Study On Super-resolution Reconstruction Method Of Hyperspectral Images Based On Depth-interpretable Network

Hyperspectral images,which combine spatial as well as spectral information in real scenes,simultaneously image target areas in tens to hundreds of consecutive and subdivided spectral bands,and play a crucial role in remote sensing,aerospace,and exploration.However,it is usually difficult to obtain images with high spatial and spectral resolution due to hardware constraints such as semiconductors and chips,so super-resolution reconstruction is an effective way to solve this problem.In recent years,super-resolution reconstruction using fusion of auxiliary images with hyperspectral images has attracted a lot of attention.There are two main categories of such fusion-based methods,traditional methods and deep learning-based methods.Traditional fusion methods are more complex and less effective,and with the introduction of deep learning,such deficiencies have been somewhat compensated.In contrast,deep learning-based methods often suffer from insufficient interpretability and insufficient utilization of some important priors.Therefore,to address the above problems,this paper proposes a hyperspectral image super-resolution reconstruction method based on deep interpretable networks in a deep learning framework.First,an iterative threshold shrinkage algorithm(ISTA)based interpretable spectral image assisted fusion method is proposed to enhance the interpretability of the network as well as to utilize the information in the transform domain of the image.In this paper,the highly interpretable ISTA unfolding network is embedded as a network module in MHF-Net,and a deep network structure with high interpretability is constructed.This method first constructs a fusion problem model based on the hyperspectral image observation model,then iteratively solves it using the ISTA algorithm,and then expands it into a deep interpretable network.Since this method adds symmetry constraints to the network model,the network can better utilize the information in the transform domain of hyperspectral images in order to extract richer features for reconstruction.Finally,the effectiveness of the network structure proposed in this study and the good reconstruction results achieved by it are demonstrated experimentally.Second,an interpretable spectral image assisted fusion method based on non-local similarity is proposed to utilize the high frequency information of hyperspectral images.In this paper,a new fusion model is proposed with the addition of non-local prior constraints on hyperspectral images.After obtaining the fusion model,the nonlocal autoregressive model is used to guide the design of the deep interpretable network,and the nonlocal regularization set becomes a trainable module within the network framework.Since the present method utilizes fewer network layers stacked to obtain richer information and makes full use of the similarity of image nonlocality,the network performs image edge reconstruction more effectively and obtains better reconstruction results.Finally,the correctness of the fusion model proposed in this method and the good performance and generalization ability of the interpretable network are demonstrated experimentally.