Deep Convolutional Networks For Snapshot Hypercpectral Demosaicking

In this paper we introduce a novel demosaicking algorithm for snashot hyperspectral images. Snapshot cameras allow real-time hyper spectral imaging in uncontrolled environments. However, snapshot cameras with K bands capture only a single frequency band per pixel leaving the rest, K - 1 values, at the mercy of demosaicking algorithms. Demosaicking is a very challenging problem as interpolating unknown values in a highly under-sampled cube of radiances is likely to lead to aliasing artifacts in the interpolated output values. Most of the existing demosaicking approaches are hand crafted interpolation based methods. Inspired by the Deep Learning based breakthroughs in various areas of computer vision, we propose a snapshot hyperspectral demosaicking approach based on deep convolutinal networks. Experiments on the CAVE and Hytexila datasets show that our proposed method outperforms the state-of-the-art methods.