Fast Optical System Identification by Numerical Interferometry

Sidharth Gupta, Rémi Gribonval, Laurent Daudet, Ivan DokmanicICASSP 2020 · OpenReview

TL;DR

This paper introduces a significantly faster method for "calibrating" optical systems, specifically those that scatter light, by numerically reconstructing their transmission matrix (TM). By cleverly designing input signals and leveraging distance geometry and FFTs, they reduce TM computation time from hours to minutes, enabling real-time imaging through scattering media.

What Problem Does This Paper Address?

Imaging through complex optical systems, especially those that scatter light (like fog or biological tissue), is challenging because the system's behavior, represented by a transmission matrix (TM), is often unknown. Furthermore, typical cameras only measure light intensity, not its phase, making the problem of determining the TM a difficult "phase retrieval" task. This phase retrieval problem, when dealing with scattering media, is computationally intensive, requiring hours or even days on powerful hardware for state-of-the-art methods. The authors aim to drastically reduce this calibration time while still achieving accurate TM reconstruction for imaging applications.

Methods

The core of their approach lies in two main innovations:

  1. Phase Retrieval via Distance Geometry (Multilateration): Instead of directly solving the quadratic phase retrieval problem, they reframe it as a "distance geometry" problem. By designing specific input "probe" signals and measuring the intensity of the scattered output, they can infer the squared distances between the output signals and known "anchor" signals in a complex plane. This is analogous to multilateration (like GPS), where knowing distances to multiple known points allows you to pinpoint your location. They show that this can be formulated as a small, efficient linear system for each row of the TM, allowing for rapid phase recovery.
  2. Fast Fourier Transform (FFT) for TM Estimation: Once the phases are recovered, the TM can be estimated by solving a linear system. To accelerate this, they design their probe signals to be circulant matrices. Circulant matrices have a special property: they can be diagonalized by the Discrete Fourier Transform (DFT), which is efficiently computed by the FFT. This allows the TM estimation to be performed with a numerical complexity that is almost linear in the size of the TM, a significant speedup over direct matrix inversion.

Datasets

Not discussed in the paper. The experiments use real optical hardware and custom-designed probe signals.

Key Insights and Ideas

  • Reframing Phase Retrieval: The most significant insight is transforming the computationally hard quadratic phase retrieval problem into a solvable distance geometry (multilateration) problem. This elegantly bypasses the need for iterative, time-consuming phase retrieval algorithms.
  • Leveraging Structure for Speed: The clever design of probe signals as circulant matrices is key to unlocking the power of FFTs for linear system inversion, drastically reducing computational complexity.
  • Intensity-Only Measurement Advantage: The method works with intensity-only measurements, which is crucial for practical applications where phase information is not directly accessible.
  • Trade-off Between Speed and Accuracy: The paper demonstrates that faster calibration doesn't necessarily mean lower accuracy. Their method achieves comparable or better results than slower methods.

Key Results

  • Massive Speedup: The proposed method reduces the time to compute a transmission matrix from hours (e.g., 3.26 hours for a 256x64 matrix using prVAMP) to just a few minutes (e.g., 6.15 minutes for a comparable matrix size on a CPU). The largest TM calibrated in their experiments (128x128) took 71.97 minutes, a task that would have taken significantly longer with previous methods.
  • Accurate Reconstruction: The numerically computed TMs were successfully used to reconstruct unseen images behind a scattering medium, with relative errors decreasing as the oversampling factor (M/N) increased.
  • Scalability: The computational time scales favorably with the oversampling factor (linearly) and the input dimension (logarithmically with FFT acceleration).

Limitations and Future Work

  • Anchor Positioning: While their method for initial anchor positioning is efficient, it relies on classical multidimensional scaling (MDS), which can still be computationally intensive for a very large number of anchors.
  • Binary Inputs: The experimental setup uses a DMD which produces binary inputs. Designing probe signals that remain binary after subtraction and can still be used for accurate phase retrieval required specific signal design strategies.
  • Future Work: The paper implies that further optimization of anchor design and signal generation could lead to even greater speedups. The successful demonstration on hardware suggests potential for real-time applications in areas like microscopy, non-invasive imaging, and optical computing.

Summary updated

Abstract

We propose a numerical interferometry method for identification of optical multiply-scattering systems when only intensity can be measured. Our method simplifies the calibration of optical transmission matrices from a quadratic to a linear inverse problem by first recovering the phase of the measurements. We show that by carefully designing the probing signals, measurement phase retrieval amounts to a distance geometry problem-a multilateration-in the complex plane. Since multilateration can be formulated as a small linear system which is the same for entire rows of the transmission matrix, the phases can be retrieved very efficiently. To speed up the subsequent estimation of transmission matrices, we design calibration signals so as to take advantage of the fast Fourier transform, achieving a numerical complexity almost linear in the number of transmission matrix entries. We run experiments on real optical hardware and use the numerically computed transmission matrix to recover an unseen image behind a scattering medium. Where the previous state-of-the-art method reports hours to compute the transmission matrix on a GPU, our method takes only a few minutes on a CPU.
BibTeX
@inproceedings{gupta2020fast,
  author = {Sidharth Gupta and Rémi Gribonval and Laurent Daudet and Ivan Dokmanic},
  title = {{Fast Optical System Identification by Numerical Interferometry}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=bEXQhr4yBZ},
}
Figure