Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays

Shashwath Bharadwaj, Ruangrawee Kitichotkul, Akshay Agarwal, Vivek K. GoyalCVPR 2025 · DOI

Introduction

Scaling single-photon detector arrays, such as Superconducting Nanowire Single-Photon Detectors (SNSPDs), is a critical challenge for advancing low-light applications like Lidar and biological imaging. While these sensors offer unparalleled sensitivity, they suffer from a significant "wiring bottleneck." In cryogenic environments, providing a dedicated readout wire for every pixel in a large array generates excessive heat, which can disrupt superconductivity.

To overcome this, researchers employ row-column multiplexing, reducing the required wiring for an $n \times n$ array from $n^2$ to $2n$. However, this efficiency comes at the cost of spatial ambiguity: when multiple photons strike the array simultaneously, the readout cannot uniquely determine their locations, leading to "ghost" artifacts in reconstructed images.

The Multiplexed Readout Challenge

In a multiplexed system, a photon hit at a specific pixel triggers both its corresponding row and column wires. If two or more photons arrive within the same acquisition frame, the resulting row-column trigger pattern can be ambiguous. For instance, hits at (1,1) and (2,2) produce the same readout as hits at (1,2) and (2,1).

Overview of the measurement process and a comparison of reconstruction methods. The proposed multiphoton estimator (bottom right) achieves low bias and low variance compared to conventional approaches.

Conventional reconstruction methods typically handle this ambiguity poorly. A Naive Estimator (NE) might assume photons hit every possible intersection of triggered rows and columns, creating significant "ghosting" and high bias. Alternatively, a Single-Photon Estimator (SPE) might simply discard any frame with more than one detection. While accurate, the SPE wastes a vast amount of data as light intensity increases, leading to high noise (variance).

Proposed Probabilistic Reconstruction

This paper introduces a Multiphoton Estimator (ME) that treats coincidence resolution as an inverse imaging problem. Instead of discarding ambiguous data, the ME uses a probabilistic framework to "unmix" these frames.

The algorithm first uses unambiguous single-photon frames to generate an initial scene estimate. It then calculates the likelihood of various photon combinations for each ambiguous frame and redistributes the counts to the most probable pixel locations. This approach allows the system to recover significant spatial information that was previously considered "trash." The researchers demonstrated this framework on a $32 \times 32$ array, successfully resolving up to four simultaneous photon hits.

Key Results and Performance

The proposed Multiphoton Estimator significantly outperforms conventional methods across several metrics. Qualitatively, it eliminates the horizontal and vertical "streaks" characteristic of naive reconstruction. Quantitatively, the ME provides a 3 to 4 dB boost in Peak Signal-to-Noise Ratio (PSNR) under optimal light conditions.

Mean Squared Error (MSE) vs. Mean Photons Per Frame. The proposed multiphoton estimators (2, 3, and 4-photon) shift the optimal operating point to higher light intensities compared to the single-photon and naive methods.

A critical insight is the shift in the system's "sweet spot." Every imaging system has an optimal incident light level; too little light results in shot noise, while too much leads to unresolvable coincidences. As shown in the performance curves, the ME shifts this optimal point to higher light intensities, allowing the sensor to operate effectively in much brighter conditions. Furthermore, the ME achieves the same Mean Squared Error (MSE) as other methods with approximately 4 times fewer readout frames, representing a substantial gain in acquisition speed.

Conclusion and Future Directions

By resolving the spatial ambiguities of multiplexed readouts, this probabilistic framework enables high-fidelity single-photon imaging with drastically reduced wiring requirements. Future work may involve incorporating spatial priors or deep learning to further enhance reconstruction quality, as well as "tiling" the sensor into smaller blocks to scale the computational model to megapixel resolutions. This advancement paves the way for large-scale, high-speed single-photon camera systems.

Summary updated

BibTeX
@inproceedings{bharadwaj2025image,
  author = {Shashwath Bharadwaj and Ruangrawee Kitichotkul and Akshay Agarwal and Vivek K. Goyal},
  title = {{Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.01065},
}
Figure