Research On Compressed Sensing Reconstruction Method Of Hyperspectral Image Based On Deep Learning

Hyperspectral image(HSI)is a three-dimensional image composed of a series of band images.Since HSI can provide rich spectral and spatial information,a series of image analysis can be done to extract useful information in HSI.However,the large amount of data in HIS may bring out a heavy burden on HSI storage and transmission.Therefore,how to recover hyperspectral images with only a small amount of observations is an important issue for hyperspectral applications.Compressed Sensing(CS)samples the signal at a low sampling rate and reconstructs the signal through a reconstruction algorithm.However,the traditional compressed sensing reconstruction algorithms require many iterations to solve the reconstruction problems,which increase the calculation time.The deep learning technology can quickly reconstruct an image by using the trained model and avoid the parameter settings in the traditional compressed sensing reconstruction algorithm.Therefore,this thesis proposes a patch-based residual network for compressively sensed hyperspectral images reconstruction and a spectral-spatial residual dense network for hyperspectral image reconstruction.The main researches and innovations of this thesis are as follows:1)This thesis studies the current research on compressed sensing from three aspects: compressed sensing theory,measurement matrix design and reconstruction algorithm design.Three existing algorithms for HSIs reconstruction are introduced including: Denoising-based Approximate Message Passing Algorithm(D-AMP),Total Variation Augmented Lagrangian Alternating Direction Algorithm(TVAL3)and ReconNet.We compare these reconstruction algorithms by conducting experiments on a public data set.The experimental results show that ReconNet has the highest reconstruction quality at a low sampling rate of 0.1,and requires the least time to reconstruct an image among all the compared algorithms.2)This thesis proposes a patch-based residual network for compressively sensed HSIs reconstruction.Since the convolutional neural network requires the input to be an image,and the observed signal is a one-dimensional measurements,we use the fully connected layer to generate the initial image.But if we reconstruct the whole imgae,it will bring millions of parameters due to the fully connected layer.So we use a patch-based reconstruction method.Our method consists of two residual networks: One residual network is a reconstruction network for compressive sensing reconstruction and the other residual network is a deblocking network for removing the blocky effect,which is caused by patch-based reconstruction.The experimental results show that the patch-based reconstruction network is better than D-AMP,TVAL3 and ReconNet.In addition,it can effectively reconstruct HSI and greatly reduce the computation time of reconstructing an image.Moreover,the performance of deblocking can be enhanced by combining more patches into a larger patch fed into the deblocking network.3)This thesis proposes a spectral-spatial residual dense network for hyperspectral image reconstruction.This method adopts the idea of residual dense blocks,which can extract multi-layer features.In addition,to characterize the strong correlation between hyperspectral adjacent bands,we construct a spectral difference reconstruction network.This network is added to the dense residual netwok for reconstruction and the adjacent inter-spectral difference regular term is added to total loss.The experimental results show that our algorithm has better spatial quality than D-AMP,TVAL3,ReconNet and patch-based residual network for compressively sensed HSIs reconstruction at a low sampling rate of 0.01.Although the spatial resolution is not as good as patch-based residual network for compressively sensed HSIs reconstruction at high sampling rates,it is far superior to the patch-based residual network for compressively sensed HSIs reconstruction and other existing compressed sensing reconstruction algorithms at spectral resolution.And the speed of reconstructing an image is faster than D-AMP,TVAL3 and ReconNet.