Research On Sampling And Reconstruction Techniques Of Hyperspectral Images Based On Compressed Sensing

The sheer volume of data obtained by hyperspectral remote sensing imaging poses great pressure to the storage and transmission of imaging system as well as subsequent precessing,affecting the in-depth application of hyperspectral imaging.Based on the premise that the signal has sparseness or compressibility,the compressed sensing theory utilizes the direction acquisition of the signal at a lower sampling rate than the Nyquist rate with a measurement matrix,and reconstructing the original signal from a small number of measurements by solving an optimized problem.This data compression method provides the possibility of further development of hyperspectral remote sensing imaging.Analyzing the basic framework of compressed sensing theory and the characteristics of hyperspectral images,the key technologies including the sparse representation,measurement matrix design,reconstruction model establishment and inplementation of reconstruction algorithms aimed to hyperspectral images,are systematically and deeply investigated in this dissertation.The main research contents and contributions of this dissertation are summarized as follows.Sparse representation of hyperspectral images: Based on the study of basic theory of sparse representation and features of hyperspectral images,Gabor redundant dictionary is constructed and the sparse representation of hyperspectral images is realized by orthogonal matching pursuit method.In order to reduce the computational complexity of orthogonal matching pursuit method,a sparse representation algorithm based on particle swarm optimization is proposed.The matching process in orthogonal matching pursuit method is optimized by the particle swarm optimization,which ameliorates the computation efficiency of orthogonal matching pursuit method dealing with the sparse representation based on redundant dictonary.The experimental results show that,compared with orthogonal wavelet transform basis,the Gabor redundancy dictionary has the stronger sparse representation capability for hyperspectral images.The proposed sparse representation algorithm has a higher computational efficiency than the orthogonal matching pursuit method.Compressed sensing measurement matrix of hyperspectral images: Analyzing the construction methods of the commonly used measurement matrices and the difficulty degree of physical hardware realization,a measurement matrix based on scrambled block Hadamard ensemble is designed.Learning from the concept of random structured matrix and using the block diagonal matrix composition,Hadamard matrix is choosed as the diagonal element to save the storage space.Using central limit theorem,the designed matrix is proved to have Gaussian behavior with the advantages of random Gaussian measurement matrix,which is suitable for compressed sensing of hyperspectral images.The experimental results show that,compared with the Gaussian random matrix,the designed measurement matrix has high sampling efficiency and is suitable for hyperspectral compressed imaging.Compressed sensing reconstruction of hyperspectral images based on circulant sampling: Based on the analysis of spatial and spectral correlation of hyperspectral images,a three dimensional compressed sensing reconstruction algorithm for hyperspectral images is proposed.During the sampling process,three dimensional circulant sampling for the whole hyperspectral data cube is introduced.Considering the gradient sparsity of different band images in the reconstruction process,an optimal reconstruction model based on three dimensional total variation minimum is established and solved by the Augmented Lagrangian multiplier method to obtain the reconstructed images.With the idea of nonuniform sampling and residual reconstruction,a compressed sensing reconstruction algorithm based on spectral prediction is proposed.In the sampling process,the hyperspectral images are grouped into reference band images and non-reference band images,which are sampled separately at different sampling rates.In the reconstruction process,the reference band images are reconstructed by general reconstruction algorithm.Referring to the non-reference band images,spectral prediction is introduced to increase the available information for reconstruction process,and the reconstruction of non-reference band images is achieved in an iterative way.Experimental results show that the reconstruction accuracy of hyperspectral images could be improved by utilizing the the gradient sparsity or spectral prediction.Compressed sensing reconstruction of hyperspectral images based on multihypothesis prediction: Considering the spatial correlation within one band image and the spectral correlation between different band images,a compressed sensing reconstruction algorithm based on spatial-spectral multihypothesis prediction for hyperspectral images is proposed.A spatial-spectral multihypothesis prediction model is constructed,and the multihypothesis prediction matrix containing spatial prediction information and spectral prediction information is established.The residual reconstruction is achieved in an iterative way.Based on the characteristics of block compressed sampling,a global measurement matrix as well as its corresponding reconstruction model are constructed,and a reconstruction algorithm based on spatial smoothness feature and spectral correlation is proposed.The compound regular optimization reconstruction model using total varation and multihypothesis prediction residual as regularization items is established and solved by the Augmented Lagrangian multiplier method and the alternating direction method.The experimental results show that the two reconstruction algorithms could both improve the reconstruction accuracy,fully reveal that it is an an effective way to improve the accuracy by exploring the characteristics of hyperspectral images.Compressed sensing reconstruction of hyperspectral images based on spectral unmixing: Based on the analysis of spectral mixing characteristics of hyperspectral images,a compressed sensing reconstruction algorithm for hyperspectral images based on spectral unmixing is proposed.The basis of this algorithm is that the hyperspectral data could be expressed as the product of endmember matrix and abundance matrix.The coherent sampling matrix is designed to obtain the spatial measurement data which still satisfies the linear mixing model.The spectral sampling is performed to obtain the spectral measurements.In the reconstruction process,the joint optimization model for endmember extraction and abundance estimation is established using the idea of variable splitting.And then the resulted problem is solved by an alternative iterative way to obtain the whole reconstructed hyperspectral images.The experimental results show that the reconstruction algorithm could effectively improve the reconstruction accuracy as well as reduce the computational complexity.