Research On Compressed Sensing For Hyperspectral Images Algorithm With Deep Learning

Hyperspectral images consist of a series of spectral vectors with continuous bands.Hyperspectral images contain more information and can be used in a wider range of applications than traditional RGB images.However,the large amount of data of hyperspectral images brings difficulties to storage and transmission.Compressed sensing theory can sample hyperspectral images at a very low sampling rate,compress the size of the image by eliminating the redundant information in the hyperspectral image,and then reconstruct the image through the reconstruction network.However,traditional hyperspectral image compressed sensing algorithms still use an iterative method to reconstruct compressed data,which makes the reconstruction time of these algorithms too long.The combination of deep learning and compressed sensing can solve the problems of high hardware pressure and lack of timeliness faced by current hyperspectral image imaging systems.Therefore,based on the method of deep learning,this paper studies the compressed sensing algorithm of hyperspectral images.The main work and achievements include:(1)A multi-scale dense compression sensing network for hyperspectral image(MDCSN)is proposed in this paper.This algorithm aims to solve the problems of low reconstruction quality and long decoding time of the current traditional hyperspectral image compressed sensing algorithm.The algorithm designs a three dimensional compressed data sampling module to save more hyperspectral image information.At the same time,it designs a multi-scale dense feature extraction module,which uses the idea of Inception and Shortcut to extract and restore hyperspectral image features.Experiments show that the MDCSN algorithm is superior to most hyperspectral image compressed sensing algorithms in terms of reconstruction quality,and achieves an order of magnitude improvement in decoding time.(2)A compressive sensing based on spectral multi-scale dense residuals for hyperspectral Imaging(CSSMDR)is proposed in this paper.The algorithm aims to improve the reconstruction quality and hardware suitability of the model.The improved sampling module conforms to the working mode of the push-sweep hyperspectral sensor,and obtains better spatial dimension reconstruction quality by improving the ratio of spatial dimension and spectral dimension.Multi-scale convolution is added to the sampling module to eliminate spectral dimension redundancy in advance.Experiments show that the CSSMDR algorithm is superior to other hyperspectral image compressed sensing algorithms in terms of reconstruction quality and decoding time.