Study On Reconstruction Of Hyperspectral Images Based On Dictionary Learning And Compressed Sensing

The problems about hyperspectral images’storage and transfer are solved by compressed sensing. Though the number of the existing researches about gray images is large, the researches about hyperspectral image compressed sensing are not mature. In this dissertation, the exiting researches of hyperspectral image compressed sensing and the incomplete hyperspectral image reconstruction based on the dictionary learning was analyzed, and then the reconstruction algorithm of hyperspectral image was studied. Mainly completed the following work:First of all, a reconstruction algorithm based on vary sampling rate recovery is proposed to recover hyperspectral images, combined with the characteristic of hyperspectral image and the knowledge of compressed sensing. At the data sampling stage, hyperspectral image is divided into group according to the correlation among the spectral band. In one group, the sampling rate of the reference image is higher; the sampling rate of non-reference images is low. At the receiver, recover the reference image with smooth l0algorithm, and then recover the non-reference image based on the spectral correlation. Experiments show that such measure and reconstruct methods can gain good recovery results than the method which using the constant sampling rate.Secondly, a novel spatial-spectral dictionary learning algorithm is proposed. Hyperspectral image can be divided into3D small overlapping cube blocks base on the characteristics that it contained significant spatial and spectral correlations. A spatial-spectral dictionary which can represent these blocks sparsely is learned. First, assume the dictionary is fixed, calculate the sparse coefficients using non-negative orthogonal match pursuit; second, assume the coefficients are fixed, update the dictionary using gradient descent method, alternate between the two steps until the dictionary is converged. The dictionary learned from this model is consistent with the characteristics of hyperspectral images, at the spectral direction, the atoms are the spectral curves of the real materials; at the spatial direction, the atoms are ordinary2D spatial block dictionaries.At last, a reconstruction algorithm based on the spatial-spectral dictionary is proposed to recover incomplete hyperspectral images. At the data sampling stage, using random uniform matrix to sample the hyperspectral image. At the receiver, the data is divided into3D small overlapping cube blocks, and then recover the hyperspectral image with smooth l0algorithm. Experiments show that this algorithm can gain good recovery results with a low sampling rate.