Maximum Likelihood Reconstruction for Multi-Look Digital Holography with Markov-Modeled Speckle Correlation

Xi Chen, Arian Maleki, Shirin JalaliarXiv 2026 · arXiv

TL;DR

Multi-look digital holography reduces speckle noise by combining multiple imaging shots, but real-world hardware limitations often introduce unwanted correlation between these shots. This paper introduces a first-order Markov process to explicitly model this inter-look correlation and proposes a scalable, deep-prior-regularized maximum likelihood reconstruction algorithm. The resulting framework recovers high-quality, speckle-free images even under strong hardware-induced correlation.

What Problem Does This Paper Address?

Coherent imaging systems, such as digital holography and holographic displays, suffer from "speckle noise"—a grainy, destructive interference pattern caused by random phase variations of coherent light waves. To mitigate this, practitioners use multi-look acquisition, which captures multiple measurements (looks) of the same object under varying optical conditions and combines them to average out the noise.

Standard reconstruction algorithms operate under a major assumption: that the speckle pattern in each look is completely independent. In practice, however, hardware constraints—such as the finite spatial resolution or phase quantization of spatial light modulators—limit measurement diversity. This introduces statistical correlation between successive looks. When conventional algorithms ignore this correlation, their reconstruction performance degrades significantly. This paper addresses the challenge of accurately reconstructing speckle-free images when multi-look measurements are statistically dependent.

Methods

The authors propose a novel statistical framework to model and exploit inter-look dependencies:

  1. Markov Speckle Modeling: The speckle field across successive looks is modeled as a first-order Markov process, where each look depends on the previous one via a correlation coefficient $\alpha \in [0, 1]$.
  2. Approximate MLE Formulation: Because additive sensor noise prevents the final measurements from forming a strict Markov chain, the authors derive a first-order Markov approximation of the joint distribution. This yields a constrained Maximum Likelihood Estimation (MLE) problem.
  3. Correlation Estimation: The authors develop an analytical method to estimate the unknown correlation coefficient $\alpha$ directly from the raw measurement data, removing the need for prior calibration.
  4. Scalable Optimization (PGD + DIP): To solve the complex, non-convex MLE optimization problem, the authors deploy Projected Gradient Descent (PGD). They use an untrained neural network as a Deep Image Prior (DIP) within the projection step, enforcing structural regularization without requiring a massive pre-trained dataset.
  5. Matrix-Free Computation: Computing the exact gradient of the likelihood function requires computationally prohibitive matrix inversions. The authors bypass this by using a nested Conjugate Gradient (CG) method for vector terms and a randomized Monte Carlo approximation for matrix diagonal terms, leveraging fast Fourier transforms (FFTs) to keep the algorithm scalable to high-resolution images.
[Multi-Look Measurements] ──> [Estimate Correlation (α)] ──> [PGD Optimization Loop]
                                                                     │
   ┌─────────────────────────────────────────────────────────────────┘
   ▼
[Compute Likelihood Gradient] ──> [DIP Projection (Untrained NN)] ──> [Speckle-Free Image]
  (Nested CG + Monte Carlo)

Datasets

The method was evaluated using simulated multi-look holographic measurements generated from standard imaging benchmarks:

  • Peppers ($256 \times 256$ resolution)
  • Barbara ($256 \times 256$ resolution)

The forward imaging model simulated realistic hardware constraints using both circular and annular aperture masks under varying noise levels. No new physical datasets were introduced.

Key Insights and Ideas

  • Embracing the Redundancy: Instead of trying to physically eliminate hardware correlation (which is often costly or impossible), mathematically modeling the correlation allows the algorithm to understand what information is redundant across looks.
  • The Power of Untrained Priors: Integrating DIP as the projection operator allows the framework to exploit deep structural priors natively, making it highly generalizable to different image types without training on specific datasets.
  • Overcoming the Inversion Bottleneck: The exact gradient of the Markov-modeled likelihood is mathematically daunting. Turning an intractable $O(n^3)$ matrix inversion problem into a matrix-free pipeline using nested CG and Monte Carlo sampling is the key algorithmic "aha moment" that makes this approach viable for practical deployment.

Key Results

  • Robustness to High Correlation: In extreme correlation regimes ($\alpha = 0.8$), the proposed method significantly outperformed baselines that assume independent looks, preserving sharp edges and fine details where the baseline failed.
  • Near-Ideal Performance: Across various aperture types and noise levels, the proposed method achieved Peak Signal-to-Noise Ratio (PSNR) values within 0.5 dB of the theoretical "upper bound" (the ideal, unachievable scenario where looks are perfectly independent).
  • Accurate Parameter Tracking: The data-driven estimator for $\alpha$ proved highly stable across 50 independent simulation runs (e.g., yielding an estimate of $0.7883 \pm 0.0013$ for a true value of $0.8$).

Limitations and Future Work

While the mathematical framework is robust, the authors note a few limitations and areas for future exploration:

  • Simulation-Only Validation: The method has currently only been validated on simulated data using a controlled forward model.
  • Real-World Model Mismatches: Experimental data from physical holographic setups may introduce unmodeled complexities, such as imperfect system calibration, non-Gaussian sensor noise, or higher-order (non-Markovian) correlation structures.
  • Future Directions: The primary next step is to adapt and validate the framework on real-world experimental digital holography datasets, extending the noise models to handle physical system discrepancies.

Summary updated

Abstract

Multi-look acquisition is a widely used strategy for reducing speckle noise in coherent imaging systems such as digital holography. By acquiring multiple measurements, speckle can be suppressed through averaging or joint reconstruction, typically under the assumption that speckle realizations across looks are statistically independent. In practice, however, hardware constraints limit measurement diversity, leading to inter-look correlation that degrades the performance of conventional methods. In this work, we study the reconstruction of speckle-free reflectivity from complex-valued multi-look measurements in the presence of correlated speckle. We model the inter-look dependence using a first-order Markov process and derive the corresponding likelihood under a first-order Markov approximation, resulting in a constrained maximum likelihood estimation problem. To solve this problem, we develop an efficient projected gradient descent framework that combines gradient-based updates with implicit regularization via deep image priors, and leverages Monte Carlo approximation and matrix-free operators for scalable computation. Simulation results demonstrate that the proposed approach remains robust under strong inter-look correlation, achieving performance close to the ideal independent-look scenario and consistently outperforming methods that ignore such dependencies. These results highlight the importance of explicitly modeling inter-look correlation and provide a practical framework for multi-look holographic reconstruction under realistic acquisition conditions. Our code is available at: https://github.com/Computational-Imaging-RU/MLE-Holography-Markov.
BibTeX
@misc{chen2026maximum,
  author = {Xi Chen and Arian Maleki and Shirin Jalali},
  title = {{Maximum Likelihood Reconstruction for Multi-Look Digital Holography with Markov-Modeled Speckle Correlation}},
  year = {2026},
  eprint = {2604.20154},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.20154},
}

Loading related content...

Figure