Application of a Deep Neural Network to Phase Retrieval in Inverse Medium Scattering Problems

Soojong Lim, Jaemin ShinComput 2021 · OpenReview

TL;DR

This paper tackles the challenging problem of reconstructing the refractive index of a medium from incomplete (phaseless) scattering data, a crucial task in applications like nondestructive testing of optical fibers. They propose a deep neural network (DNN) approach to approximate the inverse mapping, effectively overcoming the ill-posed nature of the problem and demonstrating promising results in numerical experiments.

What Problem Does This Paper Address?

The core problem addressed is inverse medium scattering with phaseless data. Imagine trying to understand the internal structure of an object (like an optical fiber) by sending waves at it and observing how they scatter. Ideally, we'd want to know both the amplitude and the phase of the scattered waves to perfectly reconstruct the object's properties (in this case, the refractive index, denoted by q). However, measuring the phase information is often difficult, expensive, or even impossible, especially at high frequencies. This leaves us with only the amplitude of the scattered waves, leading to a "phaseless" inverse problem.

This problem is particularly challenging because:

  • Non-uniqueness: Even with complete data, phase retrieval can have multiple solutions.
  • Ill-posedness: Small errors in the measured amplitude data can lead to vastly different and incorrect reconstructions of the refractive index. This sensitivity makes traditional methods struggle.

The authors are motivated by the need for nondestructive testing of optical fibers, where manufacturing defects can lead to variations in the refractive index, impacting performance. They aim to recover the refractive index profile q(r) from the modulus of the total field data at a specific boundary.

Methods

The paper proposes using a Deep Neural Network (DNN), specifically a Multilayer Perceptron (MLP), to learn the inverse mapping from measured data to the refractive index. Here's a breakdown of their approach:

  1. Problem Formulation: They start with the 2D Helmholtz equation governing wave scattering and define an operator T that maps the refractive index q to the measured amplitude of the total field u. The goal is to find T⁻¹(u).
  2. Radial Symmetry Assumption: To simplify the problem and make it tractable, they assume the refractive index q is radially symmetric, meaning q only depends on the radial distance r (q = g(r)). This significantly reduces the complexity of the inverse problem.
  3. Discretization: The function spaces for q and the measured data S are discretized into finite-dimensional spaces. The refractive index q(r) is represented by a set of coefficients C = (C₁, C₂, ..., Cм)ᵀ using a basis of sine functions. The measured data S is represented by its Fourier coefficients.
  4. DNN Training: An MLP is trained to approximate the inverse operator T⁻¹. The input to the network is the discretized measurement data (Fourier coefficients of S), and the output is the set of coefficients C representing the refractive index q. The network learns to predict these coefficients from the data.
  5. Loss Function and Optimizer: They use the mean squared error as the loss function and the Adam optimizer with a learning rate of 0.00001 for training.
  6. Regularization via DNN: The DNN acts as a form of regularization, converting the ill-posed inverse problem into a well-posed learning problem by leveraging "group data fidelity" (i.e., learning from many data pairs).

Datasets

The authors generated their own datasets for training and evaluation.

  • Dataset Size: A total of 12,000 numerical data sets were generated.
  • Data Split: These were split into 8000 training sets, 2000 development sets, and 2000 evaluation sets.
  • Parameters: The data was generated for various parameters, including:
  • Boundary radius R = 1
    • Number of Fourier coefficients for q (related to M) = 5
    • Number of Fourier coefficients for S (related to N) = 8
    • Different wavenumbers k (e.g., 5, 7, 10, 20).
  • Noise: The evaluation included datasets with varying levels of Gaussian noise (0.5% to 5%) to assess the model's robustness.

No new public benchmarks were introduced; the datasets were generated specifically for this study.

