Compressed Sensing Reconstruction of Hyperspectral Images Based on Spatial-spectral Multihypothesis Prediction

Compressed Sensing (CS) reconstruction of hyperspectral images driven by spatial-spectral multihypothesis prediction is proposed in order to take full advantage of spatial and spectral correlation of hyperspectral images. The hyperspectral images are grouped into reference band images and non-reference band images, and the reference band images are reconstructed by Smoothed Projected Landweber (SPL) algorithm. For the non-reference band images, the spatial-spectral multihypothesis prediction model is introduced to improve the reconstruction accuracy. Multihypothesis predictions drawn for an image block of non-reference band image are made not only from spatially surrounding image blocks within an initial non-predicted reconstruction of non-reference band image, but also from the corresponding position and neighboring image blocks within the reconstruction of reference band image. The resulting predictions are used to generate residuals in the projection domain, and the residuals are reconstructed to revise the prediction values. The residuals being typically more compressible than the original images and the iterative execution mode lead to improved reconstruction quality. Tikhonov regularization is utilized to solve the weight coefficients of multihypothesis prediction and structural similarity is used as a criterion to decide whether to change the search window size or not. Cross validation is presented to compute the criterion parameter of iteration termination. Experimental results demonstrate that the proposed algorithm outperforms alternative strategies only using spatial correlation or spectral correlation to predict or not employing prediction and the peak signal-to-noise ratio of its reconstructed images is increased by more than 2 dB.