Hyperspectral Image Compression and Reconstruction Based on Block-Sparse Dictionary Learning

A large amount of hyperspectral image (HSI) data poses a significant challenge for transmission and storage. A new signal processing mechanism—compressed sensing (CS)—is appropriate for processing signals with a massive amount of data and can achieve high reconstruction accuracy. According to the structural properties of HSI, the same ground features show the same spectral properties. In this paper, an approach is proposed to compress and reconstruct HSI based on CS and block-sparse dictionary learning. Primarily, a dictionary of a given set of signal is trained and prior knowledge is not required on the association of the training dataset into groups. Then, a measurement matrix is used to compress an HSI cube to reduce the data volume of the signal. Finally, we use the trained block-sparse dictionary to reconstruct the image, along with the HSI feature classification information. Our experimental results showed that, for block-sparse HSI data, the proposed approach significantly improved the performance compared with other related state of the art methods.