Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Wavefront Randomization Improves Deconvolution

Amit Kohli, Anastasios N. Angelopoulos, Laura Waller

TL;DR

The paper tackles blur caused by optical aberrations that introduce zeros in the transfer function, hindering deconvolution. It proposes wavefront randomization by inserting a random phase mask in the pupil plane to induce an aberration-invariant MTF, and analyzes two mask models: Uniform and Bernoulli. For the uniform mask, the study proves $H_n \stackrel{d}{=} \frac{1}{\lfloor N/2 \rfloor} \left| \sum_{j=n}^{N-1} e^{i(\phi_j - \phi_{j-n} + W_j - W_{j-n})} \right|$, showing the transfer function becomes independent of the aberrations and concentrates around its mean. The binary mask yields aberration-invariant second-moment behavior and a lower bound on the mean MTF, suggesting robust performance; simulations indicate masked systems improve deconvolution quality across aberration types and noise levels, offering a practical path to more reliable imaging with simple hardware.

Abstract

The performance of an imaging system is limited by optical aberrations, which cause blurriness in the resulting image. Digital correction techniques, such as deconvolution, have limited ability to correct the blur, since some spatial frequencies in the scene are not measured adequately (i.e., 'zeros' of the system transfer function). We prove that the addition of a random mask to an imaging system removes its dependence on aberrations, reducing the likelihood of zeros in the transfer function and consequently decreasing the sensitivity to noise during deconvolution. In simulation, we show that this strategy improves image quality over a range of aberration types, aberration strengths, and signal-to-noise ratios.

Wavefront Randomization Improves Deconvolution

TL;DR

The paper tackles blur caused by optical aberrations that introduce zeros in the transfer function, hindering deconvolution. It proposes wavefront randomization by inserting a random phase mask in the pupil plane to induce an aberration-invariant MTF, and analyzes two mask models: Uniform and Bernoulli. For the uniform mask, the study proves , showing the transfer function becomes independent of the aberrations and concentrates around its mean. The binary mask yields aberration-invariant second-moment behavior and a lower bound on the mean MTF, suggesting robust performance; simulations indicate masked systems improve deconvolution quality across aberration types and noise levels, offering a practical path to more reliable imaging with simple hardware.

Abstract

The performance of an imaging system is limited by optical aberrations, which cause blurriness in the resulting image. Digital correction techniques, such as deconvolution, have limited ability to correct the blur, since some spatial frequencies in the scene are not measured adequately (i.e., 'zeros' of the system transfer function). We prove that the addition of a random mask to an imaging system removes its dependence on aberrations, reducing the likelihood of zeros in the transfer function and consequently decreasing the sensitivity to noise during deconvolution. In simulation, we show that this strategy improves image quality over a range of aberration types, aberration strengths, and signal-to-noise ratios.
Paper Structure (15 sections, 8 theorems, 42 equations, 4 figures)

This paper contains 15 sections, 8 theorems, 42 equations, 4 figures.

Table of Contents

  1. INTRODUCTION
  2. Background
  3. Theory
  4. Random Masks
  5. Uniform mask
  6. Binary mask
  7. Imaging Simulations
  8. Discussion
  9. Binary mask
  10. Proofs of Main Results
  11. (1) $\mathbf{j \neq k-n}$ and $\mathbf{k \neq j-n}$
  12. (2) $\mathbf{j = k-n}$ or $\mathbf{k = j-n}$
  13. (1) $\mathbf{(j',k') = (k,j)}$
  14. (2) $\mathbf{j=k-n}$ and not (1)
  15. Technical lemmas

Key Result

Theorem 1

Consider a masked pupil function $\tilde{P}_n$ as in eq:pupil-masked with arbitrary aberrations $\phi_n$ and $W_n \overset{i.i.d.}{\sim} \mathrm{Unif(0, 2\pi)}$. Then, and where $C(N,n) = \lfloor N/2 \rfloor - n$ and the $\Delta_n : \mathbb{R}^N \to \mathbb{R}^{C(N,n)}$ function computes the vector $\Delta_n(w) = (w_n - w_0, \ldots, w_j - w_{j-n}, \ldots, w_{N-1} - w_{N-n-1})$.

Figures (4)

  • Figure 1: Simulation of a spherically aberrated imaging system with and without wavefront randomization. With no randomization (top row) the system has an MTF with severe nulls and a large blob-like point spread function (PSF). The image is a blurry, noisy version of the scene and the deconvolved image has noise-induced patterned artifacts. With a random mask (bottom row), the wavefront is randomized, causing the modulation transfer function (MTF) to become flatter, with no nulls. The corresponding point spread function (PSF) is a speckle pattern with small features. The image is random, but the deconvolved image is much closer to the ground truth. Noise is white, additive Gaussian.
  • Figure 2: Simulation of MTFs with and without a uniform random mask. Each row represents a different aberration type, and each column represents a different aberration strength. Within each individual plot is the MTF of the system with no mask (red), the empirical distribution of MTFs from many draws of a uniform mask (green), and the average MTF of those draws (blue). As expected by Theoreom \ref{['thm:uniform-mtf']}, the MTF distribution and average from uniform masks do not change with aberration type or strength, whereas the MTF without a mask does so drastically. Also note how the MTF distribution is concentrated around the average, signifying that the MTF is reliably null-free.
  • Figure 3: Simulation of image reconstructions and their SSIM scores for different noise levels and aberration strengths/types. a) Deconvolutions with (bottom) and without (top) a uniform random mask for increasing noise power. The masked case degrades more slowly. b) A plot of SSIM scores on the deconvolution from a) against SNR; the masked deconvolution degrades more gradually. c) Deconvolutions with and without a mask for increasing levels of spherical aberration and d) astigmatism. Deconvolution with the mask is less sensitive to noise for large aberrations. SSIM scores are shown on each image reconstruction.
  • Figure 4: Simulation of MTFs with and without a binary random mask. Each row represents a different aberration type, and each column represents a different aberration strength. Within each individual plot is the MTF of the system with no mask (red), the empirical distribution of MTFs from many draws of a binary mask, and the average MTF of those draws (blue). The MTF distribution and average from binary masks do change with aberration type and strength, but very little. Moreover, the binary mask MTF distribution is concentrated around the average, signifying that the MTF is reliably null-free.

Theorems & Definitions (15)

  • Theorem 1: Aberration invariance: uniform mask
  • Theorem 2: Approximate aberration invariance: binary mask
  • Theorem 3
  • Corollary 1: Aberration invariance: squared MTF of binary mask
  • proof : Proof of Theorem \ref{['thm:uniform-mtf']}
  • proof : Proof of Theorem \ref{['thm:binary-mtf']}
  • proof : Proof of Theorem \ref{['thm:binary-squared-mtf']}
  • Lemma 1
  • proof
  • Lemma 2: Translation invariance of uniform phasors.
  • ...and 5 more