A contrastive learning-based iterative network for remote sensing image super-resolution

Many deep convolutional neural network(CNN)-based methods have achieved significant success in noise-free image super-resolution(SR) tasks. However, these methods produce unsatisfactory results for noisy remote sensing imagery. Recently, some practical SR models have been proposed to eliminate the negative impact of noise during reconstruction process, but they still have the problem of insufficient or excessive denoising. To address this issue, this article proposes a contrastive learning-based iterative network(CLIN) for noisy remote sensing image SR. Specifically, CLIN adopts an iterative cooperation approach, which includes an evaluator and a reconstructor. First, the evaluator evaluates the noise levels of low resolution(LR) images. Then the reconstructor utilizes LR images and their noise levels to reconstruct the SR images, which are returned to the evaluator for noise evaluation again. Furthermore, in order to make the reconstructor retain more spatial details, we design a global feature fusion block in the reconstructor to fuse the local features and the global features. To further suppress the noise, we propose a novel contrastive penalty strategy to train our model away from the noise domain. Compared with state-of-the-art SR methods, the peak signal to noise ratio (PSNR) improvements of our approach are about 0.04-0.78 dB on RSSCN7 dataset with a scale factor of 2. Qualitative and quantitative experiments on several noisy satellite image datasets demonstrate that the proposed CLIN achieves promising performance under different noise levels.