MFEM: Multiscale Frequency-Enhanced Mamba for Lightweight Remote Sensing Image Super-Resolution

In recent years, remarkable progress has been achieved in lightweight remote sensing image super-resolution (RSISR) techniques. However, these methods, to maintain lightweight property and efficient computation, are often affected by inadequate feature extraction and limited receptive fields, which restrict the precise reconstruction of structurally complex remote sensing images (RSIs). To address this issue, we propose a multiscale frequency-enhanced Mamba (MFEM), which uses a unique multiscale convolutional structure to enhance feature extraction and uses a frequency-enhanced Mamba to capture long-range dependencies with linear complexity. Specifically, we design a multiscale blueprint residual block (MBRB) with three specifications, which allows the network to efficiently extract multiscale features of RSIs while maintaining lightweight property using different specifications of MBRB at varying network depths. Furthermore, we develop a frequency mamba enhancement block (FMEB), which enhances Mamba from within through frequency analysis to globally model multilevel features while improving local information loss. Extensive experiments demonstrate that our method achieves better super-resolution (SR) performance while maintaining lightweight property.