Abstract: Aiming at the problem of information loss, such as details of remote sensing images reconstructed by a super-resolution algorithm, in order to ensure that remote sensor reconstruction images contain more texture and high-frequency information, a remote-sensitive image super resolution network is proposed based on a pyramid- based attention mechanism and the generation of confrontational networks. Firstly, a new pyramidal dual attention module is designed, including channel attention network and spatial attention network. Pyramid pooling is used instead of average pooling and maximum pooling in the channel attention network structure to enhance the feature representation capability from the perspective of global and local information. The spatial attention network structure adopts large scale convolution to expand the integration capability of local information, which can effectively extract texture, high frequency and other information. Secondly, the dense multi-scale feature module is designed to extract feature information at different scales using asymmetric convolution, and the extraction accuracy of texture, high frequency and other information is enhanced by fusing multi-level scale features through dense connection. Experimental validation is performed on the publicly available NWPU- RESISC45 dataset, and the experimental analysis shows that the algorithm outperforms the comparison methods in both subjective visual effect and objective evaluation metrics, and the reconstruction performance is relatively good. 

Key words: remote sensing images, super resolution, the pyramid ofdouble attention, dense multi-dimensional characteristics, asymmetric convolution