TransCycleGAN: An Approach For Remote Sensing Image Super-Resolution.

Satellite images are essential for providing geoinformation in Earth Science. Limited imaging devices make high-resolution images hard to obtain, bringing difficulties for precise predictions in many applications. In this paper, a novel unpaired TransCycleGAN network is proposed to super-resolve remote sensing images using pseudo-supervision. Specifically, TransCycleGAN is based on CycleGAN with an integrated effective degradation-removal Transformer module. Benefiting from pseudo-supervision, unpaired clean high-resolution and naturally degraded low-resolution images can be used to train our model. The introduced transposed self-attention mechanism allows for capturing global interactions between contexts. Extensive experiments on remote sensing benchmark datasets demonstrate the superiority of our proposed method against existing state-of-the-art algorithms in super-resolving remote sensing images.