- The paper introduces a Poisson wavefront imaging method that accurately estimates phase in photon-starved conditions using SLM phase diversity with TV regularization.
- It models wave propagation using the angular spectrum method and employs an ADMM-based optimization, achieving up to 1.8× improvement in spatial resolution compared to classical techniques.
- Experimental results demonstrate reduced RMSE and enhanced Fisher information, confirming superior performance over traditional methods in low-photon scenarios.
Poisson Wavefront Imaging in Photon-Starved Scenarios
Motivation and Context
Accurate wavefront sensing in photon-starved regimes is central to multiple high-impact applications, including adaptive optics for astronomy, high-resolution microscopy, and semiconductor metrology. Traditional techniques, e.g., Shack-Hartmann wavefront sensors (SHWFS), are fundamentally constrained in low-light by read-noise, limited spatial resolution, and biased estimation under non-Gaussian statistics. Established inverse approaches, like Gerchberg-Saxton (GS) phase retrieval and variational methods, employ fidelity terms mismatched to the Poissonian statistics endemic to weak optical signals, leading to suboptimal performance.
Recent simulation advances suggest that incorporating correct photon-counting models (Poisson likelihoods) and advanced regularization (such as total variation, TV) could unlock much higher fidelity in wavefront imaging under severe photon constraints. However, experimental validation and systematic analysis of theoretical limits have lagged behind, especially in practical regimes featuring dynamic targets and SLM-enabled phase diversity, with implications for real-world imaging.
Methodology
The proposed Poisson Wavefront Imaging (PWI) method leverages optimization-based phase retrieval with explicit modeling of Poisson photon statistics and SLM-based phase diversity. The overall architecture consists of:
- Hardware configuration: The system exploits a spatial light modulator (SLM) to introduce deterministic and known phase diversity across a sequence of exposures. Each SLM modulation generates a distinct forward model, collectively increasing Fisher information in the photon-starved regime.
- Forward model: Wave propagation between major planes (target, SLM, and detector) is modeled using the angular spectrum method, allowing a precise, physically-consistent relationship between the phase object and measured intensities under various coded SLM patterns.
- Inverse algorithm: Phase retrieval is formalized as minimization of a composite loss: a Poisson negative log-likelihood fidelity term enforces consistency with measured photon statistics, while a TV prior imposes smoothness and edge preservation on the estimated phase. The resulting constrained optimization is solved via the ADMM, yielding efficient and modular updates for the field at each plane and the auxiliary TV variable.
The method is implemented both in simulation and in hardware, using an emCCD camera and a well-characterized SLM to measure photon-starved images of phase targets placed in both the image and pupil planes. Experimental photon counts and exposure times are rigorously quantified.
A critical innovation is the analytical calculation of Fisher information and the Cramér-Rao lower bound (CRLB) for SHWFS and PWI under the same photon budget. By evaluating the diagonal elements of the phase parameter CRLB, the study demonstrates that SLM-generated phase diversity fundamentally increases phase sensitivity: PWI reaches a lower theoretical error floor than SHWFS, especially for high-resolution imaging tasks. This is especially pronounced when the SLM patterns are designed for maximal information extraction rather than simple phase stepping.
PWI’s MAP estimator can, through use of the TV prior, further lower empirical mean squared error below the unbiased CRLB, as the regularizer encodes strong prior information about expected phase structure.
Experimental Results
Quantitative and qualitative experimental evaluation solidifies the advantages of the PWI framework:
- Comparison with GS and plain Poisson estimation: In reconstructions of a standard USAF phase target, PWI with the TV prior achieves reliable phase estimation down to less than 2 photons per pixel, where classical GS completely fails. The TV-regularized approach yields up to 1.6× reduction in RMSE compared to GS, and approximately 1.8× improvement in spatial resolution at a fixed photon budget.
- Generalization across system configurations: When the phase target is placed in the pupil plane, where per-pixel Poisson fidelity brings reduced marginal benefit due to Fourier-domain mixing, the TV prior predominates in performance gain. Both GS and PWI without TV perform similarly in this case, but PWI with TV still achieves a consistent 1.6× RMSE reduction.
- Phase diversity and Fisher information enhancement: The increase in Fisher information with SLM-coded phase diversity is confirmed experimentally, supporting the theoretical claim that physically-programmed diversity is a crucial driver of performance in photon-limited scenarios.
- Quantitative Correlation: Across all tested photon regimes and phase targets, PWI (especially Poisson+TV) maintains higher correlation to ground truth compared to GS, as measured by normalized inner product metrics.
Implications and Further Directions
PWI demonstrates that ADMM-based Poisson MAP estimators, together with physically-programmable phase diversity and TV priors, constitute a robust pipeline for low-photon wavefront estimation. Empirically, this enables faster measurement, reduced sample damage, and improved resolution in photon-limited microscopy, real-time astronomy, and semiconductor inspection.
The dependence of Poisson modeling efficacy on the optical system matrix (here, distinct between image and pupil planes) indicates that future advances should target adaptive prior and fidelity function selection based on precise task configuration. Incorporating learned deep priors (conditional on target domain data), sophisticated noise models (e.g., Poisson-Gaussian mixtures), and advanced denoisers (BM3D or neural) within the PWI framework could further lower attainable error, provided the statistical mismatch is properly controlled.
Extensions to broadband or partially coherent imaging would require explicit handling of spectral variation and speckle decorrelation---possible through multi-wavelength or spectral slicing strategies. The rapid algorithmic convergence and efficiency of the ADMM approach positions PWI as a practical method for large-format, real-time or high-throughput systems.
Conclusion
Poisson Wavefront Imaging, based on SLM-driven phase diversity and Poisson-MAP inference with TV regularization, approaches the theoretical performance boundary for low-photon phase retrieval. It outperforms classical and even modern simulation-based techniques in both accuracy and speed when validated experimentally under severe photon constraints. The method’s efficacy is theoretically underpinned by Fisher information analysis and robust to varying optical system layouts, offering a superior route for next-generation photon-limited imaging systems in science and engineering (2512.12401).