Multi-angle Remote Sensing Images Super-Resolution Reconstruction Using Dynamic Upsampling Filter Deep Network

  Objectives  Deep learning based on super-resolution reconstruction technology has been widely used in multi-temporal hyperspectral images, high-resolution image reconstruction. Multi-angle remote sensing images have rich complementary information, which is suitable for super-resolution reconstruction.

  Methods  An end-to-end super-resolution reconstruction method based on dynamic upsampling filter network is proposed for high-resolution multi-angle remote sensing images. The network of the method includes an end-to-end two-way network, in which one branch is used dynamic upsampling filter block to improve the image resolution. Another branch network is used to learn the high-frequency information in the image. In order to verify the effectiveness of the proposed method, 2, 3 and 4 times super-resolution reconstruction simulation experiments and real experiments are carried out with Worldview-2 multi-angle remote sensing images from Atlanta, America and Rio de Janeiro, Brazil, respectively. Several groups of comparative experiments are carried out.

  Results and Conclusions  Experimental results show that the proposed method can effectively improve the spatial resolution of the target image while taking into account the angle dimension information of multi angle images, and effectively maintain the image details.