💻 computer science

Monte Carlo Maximum Likelihood Reconstruction for Digital Holography with Speckle

This paper proposes a scalable Monte Carlo-based maximum likelihood reconstruction framework (PGD-MC) that utilizes randomized linear algebra to enable computationally efficient, physically accurate digital holography reconstruction with finite aperture modeling and speckle mitigation.

Original authors: Xi Chen, Arian Maleki, Shirin Jalali

Published 2026-02-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Xi Chen, Arian Maleki, Shirin Jalali

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a high-definition photo of a beautiful, shimmering lake at sunset. However, there are two major problems:

The "Glitter" Problem (Speckle Noise): Because the water is moving and reflecting light in a complex way, the image isn't smooth. Instead, it’s covered in tiny, distracting "glitter" or graininess. In science, this is called speckle noise, and it happens in many high-tech imaging systems like medical ultrasounds or radar.
The "Dirty Lens" Problem (Aperture): Your camera lens isn't perfect. It might have a circular shape or even a hole in the middle (like a donut). This "aperture" limits how much light gets in and blurs the details, making the math required to "fix" the image incredibly messy.

The Old Way: The "Shortcut" Approach

For years, scientists have tried to fix this by taking "shortcuts." To deal with the math, they would pretend the lens was perfect or pretend the "glitter" wasn't there.

Think of it like trying to clean a window by just wiping it with a dry cloth. It works okay for a little bit of dust, but if the window is covered in thick mud (complex noise) and the window frame is oddly shaped (the aperture), the shortcuts fail. You end up with a blurry, "over-smoothed" image that looks more like a watercolor painting than a real photo.

The New Way: The "Smart Detective" (PGD-MC)

The authors of this paper have created a new method called PGD-MC. Instead of taking shortcuts, they decided to face the math head-on.

The problem is that the "correct" math is so massive and heavy that even the world's fastest supercomputers would choke on it. It’s like trying to solve a billion-piece jigsaw puzzle all at once.

Here is their clever trick:

The Monte Carlo Method (The "Statistical Scout"): Instead of trying to calculate every single tiny piece of the puzzle at the same time, they use "scouts." They send out a few random samples (the Monte Carlo part) to get a "good enough" idea of what the whole picture looks like. It’s like looking at a few random pixels to guess the color of the entire sky—it’s much faster than checking every single atom of air!
The Conjugate Gradient (The "Efficient Navigator"): To make sure those scouts don't lead them astray, they use a very smart navigation system. Instead of wandering aimlessly, this system calculates the most direct path to the correct answer, saving massive amounts of time and energy.

Why does this matter?

By using these two tricks together, the researchers created a "Smart Detective" that can:

See through the glitter: It removes the speckle noise without making the image look blurry or "fake."
Respect the lens: It works perfectly even if the camera lens has a weird shape (like an annular/donut shape).
Scale up: It can handle much higher-resolution images (like 512x512 pixels) without crashing the computer.
Use "Smart Filters": It can plug in existing AI "denoisers" (like the ones in your smartphone) to make the final image look even cleaner.

In short: They found a way to solve a "mathematically impossible" problem by using smart sampling and efficient navigation, allowing us to see much clearer, sharper images in technologies like medical imaging and digital holography.

Technical Summary: Monte Carlo Maximum Likelihood Reconstruction for Digital Holography with Speckle

1. Problem Statement

Coherent imaging systems, such as digital holography, suffer from speckle noise, which is statistically modeled as multiplicative noise. Mathematically, the acquisition is modeled as $y = Ag + z$, where $g$ is the complex reflectance image containing speckle (modeled as $g \sim \mathcal{CN}(0, \text{diag}(x))$ ) and $z$ is additive sensor noise.

The principled way to reconstruct the speckle-free reflectivity $x$ is through Maximum Likelihood Estimation (MLE). However, the MLE objective function involves the log-determinant and the inverse of the covariance matrix $\Sigma(x) = AX^AH + \sigma_z^2 I$ . For high-resolution images, the dimensionality of these matrices makes direct inversion and determinant calculation computationally prohibitive. Consequently, existing methods often rely on sub-optimal approximations (e.g., assuming $A^HA \approx I$ ) or computationally heavy iterative matrix inversion methods (e.g., Newton-Schulz), which limit scalability and accuracy when modeling physically accurate apertures (like circular or annular masks).

2. Methodology: PGD-MC

The authors propose a new framework termed Projected Gradient Descent with Monte Carlo estimation (PGD-MC). The core innovation is a randomized linear algebra approach that enables scalable MLE optimization without explicit matrix inversions.

Key Algorithmic Components:

Projected Gradient Descent (PGD): The problem is formulated as a constrained MLE: $\hat{x} = \arg \min_{x \in \mathcal{X}} f_L(x)$ , where $\mathcal{X}$ is a constraint set defined by image priors (denoisers). The algorithm alternates between a gradient descent step (data fidelity) and a projection step (incorporating priors).
Monte Carlo (MC) Trace Estimation: To avoid the $O(n^3)$ cost of computing the diagonal of $A^H\Sigma(x)^{-1}A$ (required for the gradient), the authors use a randomized approach. By generating $K$ random vectors $v \sim \mathcal{N}(0, I)$ and solving for $u$ such that $\Sigma(x)u \approx Av$ , they obtain an unbiased estimator of the diagonal entries via the identity $E[A^Hu \odot v]$ .
Conjugate Gradient (CG) Method: Instead of inverting $\Sigma(x)$ , the algorithm uses the CG method to solve the linear systems required for both the MC estimation and the likelihood gradient.
Exploiting Fourier Structure: Since the sensing matrix $A$ in digital holography is a Fourier-domain mask, the authors implement all operations using Fast Fourier Transforms (FFTs). This allows the algorithm to operate in the 2D spatial domain, avoiding explicit vectorization and massive matrix-vector multiplications.

3. Key Contributions

Scalable MLE Framework: Developed a method to solve the exact MLE objective for coherent imaging without the simplifying assumptions (like $A^HA \approx I$ ) used in prior works.
Randomized Matrix-Free Gradient: Introduced a combination of Monte Carlo trace estimation and CG that bypasses the need to form or invert high-dimensional covariance matrices.
Plug-and-Play (PnP) Integration: The framework is highly flexible, allowing the integration of various regularizers, including classical denoisers (BM3D), deep learning denoisers (DnCNN), and untrained neural networks (Deep Decoder).
High-Resolution Scalability: Demonstrated that the method scales effectively to high-resolution settings (up to $512 \times 512$ ) by leveraging FFT-based computations.

4. Results

Superior Accuracy: Extensive experiments show that PGD-MC consistently outperforms the state-of-the-art CPnP-EM method in terms of PSNR and SSIM across various aperture models (circular and annular), noise levels, and numbers of looks ( $L$ ).
Robustness: Unlike prior methods that become unstable at low noise levels or with restrictive apertures (where the $A^HA \approx I$ approximation fails), PGD-MC remains stable and accurate.
Efficiency: The per-iteration runtime of PGD-MC is lower than CPnP-EM, and this computational advantage increases as the image resolution grows.
Prior Performance: The method effectively leverages different priors; specifically, the Deep Decoder provided the highest visual quality and structural preservation.

5. Significance

This work bridges the gap between statistical rigor and computational feasibility in coherent imaging. By providing a way to implement "true" Maximum Likelihood reconstruction for high-resolution digital holography, it allows researchers to use physically accurate aperture models without the prohibitive cost previously associated with them. This is particularly significant for practical optical systems where central obscurations (annular apertures) and finite apertures are the norm rather than the exception.