Compressive Spectral Video Sensing using the Convolutional Sparse Coding framework CSC4D

Spectral Videos (SV) contain a scene’s spatial–spectral-time information. Just as with Spectral Images (SI), SVs require expensive sensing hardware, storage plus high frame ratios. Although Super Resolution techniques improve the quality of low-resolution SVs, Compressive Spectral Video Sensing (CSVS) senses high-quality SVs by extending the Compressive Sensing Image (CSI) techniques. CSI uses the universal Sparse Signal Representation (SSR) model for SVs and SIs despite the limited quality of the recovered signals. On the other hand, dictionaries synthesis models are used successfully for representing SIs, SVs, and in CSI. This work proposes the 4D convolutional sparse representation (CSC4D) for recovering full-resolution SV from CSVS measurements. It is based on a multidimensional formulation of the CSC model, profiting from its robustness without additional optical flow information. Extensive numerical simulations (two CSI architectures and noise models) show that the proposed CSC4D+CSVS improves the state-of-the-art in both quality and border sharpness by up to 1.5 dB.