Spatial-Frequency Synergy for Remote Sensing Image Super-Resolution with Holistic Feature Enhancement

Xiaole Zhao, Zitao Dai, Xun Xu, Xiaobo Zhang, Xun Gong, Yan Yang, Tianrui Li

Keywords: Super-Resolution, Remote Sensing, Holistic Feature Enhancement, Dual-Domain Feature Interaction

TL;DR: A novel lightweight SR model for remote sensing images promoted by dual-domain interaction and holistic feature enhancement.

Abstract: High-resolution (HR) remote sensing images are essential for various applications of Earth observation, but hardware limitations often give rise to degraded and low-resolution (LR) acquisitions. Super-resolution (SR) has currently emerged as a popular manner to ease this issue. However, most existing SR methods fail to effectively exploit the synergism between frequency and spatial information, while also suffering from inadequate feature enhancement. In this work, we present a novel model for remote sensing image SR, termed as Spatial-Frequency Synergy Network (SFSN). Firstly, it holistically boosts hierarchical features from both the channel and spatial dimensions, through Adaptive Channel Shifting (AdaCS) and Multi-Scale Large Kernel Attention (MS-LKA), respectively. Meanwhile, a Dual-Domain Interaction Attention (DDIA) is devised to explicitly model the mutual effect between spatial and frequency domains, enabling synergic feature refinement and HR detail recovery. It also delivers a versatile solution for bridging the spatial-frequency domain gap in remote sensing SR tasks. Extensive experiments on benchmark datasets have demonstrated the superiority of our SFSN over advanced SR models quantitatively and qualitatively, while maintaining considerably low overhead.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 11487