Compressive hyperspectral imaging with non-zero mean noise

Compressive hyperspectral imaging (CHI) has attracted widespread attention due to its advantage of snapshot by encoding the 3D spectral image into a 2D measurement. The bottleneck of CHI for the real application now lies in the limited reconstruction quality, for which one of the fundamental reason is the inaccurate modeling of the measurement noise. Our key observation is that in practical scenarios, the measurement is inevitably contaminated with a positive offset (i.e., noise mean) due to the unideal imaging conditions (e.g. stray light and dark current), resulting in serious degradations of the reconstruction quality. In this paper, we propose to model the real noise with non-zero mean that generalizes the traditional zero mean noise to faithfully characterizing the optical imaging principle, and then introduce a novel reconstruction method with a goal to boost the reconstruction quality. During the reconstruction, the noise mean is estimated gradually to its convergence by measuring the deviation of the intermediate reconstruction result from the system measurement. We demonstrate the superior performance of our method by experiments on synthetic data and real capture data with a hardware prototype.