Research On Hyperspectral Image Compressed Sensing Based On Spectral-Spatial Joint Compressed Model

With the development of remote sensing technology,hyperspectral image related technology has made great progress.Due to the characteristics of combining spectral information and spatial information,hyperspectral images are widely used in mineral exploration,border surveillance,search and rescue,military reconnaissance and other fields.Hyperspectral images can collect more accurate spectral information,which also generates a huge amount of data.Therefore,how to make full use of the prior knowledge of hyperspectral images to compress them and reconstruct high-quality images has become an urgent problem to be solved.Compressed sensing and convolutional neural networks,as two hot technologies in recent years,provide new ideas for hyperspectral image compression.This dissertation takes compressed sensing and convolutional neural networks as the starting point,and conducts research on hyperspectral image compression.First,a hyperspectral image compression sensing algorithm based on low rank total variation sparse constraint has been proposed.At the compression side,to solve the problem that existing algorithms do not consider the correlation between spectral domain,a compression method is designed that the sparse transforms are applied in both spectral and spatial domains.On the reconstruction side,in order to make full use of the prior knowledge of hyperspectral images,the low-rank three-dimension total variation sparse model has been proposed and the reconstruction model is solved by the fast iterative shrinkage thresholding algorithm.The experimental results prove that the compressed sensing algorithm for hyperspectral images can effectively improve the quality of reconstructed hyperspectral images while reducing the sampling ratio.Secondly,we combines compressed sensing and deep neural network technology to design an end-to-end hyperspectral image compressed sensing network based on Convolutional Neural Network.At the compression side,our method uses two types of convolution to remove redundant information between the hyperspectral image spectral and spatial domains,respectively.At the reconstruction side,corresponding to the compression side,two kinds of convolutions are used to decompress the observation data to complete the reconstruction initialization.At the same time,a residual network that combines multi-scale spectral and spatial information is designed to improve the quality of the reconstructed hyperspectral image.In terms of loss function,compared with the traditional mean square error loss function,a three-dimensional total variation loss term is added to fully mine the high-level structural information of hyperspectral images to further improve the reconstruction quality.Experiments conducted on many hyperspectral image data sets show that out algorithm can not only provide a reconstructed hyperspectral image closer to the image before compression,but also the reconstruction speed is greatly improved compared to the iterative hyperspectral image compression sensing algorithm.