CFMB-T: A cross-frequency multi-branch transformer for low-quality infrared remote sensing image super-resolution

Abstract

Aiming at the problems of low resolution and fuzzy details of infrared remote sensing images, paper proposed a multi branch super-resolution method based on wavelet decomposition and a dataset processing method based on real degradation process. To improve the super-resolution reconstruction effect, the modulation transfer function is used to add the degradation process of the real environment to the dataset at first, and based on this, CFMB-T is proposed. CFMB-T took SwinIR as the architecture, separated the frequency information of the infrared remote sensing image through wavelet decomposition, and introduced a degenerate prior to focus on the recovery of high-frequency information. Moreover, CFMB-T also introduced a recursive module to ensure the ability of feature extraction while reducing model parameters. The detailed experiments on degraded datasets show that CFMB-T achieves better reconstruction results than other super-resolution models. Compared with SwinIR, with limited data and training rounds, the PSNR and SSIM of CFMB-T are improved by 1.07% and 0.71%, respectively, over SwinIR on the 4x super-resolution task, and 3.28% and 0.09%, respectively, on the 2x super-resolution task. The results demonstrate its great potential for infrared remote sensing super-resolution reconstruction.