Perception inspired deep neural networks for spectral snapshot compressive imaging

We consider the inverse problem of coded aperture snapshot spectral imaging (CASSI), which captures the spatio-spectral data-cube using a snapshot 2D measurement and reconstructs the 3D hyperspectral images using algorithms. Recent advances of deep learning have boosted the image quality of the reconstructed hyperspectral images significantly, and this leads to an end-to-end real-time capture and reconstruction system. However, the network design for CASSI reconstruction is still at the incubation stage and usually an off-the-shelf network is employed and re-purposed. In this work, from a different perspective, inspired by the fact that most existing hyperspectral images are still in the visible bandwidth, we introduce the perceptual loss into the deep neural network for CASSI reconstruction. Extensive results on both simulation and real data demonstrate that with this small change, the reconstructed image quality can be improved dramatically using the same network.