Block-based compressive low-light-level imaging

Abstract

In this paper, block-based compressive low-light-level imaging (BCL-imaging) is studied. To obtain larger measurement SNR (signal to noise ratio), instead of object pixels, linear combinations of pixels, referred to as features, are collected. PCA and Hadamard features are studied. Measurement SNR and reconstruction error are analyzed to quantify BCL-imaging performance. Compared with conventional imaging, BCL-imaging presents better reconstruction quality. Between PCA and Hadamard projections, PCA has smaller reconstruction error. However, after sorting the projection vectors using measurement SNR, Hadamard can obtain similarly performance as PCA. Biased vector and dual-measurements are studied with experimental results for the implementation of both projections in the end of this paper.