Key Insights and Ideas

  • DNNs as Regularizers for Ill-Posed Problems: The paper highlights how DNNs can effectively act as implicit regularizers for severely ill-posed inverse problems. By learning from a large dataset of input-output pairs, the DNN can find a stable mapping that traditional regularization methods might struggle with.
  • Radial Symmetry Simplification: The assumption of radial symmetry is crucial. It significantly reduces the dimensionality of the unknown parameter (q) and simplifies the forward operator T, making the inverse problem more amenable to DNN learning.
  • Trade-off in Network Complexity (M): While increasing the dimension M (number of basis functions for q) can improve resolution, it also makes the DNN harder to train due to increased nonlinearity and sensitivity. This suggests a delicate balance in choosing the network architecture and parameter representation.
  • Wavenumber Dependence: The performance of the DNN is shown to be dependent on the wavenumber k. Lower wavenumbers exhibit higher nonlinearity, while higher wavenumbers lead to oscillatory behavior and potential local minima issues. This indicates that the optimal DNN architecture or training strategy might need to be adapted for different frequency regimes.
  • Handling Noise: The study demonstrates that the trained DNN can handle noisy measurements reasonably well, with relative errors increasing gradually with noise levels, suggesting robustness.

Key Results

  • Successful Reconstruction: The proposed DNN model successfully reconstructs the refractive index q from phaseless scattering data in numerical experiments.
  • Relative Errors: For the radial symmetry case (X₅ with k=7), the relative errors ||P₅q - q̂||₂ / ||P₅q||₂ (where P₅ is the projection operator onto X₅ and q̂ is the recovered index) were generally low, ranging from approximately 0.0164 to 0.0574 for different radial profiles.
  • Noise Robustness: Figure 4 shows that relative errors increase with noise levels but without sudden jumps, indicating the model's ability to cope with noise. Figure 5 further illustrates good reconstructions even at moderate noise levels (SNRs around 23-33 dB).
  • Wavenumber Performance: The model performs better for intermediate wavenumbers (e.g., k=5, 7, 10) compared to very low or very high ones. Performance degrades for k < 5 due to nonlinearity and for k = 20 due to oscillatory behavior and data accumulation.
  • Coefficient Recovery: Table 2 shows that the DNN can recover coefficients c_m for 1 ≤ m ≤ 5 when the true q belongs to X₅, and it effectively ignores higher-order coefficients, acting as a projection.

Limitations and Future Work

  • Radial Symmetry Assumption: The current work relies heavily on the assumption of radial symmetry for the refractive index. This limits its direct applicability to non-radially symmetric objects.
  • Wavenumber Dependence: The performance varies significantly with the wavenumber k. The model's effectiveness at very low or very high frequencies needs further improvement.
  • Network Complexity and Training: For higher-dimensional representations of q (larger M), training the DNN becomes more challenging due to increased nonlinearity and sensitivity.
  • Future Work:
    • Developing new network architectures or training strategies for larger M to improve resolution.
  • Investigating methods to handle non-radially symmetric refractive indices.
  • Exploring techniques to improve performance across a wider range of wavenumbers.
  • Further analysis of the ill-posedness and the role of DNNs in overcoming it.

Summary updated

Abstract

We address the inverse medium scattering problem with phaseless data motivated by nondestructive testing for optical fibers. As the phase information of the data is unknown, this problem may be regarded as a standard phase retrieval problem that consists of identifying the phase from the amplitude of data and the structure of the related operator. This problem has been studied intensively due to its wide applications in physics and engineering. However, the uniqueness of the inverse problem with phaseless data is still open and the problem itself is severely ill-posed. In this work, we construct a model to approximate the solution operator in finite-dimensional spaces by a deep neural network assuming that the refractive index is radially symmetric. We are then able to recover the refractive index from the phaseless data. Numerical experiments are presented to illustrate the effectiveness of the proposed model.
BibTeX
@inproceedings{lim2021application,
  author = {Soojong Lim and Jaemin Shin},
  title = {{Application of a Deep Neural Network to Phase Retrieval in Inverse Medium Scattering Problems}},
  booktitle = {Comput},
  year = {2021},
  url = {https://openreview.net/forum?id=b03h2hT750},
}
Figure