| Aiming at the problems of easy loss of image detail information in hyperspectral image superresolution process,inability to effectively extract high-frequency information and low-frequency information in hyperspectral images,and the strong correlation between hyperspectral image spectral band,the spatial spectrum attention combined with hyperspectral image super-resolution method is proposed.First of all,in the design of the overall network,the use of hyperspectral image spectral bands between the strong correlation,the use of the number of spectral bands in the image for grouping strategy,so as to carry out super-resolution algorithm,for each group in the feature extraction stage using convolution of different sizes of convolutional kernels,to obtain multi-scale receptive field features,better extraction of low-frequency information and lowfrequency information in low-resolution images,help to retain the feature information of the original image;and then the acquired image features are enhanced by the attention mechanism of "space-spectrum" combination,the spectral dimension information is used to assist the spatial dimension information to reconstruct the features;finally,the features of each group are fused,and the checkerboard effect is alleviated by the pixel-level deconvolution layer,and the clear highresolution image is output.Comparing the proposed algorithm with the latest six super-resolution algorithms,the experimental results show that the space spectral attention combined with hyperspectral image super-resolution methods in this paper reached 39.8697,31.9422 and 39.1729 on the three public datasets of Chikusei,Pavia center scene and CAVE,respectively,and the structural similarity reached 0.9376,0.8786 and 0.9572,respectively,which has a clear performance advantage over the latest six super-resolution algorithms.Compared with other algorithms,the proposed algorithm combines the advantages of multi-receptive field feature extraction module and space spectrum combined with attention mechanism,and the image detail characteristics after super resolution have been significantly improved,and the overall image quality has been further improved.There are 35 figures,6 tables and 66 references in this paper. |
