Estimation-theoretic analysis of lensless imaging

Leyla Kabuli, Nalini Singh, Laura Waller

Abstract

We analyze lensless imaging systems with estimation-theoretic techniques based on Fisher information. Our analysis evaluates multiple optical encoder designs on objects with varying sparsity, in the context of both Gaussian and Poisson noise models. Our simulations verify that lensless imaging system performance is object-dependent and highlight tradeoffs between encoder multiplexing and object sparsity, showing quantitatively that sparse objects tolerate higher levels of multiplexing than dense objects. Insights from our analysis promise to inform and improve optical encoder designs for lensless imaging.

Introduction

Lensless imaging represents a paradigm shift in optical design, replacing bulky refractive lenses with thin "optical encoders" such as phase masks or diffusers. These systems rely on "computational decoders"—algorithms or neural networks—to reconstruct the final image from the encoded measurements. However, evaluating these systems is challenging because hardware performance (the mask) is often conflated with software performance (the reconstruction algorithm). If a reconstruction is poor, it is unclear whether the mask failed to capture sufficient information or if the algorithm was simply suboptimal.

Methodology: Decoupling Hardware and Software

To address this, researchers at UC Berkeley have treated lensless imaging as a parameter estimation problem. By calculating the Fisher information and the associated Cramér-Rao Bound (CRB), they determine the theoretical minimum variance any unbiased reconstruction algorithm could achieve for a given hardware design. This framework allows for the analysis of the optical encoder in isolation.

Overview of the lensless imaging pipeline and the various point spread functions (PSFs) analyzed.

The study models the system as a convolution of the object with a Point Spread Function (PSF). Seven different PSFs were tested, representing increasing levels of "multiplexing"—the degree to which light from a single object point is spread across the sensor. These range from a simple single lens to complex diffusers. The analysis considers two primary noise sources: Gaussian "read noise" from sensor electronics and Poisson "shot noise" inherent to the physics of light.

Results: The Impact of Object Sparsity and Noise

A key finding of the study is that the theoretical performance limit depends heavily on the noise model and the scene content. In the presence of Gaussian noise, the CRB depends solely on the mask design. However, under Poisson noise, the performance becomes object-dependent.

Cramér-Rao Bound (CRB) analysis for dense and sparse objects under Poisson noise, showing how sparsity mitigates the performance degradation caused by high multiplexing.

As shown in Figure 3, sparse objects (such as fluorescent beads) are remarkably resilient to high levels of multiplexing. While dense scenes show a significant increase in variance (worse performance) as multiplexing increases, sparse objects maintain a low and nearly constant CRB across multiple lenslet designs. Performance only degrades significantly for sparse objects when moving to the extremely high multiplexing of a diffuser. Additionally, the CRB is not uniform across the image; for multi-lenslet systems, variance is higher in the center where multiple "copies" of the object overlap.

Discussion and Future Work

The "decoder-independent" benchmark provided by the CRB allows for a fundamental comparison of optical masks. The results highlight a critical trade-off: while multiplexing is necessary for lensless imaging to work, it generally increases estimation variance, especially for dense objects.

One limitation is that the CRB applies to unbiased estimators, whereas many modern AI-based decoders use priors to improve image quality. Future work will explore the Biased CRB and the Van Trees inequality to better model these advanced decoders. Ultimately, these metrics could serve as loss functions for the automated design of task-specific optical encoders for biological samples.

Summary updated

Abstract

We analyze lensless imaging systems with estimation-theoretic techniques based on Fisher information. Our analysis evaluates multiple optical encoder designs on objects with varying sparsity, in the context of both Gaussian and Poisson noise models. Our simulations verify that lensless imaging system performance is object-dependent and highlight tradeoffs between encoder multiplexing and object sparsity, showing quantitatively that sparse objects tolerate higher levels of multiplexing than dense objects. Insights from our analysis promise to inform and improve optical encoder designs for lensless imaging.
BibTeX
@misc{kabuli2025estimation,
  author = {Leyla Kabuli and Nalini Singh and Laura Waller},
  title = {{Estimation-theoretic analysis of lensless imaging}},
  year = {2025},
  eprint = {2501.14727},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.14727},
  howpublished = {arXiv},
}
Figure