Polarization prior to single-photon counting image denoising.

Abstract

Single-photon counting (SPC) imaging technique, which can detect targets in extremely low light levels, has attracted considerable research interest in recent years. To reduce the influence of noise under the low light condition, traditional approaches typically seek various priors from images themselves to construct denoising models, leading to inferior performance as the signal and noise cannot be efficiently distinguished. To address this challenging problem, in this study we propose a novel polarization prior to SPC image denoising based on the observation that a special polarization SPC (PSPC) image has a higher SNR than the SPC image. It enables us to construct a polarization prior to the PSPC image that can transfer efficient targets' spatial details to the denoised SPC image, and hence improves the denoising performance. Specifically, we group similar patches of the PSPC image to form 'anti-noise' dictionaries with high SNR. Then we construct a non-local prior-oriented sparse representation constraint based on the fact that each noisy patch of the SPC image can be sparsely represented by the corresponding 'anti-noise' dictionary. According to this sparse representation constraint, we further formulate an SPC image denoising model by incorporating two terms, i.e., a negative Poisson log-likelihood function for preserving the data fidelity and a total variation constraint to reduce the influence of noise, which is solved by an efficient variable splitting method. In the experiment, we have verified the effectiveness of the proposed method from simulated and real data in terms of visual comparison and quantitative analysis, respectively.