Photon-Limited Imaging with Quanta Image Sensors Via an Unsupervised Learning Framework

Haosen Liu, Yunping Zhang, Edmund Y. Lam

Abstract: Due to their single-photon sensitivity, quanta image sensors (QIS) are designed to complement traditional image sensors for a wide range of applications in photon-limited imaging conditions. However, the binary nature of QIS data poses compatibility challenges with existing image processing tools, necessitating the development of specialized reconstruction algorithms. While training a deep neural network with paired QIS recordings and corresponding ground truth data in a supervised manner offers superior performance compared to closed-form optimization-based solutions, collecting such a dataset can be laborious or impractical in certain sce-narios. To address this issue, we propose an unsupervised framework that eliminates the reliance on clean ground truth data. Experimental results highlight the superiority of our method over other unsupervised, model-based approaches, particularly in terms of image reconstruction quality. No-tably, our proposed method demonstrates competitiveness with the supervised learning method while circumventing the need for labeled training data.