Noise-disentangled single-pixel imaging under photon-limited conditions

Abstract

Single pixel imaging (SPI) is a well-established computational imaging modality that renders as a practical and innovative tool for sensing across almost the entire spectrum range. Given the specific measurement mode (e.g., bucket in junction with differential), SPI arguably allows for promising signal-to-noise ratio that underpins an economical alternative to high-sensitivity pixelated detectors. However, this common interpretation could be totally collapsed due to the inevitable time-varying interference, making the image fidelity fall far below those using classical detectors. In this work, we present a novel learning-based SPI protocol that maintains outstanding robustness to noise under extremely low-light condition that leads to 10 photons per pixel or even less. Specifically, a noise disentanglement paradigm is implemented across the data and image domains in an unsupervised framework. In experiments, two photon-limited scenarios that have been typically found in fluorescence imaging were investigated. The results validated the compelling superiorities of our strategy over other conventional SPI algorithms. This work substantially improves the reliability and validity of SPI with downstream applications in the field of biomedicine and other scenarios that suffer from limited number of photons.