Research On Reconstruction Algorithm Of Compressive Hyperspectral Imaging With LCVR Based On Deep Learning

Hyperspectral imagers are capable of acquiring two-dimensional spatial information in hundreds of spectral bands of a target,and are widely used in remote sensing,biomedicine and precision agriculture.Compressive sensing hyperspectral imaging technology breaks the limitation of Nyquist sampling rule,significantly reduces the data acquisition volume,and overcomes the pressure on data storage and transmission caused by huge amount of data in traditional hyperspectral imaging systems,which is the main direction of current hyperspectral imaging technology development.Hyperspectral image reconstruction is one of the key issue in compressive hyperspectral imaging technology.The current main information reconstruction algorithms are divided into two categories: traditional algorithms based on optimization theory and deep learning reconstruction algorithms.The traditional reconstruction algorithms have the problems of design priors manually,long reconstruction time,and low reconstruction accuracy;the existing deep learning reconstruction algorithms are mostly based on specific imaging systems and the reconstruction effect in the spectral dimension needs to be improved.Therefore,how to reconstruct hyperspectral images quickly while maintaining reconstruction accuracy is an urgent problem.In this thesis,we investigate high-precision and lightweight compressive hyperspectral imaging reconstruction method based on liquid crystal variable retarder compressive hyperspectral imaging system with deep learning methods.The main research contents and conclusions of this thesis are as follows:(1)Compressive hyperspectral imaging reconstruction algorithm based on improved DeepCubeNetBased on the characteristics of compressive hyperspectral imaging system with LCVR,compressed measurement images are simulated and four algorithms are used to reconstruct hyperspectral image,including Tw IST,TVAL3,GPSR,and Deep CubeNet.To address the problems of long reconstruction time and low accuracy of the optimized iterative algorithm,a compressive hyperspectral reconstruction imaging algorithm(3DASPP Deep CubeNet,3DASDCNet)based on the improved Deep CubeNet was proposed by adding an optimized atrous spatial pyramid pooling module to Deep CubeNet.Compared to TVAL3,the mean peak signal-tonoise ratio(MPSNR)was improved by 12.02 d B,and the mean spectral angle mapper(MSAM)was improved by an order of magnitude,achieving a second-scale reconstruction.Compared to Deep CubeNet,MPSNR was improved by 0.45 d B,MSAM was improved by 13.2%.The results show that the reconstruction performance of 3DAS-DCNet is more superior.However,large convolutional kernels are used in 3DAS-DCNet,which increases the complexity of the model.(2)Compressive hyperspectral reconstruction imaging algorithm based on atrous convolution lightweight networkTo address the problem of increasing complexity of 3DAS-DCNet model,a compressive hyperspectral imaging reconstruction algorithm(Atrous Convolution-based LightweightNetwork,ACLNet)was constructed by optimizing 3DAS-DCNet model through spatial spectral separation feature extraction module,spatial spectral separation pooling module and 3D recurrent criss-cross attention.Compared to 3DAS-DCNet,the model size,model parameters and the FLOPs of ACLNet are reduced by 4.24 MB,1.122 M and 337.041 G,respectively,the MPSNR was improved by 1.42 d B,and MSAM was improved by 5%.The results indicate that ACLNet effectively reduces the complexity of the model and has better reconstruction results.