A Recurrently Complicated Lightweight Network for Superresolution of Remote Sensing Images

Convolutional neural network (CNN) has made significant progress in image superresolution (SR), which could thrash the limits of image spatial resolution. Recently, abundant CNN-based methods have been proposed for the remote sensing image SR; however, the usages of complex structures and coarse manners could introduce excessive learnable parameters and ignorance of heterogeneous image details, respectively. In this article, we propose a recurrently complicated lightweight network (RCL-Net) for SR image recovery, through procedures of recurrent fluctuated complexity. We design a serial of the progressive complicated block (PC-Block) in the RCL-Net, and each PC-Block is composed of three complicated lightweight branches (CL-Branches) with increasing complexities in order, for recovering heterogeneous image details. Meanwhile, the CL-Branch is integrated with a multireceptive field module (MRF-Module) to more efficiently recover intact images through forward propagation paths of heterogeneous routes and lengths, where the excessive interactive calculations between feature subparts are constrained to reduce learnable parameters. In this manner, the proposed RCL-Net achieves a tradeoff between model complexities of traditional powerful structures, such as coarse-to-fine manners, and SR performances. Plentiful experiments with excellent results grounded on popular datasets exactly demonstrate the superiority of our proposed network, which even surpasses the advanced large SR model with less than 3% learnable parameters, compared to the state-of-the-art lightweight methods.