SFDNet: spatial-frequency decoupled network for infrared and visible image fusion

Abstract

Infrared and visible image fusion aims to integrate complementary information from different modalities to generate images with both salient targets and rich textures. However, existing methods mainly rely on spatial feature modeling and lack an explicit mechanism to exploit frequency-aware representations, limiting their ability to preserve fine-grained details. To address this limitation, we propose a novel spatial-frequency decoupled learning framework, termed spatial-frequency decoupled network(SFDNet), which decomposes features into low-frequency semantic structures and high-frequency detail components for independent representation learning and adaptive fusion. Unlike conventional fusion networks that implicitly couple different frequency information, the proposed framework introduces a dedicated spatial-frequency interaction paradigm. Specifically, a dual-branch architecture is designed, where a semantic branch captures stable global structures, while a detail branch jointly models spatial structural information and frequency-aware representations. Furthermore, a spatial-frequency compensation mechanism is developed to bridge the discrepancy between spatial and frequency features, enabling complementary enhancement. In addition, a frequency-aware feature enhancement module is introduced in the reconstruction stage to adaptively modulate frequency responses, thereby improving detail preservation while maintaining structural consistency. Extensive experiments on multiple benchmark datasets demonstrate that SFDNet consistently outperforms existing methods in both visual quality and quantitative evaluations. Moreover, it improves downstream object detection performance, further demonstrating the effectiveness of the proposed spatial-frequency decoupled learning framework.