Spectral images (SIs) are a collection of images, where each image captures the scene of interest illuminated by light of a different color. Over the years, SIs have become a great asset in advancing a variety of fields, from medicine and biology to space exploration and astrophysics. Since different materials on the surface of planets, for instance, have different spectral signatures, spectral images may be used to determine the mineral composition of planets, which is key to enable sustainable interplanetary travel, and other human endeavors.
The problem is that: these images in their natural format not only may take long to collect, but they may also put a heavy load on transmission and storage. This results in the consumption of more energy, a scarcity in outer space.
To enable efficient sensing of SIs, snapshot compressive spectral imaging (SCSI) cameras have been developed. SCSI cameras capture the whole collection in a compressed format. Specifically, a coded grayscale snapshot of the scene is taken by the camera, and then the SI is recovered from the snapshot using computational photography. In doing so, an algorithm searches for the spectral image that leads to the snapshot, and simultanously complies with our prior knowledge of the scene.
Our goal here is to introduce such a knowledge, different to traditional approaches, by leveraging the intrinsic network structure of SIs. Like Facebook friends may be connected on a network by their affinities, pixels in SIs may be connected on a network by their spatio-spectral proximity. Provided such a network, we thus enable the design of low complexity algorithms that may not only operate close to real time when needed, but also allow us to reveal the color of a scene when coded grayscale is all we are allowed to obtain.