WorksMathematics
Hyperspectral imaging from diffractograms
Abstract
We consider an optical system with a periodic point spread function, such as a camera with a periodic diffractive filter array placed near the sensor. Such systems produce a monochrome image that we call ``diffractogram'', and that multiplexes the spatial and spectral information about the scene. We present a method to efficiently invert this forward model and recover the hyperspectral scene. The inverse problem is formulated as a large scale -penalized least squares problem that is solved iteratively with the Alternating Direction Method of Multipliers (ADMM). A Krylov subspace solver, such as LSQR, may require tens to hundreds of iterations to approximatively solve the augmented least-squares problem required in each ADMM iteration. Here we present an exact direct solver for the augmented least-squares problems that exploits the periodicity of the point spread function. Each direct solve costs roughly as much as three LSQR iterations, eliminating inner iterations entirely. We validate the method on both synthetic and experimental datasets.