Wavelet-based Frequency Decoupling Refinement Network for Remote Sensing Image Super Resolution

Deep learning methods have demonstrated remarkable capabilities in image super-resolution. However, remote sensing images present unique challenges due to the heterogeneous content, multi-scale characteristics, and abundance of high-frequency details. In light of recent advancements, we rethink the architecture design for super-resolution of remote sensing images, introducing a wavelet-based frequency decoupling refinement network (WFDRN) to reconstruct lost information from low resolution images. Specifically, we leverage discrete wavelet transform (DWT) to decouple the deep features into low-frequency (LF) and high-frequency (HF) component, which are subsequently processed by specialized network architectures incorporating self-attention mechanisms and vanilla convolutions, respectively. Furthermore, we cascade multi-level wavelet units to expand the receptive field and capture multi-scale contextual information, activating more pixels to contribute to reconstruction. Quantitative and qualitative experiment results on two remote sensing datasets demonstrate that our method effectively restores details, and surpasses traditional single image super-resolution methods.