Large Remote Sensing Image Super Resolution Using Self Attention Conditional Coordinate Network

Optical imagery is one of the most important sources for model remote sensing. Super resolution of the optical data could be very useful for industry and academia. Conventional solutions from computer vision only offers super resolution on small images describing relatively smaller scene and optimized towards human perception. In this paper, we proposed a new deep learning model for large remote sensing image super resolution. By using conditional coordinate and self-attention, the model could achieve arbitrary large image super resolution task with special focus on correctness of the details. Numerical evaluation of the model is carried out based on real satellite remote sensing data. Result shows a promising performance with PSNR
$> 33\ \text{dB}$
and SSIM
$> 0.93$
.