Unpaired Remote Sensing Image Super-Resolution with Multi-Stage Aggregation Networks

Abstract

Mainstream image super-resolution (SR) methods are generally based on paired training samples. As the high-resolution (HR) remote sensing images are difficult to collect with a limited imaging device, most of the existing remote sensing super-resolution methods try to down-sample the collected original images to generate an auxiliary low-resolution (LR) image and form a paired pseudo HR-LR dataset for training. However, the distribution of the generated LR images is generally inconsistent with the real images due to the limitation of remote sensing imaging devices. In this paper, we propose a perceptually unpaired super-resolution method by constructing a multi-stage aggregation network (MSAN). The optimization of the network depends on consistency losses. In particular, the first phase is to preserve the contents of the super-resolved results, by constraining the content consistency between the down-scaled SR results and the low-quality low-resolution inputs. The second stage minimizes perceptual feature loss between the current result and LR input to constrain perceptual-content consistency. The final phase employs the generative adversarial network (GAN) to adding photo-realistic textures by constraining perceptual-distribution consistency. Numerous experiments on synthetic remote sensing datasets and real remote sensing images show that our method obtains more plausible results than other SR methods quantitatively and qualitatively. The PSNR of our network is 0.06dB higher than the SOTA method—HAN on the UC Merced test set with complex degradation.