Object reconstruction from thermal and shot noises corrupted block-based compressive ultra-low-light-level imaging measurements

In this paper, block-based compressive ultra low-light-level imaging (BCU-imaging) is studied. Objects are divided into blocks. Features, or linear combinations of block pixels, instead of pixels, are measured for each block to improve system measurement SNR and thus object reconstructions. Thermal noise and shot noise are discussed for object reconstruction. The former is modeled as Gaussian noise. The latter is modeled as Poisson noise. Linear Wiener operator and linearized iterative Bregman algorithm are used to reconstruct objects from measurements corrupted by thermal noise. SPIRAL algorithm is used to reconstruct object from measurements with shot noise. Linear Wiener operator is also studied for measurements with shot noise, because Poisson noise is similar to Gaussian noise at large signal level and feature values are large enough to make this assumption feasible. Root mean square error (RMSE) is used to quantify system reconstruction quality.