Abstract
Deep learning models are state-of-the-art in compressive spectral imaging(CSI) recovery. These methods use a deep neural network (DNN) as an imagegenerator to learn non-linear mapping from compressed measurements to thespectral image. For instance, the deep spectral prior approach uses aconvolutional autoencoder network (CAE) in the optimization algorithm torecover the spectral image by using a non-linear representation. However, theCAE training is detached from the recovery problem, which does not guaranteeoptimal representation of the spectral images for the CSI problem. This workproposes a joint non-linear representation and recovery network (JR2net),linking the representation and recovery task into a single optimizationproblem. JR2net consists of an optimization-inspired network following an ADMMformulation that learns a non-linear low-dimensional representation andsimultaneously performs the spectral image recovery, trained via the end-to-endapproach. Experimental results show the superiority of the proposed method withimprovements up to 2.57 dB in PSNR and performance around 2000 times fasterthan state-of-the-art methods.