Frequency-Interaction Mamba for Remote Sensing Image Super-Resolution

Abstract

Image super-resolution (SR) aims to recover a high-resolution (HR) image from its low-resolution (LR) counterpart, yet accurately reconstructing texture details in complex regions remains challenging. This issue is particularly pronounced in remote sensing images (RSIs), which contain diverse textures and small-scale objects. LR RSIs often suffer from weakened high-frequency (HF) textures and blurred object boundaries, making it difficult to recover both global structures and fine details. Although Transformer-based methods have achieved promising performance, global self-attention usually incurs substantial computational costs. Mamba provides an efficient approach to modeling long-range dependencies, but existing Mamba-based methods remain limited in modeling HF details and 2-D spatial structures. To address these limitations, we propose a frequency-interaction Mamba (FIM) network that jointly enhances long-range dependency modeling, texture reconstruction, and spatial structure representation. Specifically, we design a FIM module (FIMM), in which a frequency enhancement module (FEM) strengthens reconstruction-relevant HF information before state-space modeling and selectively enhances informative texture responses, improving the recovery of edges, textures, and small-scale objects. In addition, we develop a spatial refinement module (SRM), which complements 2-D spatial information through global-guided attention (GGA) and an activation-gated feed-forward network (AGFN), improving the preservation of structural continuity and geometric relationships. Extensive experiments demonstrate that the proposed FIM contains only 15.4M parameters while effectively recovering both global structures and fine-grained details in RSIs.