Research On Hyperspectral Image Super-Resolution Method Based On Deep Convolutional Neural Network

Hyperspectral images usually contain tens to hundreds of spectral bands,which can provide the spatial and spectral characteristics of the target area at the same time.Due to the superiority of hyperspectral images,it is widely used in many fields.However,there are limitations in the physical hardware of hyperspectral sensors,and there is a trade-off between their spatial and spectral resolution.Therefore,the spatial resolution is usually reduced to compensate for the spectral resolution.However,the lower spatial resolution will affect the development of subsequent tasks.To address this challenge,hyperspectral image super-resolution is devoted to reconstructing lowresolution hyperspectral images into high-resolution hyperspectral images.Depending on whether auxiliary information is used,hyperspectral image superresolution methods are usually divided into two categories: fusion-based hyperspectral image super-resolution methods and single hyperspectral image super-resolution methods.In this paper,based on convolutional neural network,spectral attention fusion network(SAFN)and branch-trunk super-resolution(BTSR)network are proposed respectively.SAFN is a fusion-based hyperspectral image super-resolution method.This method is a dual-path multi-node fusion network,which can effectively extract the feature information of the two paths.It takes the residual block as the basic core,and this simple and effective network structure can fully extract spatial information.The node fusion strategy has a good convergence effect on the information of multiple paths,and can extract the shallow and deep information in each path.In addition,the spectral attention block with attention mechanism is beneficial to extract the spectral map of pixels.On the CAVE dataset,SAFN is compared with six state-of-the-art methods.Both quantitative and qualitative evaluations show that the SAFN model has a strong super-resolution capability.BTSR is a super-resolution method based on a single hyperspectral image.This method is a novel branch-trunk structure network.Different from the previous methods,there is no independent extraction of spatial and spectral information,and there is no more emphasis on the mining of spatial information.In the branch network,spatial units and spectral units with simple structures are designed to extract spatial and spectral information,respectively.In the backbone network,a spatial spectral fusion module employing hybrid 2D and 3D convolutions can pay more attention to the extraction of spatial information.In order to alleviate the influence of too many branch structures on information aggregation,a deep feature extraction module is added to ensure the efficiency of information transmission.On the Pavia Centre and Chikusei datasets,the BTSR model is compared with six representative methods,all achieving the best results.In summary,the fusion-based hyperspectral image super-resolution method SAFN and the single hyperspectral image super-resolution method BTSR proposed in this paper have achieved superior performance and have huge application prospects.