Deep Variational Inverse Scattering

AmirEhsan Khorashadizadeh, Ali Aghababaei Harandi, Tin Vlasic, Hieu Nguyen, Ivan DokmanicarXiv 2022 · OpenReview

TL;DR

This paper introduces U-Flow, a novel deep learning framework that combines the strengths of U-Nets for high-quality image reconstruction with conditional normalizing flows for robust uncertainty quantification in inverse scattering problems. U-Flow generates physically meaningful uncertainty estimates and high-quality posterior samples, outperforming existing methods in challenging ill-posed scenarios.

What Problem Does This Paper Address?

Inverse medium scattering aims to reconstruct the properties of an unknown medium (like its wave speed) from scattered wave measurements. This is crucial for applications ranging from medical imaging to geophysics. However, these problems are often ill-posed, meaning that small errors in measurements or noise can lead to wildly inaccurate or ambiguous reconstructions. Traditional methods and even current deep learning approaches typically provide only a single "best guess" reconstruction, failing to capture the inherent uncertainty. This lack of uncertainty information is a significant drawback for safety-critical applications. While some methods attempt to quantify uncertainty, they often produce low-quality samples or require computationally expensive training.

Methods

The core innovation of U-Flow lies in its hybrid architecture. It leverages a U-Net, a well-established convolutional neural network architecture known for its excellent performance in image-to-image tasks, to produce high-quality point estimates of the medium. Crucially, U-Flow then uses conditional normalizing flows (specifically, a conditional version of Glow) to model the posterior distribution of the U-Net's internal, lower-dimensional feature representations.

Here's a breakdown:

  • U-Net Encoder-Decoder: The U-Net takes the backprojected measurements (a form of pre-processed data) as input and produces multi-scale feature maps. The decoder then reconstructs a single, high-quality estimate of the medium.
  • Conditional Normalizing Flows: Instead of directly modeling the high-dimensional posterior of the medium itself, U-Flow models the posterior of the U-Net's coarsest-scale latent features. This is a key insight: by modeling a lower-dimensional latent space, the normalizing flow can learn a more accurate and stable distribution. The flow is conditional on the input measurements.
  • Amortized Inference: To efficiently train the model, U-Flow employs amortized inference. This means the model learns to infer the posterior for any given measurement in a single forward pass, rather than requiring a separate optimization step for each new measurement. This is achieved by training the model to minimize the KL divergence averaged over the entire dataset.
  • Posterior Sampling: Once trained, U-Flow can generate multiple posterior samples by sampling from the learned latent distribution and passing these samples through the U-Net decoder. These samples collectively represent the uncertainty in the reconstruction.

Datasets

The paper describes generating synthetic datasets for training and evaluation. These datasets consist of 4000 samples, each containing a set of random ellipsoidal scatterers within a 128x128 grid. The wave speed is varied, with a background speed of 1540 m/s and contrasts up to four times the background. Measurements are simulated using the time-harmonic wave equation and corrupted with Gaussian noise (30 dB SNR is mentioned for some examples). The paper uses both limited-view and full-view measurement configurations. No specific public datasets are mentioned as being used.

Key Insights and Ideas

  • Latent Space Modeling for Flows: A significant insight is that modeling the posterior distribution in a lower-dimensional latent space (like the U-Net's internal features) is more effective for normalizing flows than directly modeling the high-dimensional image space. This latent representation acts as a powerful regularizer for ill-posed problems.
  • Combining Strengths: U-Flow successfully marries the high-fidelity reconstruction capabilities of U-Nets with the probabilistic modeling power of normalizing flows, addressing the dichotomy between accurate point estimates and meaningful uncertainty quantification.
  • Physically Meaningful Uncertainty: The generated uncertainty maps are shown to be physically meaningful, highlighting regions where the model is less confident, particularly in limited-view scenarios where information is missing.
  • Computational Efficiency: The amortized inference approach makes U-Flow significantly faster than traditional iterative methods for generating reconstructions and uncertainty estimates.

Key Results

  • Posterior Sample Quality: U-Flow significantly outperforms recent conditional injective flows (like C-Trumpet) in terms of posterior sample quality.
  • Point Estimation: U-Flow achieves MMSE (Minimum Mean Square Error) estimates that are comparable in quality to a standard U-Net, while also providing uncertainty information.
  • Uncertainty Quantification (UQ): U-Flow generates physically meaningful uncertainty estimates, especially in challenging limited-view scenarios, where it correctly assigns higher uncertainty to regions with less data.
  • Speed: Reconstructions are described as "very fast—orders of magnitude faster than with the standard iterative methods."
  • Quantitative Comparison: Table I shows U-Flow achieving an SNR of 9.7 for MMSE estimates in the limited-view setup, comparable to the U-Net's 11.3, and significantly better than C-Trumpet's 5.2. For the full-view setup, U-Flow achieves 11.3, again comparable to the U-Net.

Limitations and Future Work

  • Limitations: While U-Flow performs well, the paper notes that the quality of reconstructions is still inferior to highly successful image-to-image regression models like the U-Net when only point estimation is considered. The paper also mentions that conditional normalizing flows can require significant memory and be slow to train, although U-Flow mitigates some of these issues by operating in a latent space.
  • Future Work: The authors suggest that it would be interesting to explore how the proposed model performs on other inverse problems beyond inverse scattering. They also imply that further architectural improvements or training strategies could potentially close the gap in point estimation quality with standard U-Nets.

Summary updated

Abstract

Inverse medium scattering solvers generally reconstruct a single solution without an associated measure of uncertainty. This is true both for the classical iterative solvers and for the emerging deep learning methods. But ill-posedness and noise can make this single estimate inaccurate or misleading. While deep networks such as conditional normalizing flows can be used to sample posteriors in inverse problems, they often yield low-quality samples and uncertainty estimates. In this paper, we propose U-Flow, a Bayesian U-Net based on conditional normalizing flows, which generates high-quality posterior samples and estimates physically-meaningful uncertainty. We show that the proposed model significantly outperforms the recent normalizing flows in terms of posterior sample quality while having comparable performance with the U-Net in point estimation.
BibTeX
@misc{khorashadizadeh2022deep,
  author = {AmirEhsan Khorashadizadeh and Ali Aghababaei Harandi and Tin Vlasic and Hieu Nguyen and Ivan Dokmanic},
  title = {{Deep Variational Inverse Scattering}},
  year = {2022},
  eprint = {2212.04309},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2212.04309},
}
Figure