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

1. Problem Statement

Context: In coherent imaging systems like digital holography, speckle noise (granular intensity fluctuations caused by random phase interference) severely degrades image quality. A standard mitigation strategy is multi-look acquisition, where multiple measurements are taken and averaged or jointly reconstructed.
The Challenge: Conventional multi-look methods assume that speckle patterns across different "looks" are statistically independent. However, in practice, hardware limitations (finite spatial resolution, phase quantization, imperfect diffusers) prevent full independence, leading to inter-look correlation.
Consequence: When correlation exists but is ignored, standard Maximum Likelihood Estimation (MLE) methods treat dependent data as independent, resulting in suboptimal reconstruction performance and residual speckle noise.
Goal: The authors aim to reconstruct speckle-free reflectivity from complex-valued multi-look measurements by explicitly modeling and leveraging the statistical dependence between looks.

2. Methodology

A. Stochastic Modeling (Markov Process)

The authors propose modeling the speckle field across $L$ looks as a first-order Markov process.

• Model: Let $g_\ell$ be the complex speckle field for look $\ell$.
  • $g_1 \sim \mathcal{CN}(0, X)$, where $X = \text{diag}(x)$ and $x$ is the target reflectivity.
  • For $\ell > 1$: $g_\ell = \alpha g_{\ell-1} + \sqrt{1-\alpha^2} u_\ell$, where $u_\ell \sim \mathcal{CN}(0, X)$ and $\alpha \in [0, 1]$ is the correlation coefficient.
• Measurement Model: The observed measurement is $y_\ell = A g_\ell + z_\ell$, where $A$ is the sensing matrix (Fresnel diffraction model) and $z_\ell$ is additive Gaussian noise.
• Approximation: Because the additive noise breaks the strict Markov property of the measurements $y_\ell$, the authors derive a first-order Markov approximation for the joint likelihood $p(y_1, \dots, y_L)$. This approximates the joint distribution as $p(y_1) \prod_{\ell=2}^L p(y_\ell | y_{\ell-1})$.

B. Constrained Maximum Likelihood Estimation (MLE)

The reconstruction is formulated as a constrained optimization problem:
$$\hat{x} = \arg \min_{x \in \mathcal{X}} f_L(x)$$
Where $f_L(x)$ is the negative log-likelihood derived from the Markov approximation. The objective function includes:

1. A term for the first look: $\log \det S + y_1^H S^{-1} y_1$.
2. Conditional terms for subsequent looks involving the conditional mean and covariance, where $S = AXA^H + \sigma_z^2 I$ and $M = S - \alpha^2 B S^{-1} B$ (with $B=AXA^H$).
3. Regularization: The set $\mathcal{X}$ enforces image priors. The authors use Deep Image Priors (DIP), where the projection step $\Pi_\mathcal{X}$ is implemented by optimizing a neural network to fit the current iterate, effectively acting as an implicit regularizer.

C. Efficient Algorithm: Projected Gradient Descent (PGD)

Solving the MLE problem is computationally challenging due to the need to invert large matrices ($S$ and $M$) and compute their determinants. The authors propose a scalable Projected Gradient Descent framework:

1. Gradient Computation: The gradient of the non-convex objective is derived analytically.
2. Matrix-Free Operations:
   • Conjugate Gradient (CG): Instead of explicitly forming and inverting matrices, CG is used to solve linear systems (e.g., $S^{-1}y$) by leveraging the Fourier structure of the sensing matrix $A$ (using FFTs).
   • Monte Carlo Approximation: To compute the diagonal terms required for the gradient (e.g., $\text{diag}(A^H S^{-1} A)$), the authors use a randomized linear algebra approach. They generate random probe vectors, solve linear systems via CG, and average the results to estimate the diagonal entries unbiasedly.
3. Correlation Estimation: The correlation coefficient $\alpha$ is unknown in practice. The authors derive a closed-form estimator based on the sample covariance of the measurements ($E[y_{\ell-1}^H y_\ell]$) and the trace of the signal covariance.

3. Key Contributions

1. Markov Modeling of Speckle: First to explicitly model inter-look speckle correlation in digital holography using a first-order Markov process, moving beyond the independence assumption.
2. Novel Likelihood Formulation: Derivation of a constrained MLE objective function that accounts for correlated measurements, leading to a more accurate statistical model.
3. Scalable Optimization Algorithm: Development of a matrix-free PGD algorithm that combines:
   • Conjugate Gradient solvers for matrix inversion.
   • Monte Carlo methods for diagonal matrix approximation.
   • Deep Image Priors for regularization.
   • This allows the method to scale to high-resolution images without storing massive matrices.
4. Robustness to Hardware Limits: The method remains effective even when hardware constraints prevent full decorrelation of speckle patterns.

4. Results

The authors evaluated the method using simulations with the "Peppers" and "Barbara" images under various conditions (circular/annular apertures, noise levels, and correlation coefficients $\alpha$).

• Performance vs. Correlation:
  • As the correlation coefficient $\alpha$ increases (from 0.0 to 0.8), the performance of the baseline method (assuming independence) degrades significantly.
  • The proposed method maintains high performance, staying within 0.5 dB of the ideal independent-look upper bound even at high correlation ($\alpha=0.8$).
  • In terms of PSNR and SSIM, the proposed method consistently outperforms the baseline, particularly in high-correlation regimes.
• Visual Quality: Reconstructed images using the proposed loss function show significantly reduced speckle noise and better structural preservation compared to the baseline when $\alpha > 0$.
• Efficiency: The algorithm converges in 50–100 PGD iterations. Each iteration takes approximately 20–40 seconds for a $256 \times 256$ image, demonstrating practical feasibility for high-resolution imaging.
• Parameter Estimation: The proposed estimator for $\alpha$ is shown to be accurate (mean error < 0.02) across various noise levels and aperture settings.

5. Significance

This work addresses a critical gap between theoretical multi-look imaging strategies and practical hardware limitations. By acknowledging and modeling the inevitable correlation between measurements, the authors provide a framework that:

• Improves Image Quality: Achieves near-ideal denoising performance even when hardware cannot generate perfectly independent speckle patterns.
• Enables Realistic Deployment: Provides a practical solution for digital holography systems where perfect decorrelation is impossible due to physical constraints.
• Computational Efficiency: Demonstrates that complex statistical models involving matrix inversions can be solved efficiently for large-scale imaging problems using modern optimization techniques (CG, Monte Carlo, Deep Priors).

The code is open-source, facilitating further research into correlated multi-look imaging and holographic reconstruction.