction: A future goal is to use the reconstructed phase maps to generate corrected PSFs and potentially reconstruct the aberration spot.

Summary updated

Abstract

In the adaptive optics (AO) system, to improve the effectiveness and accuracy of wavefront sensing-less technology, a phase-based sensing approach using machine learning is proposed. In contrast to the traditional gradient-based optimization methods, the model we designed is based on an improved convolutional neural network. Specifically, the deconvolution layer, which reconstructs unknown input by measuring output, is introduced to represent the phase maps of the point spread functions at the in focus and defocus planes. The improved convolutional neural network is utilized to establish the nonlinear mapping between the input point spread functions and the corresponding phase maps of the optical system. Once well trained, the model can directly output the aberration map of the optical system with good precision. Adequate simulations and experiments are introduced to demonstrate the accuracy and real-time performance of the proposed method. The simulations show that even when atmospheric conditions D/r<sub>0</sub> = 20, the detection root-mean-square of wavefront error of the proposed method is 0.1307 λ, which has a better accuracy than existing neural networks. When D/r<sub>0</sub> = 15 and 10, the root-mean-square error is respectively 0.0909 λ and 0.0718 λ. It has certain applicative value in the case of medium and weak turbulence. The root-mean-square error of experiment results with D/r<sub>0</sub> = 20 is 0.1304 λ, proving the correctness of simulations. Moreover, this method only needs 12 ms to accomplish the calculation and it has broad prospects for real-time wavefront sensing.
BibTeX
@inproceedings{guo2019improved,
  author = {Hongyang Guo and Yangjie Xu and Qing Li and Shengping Du and Dong He and Qiang Wang and Yongmei Huang},
  title = {{Improved Machine Learning Approach for Wavefront Sensing}},
  booktitle = {Sensors},
  year = {2019},
  doi = {10.3390/s19163533},
}

Loading related content...

Figure