Infrared and visible image fusion network based on spatial-frequency domain cross-learning

Abstract

The goal of infrared and visible image fusion is to combine complementary information from source images to generate fused images with high contrast, which can highlight salient targets while preserving rich texture details. Most deep learning-based fusion methods focus solely on the spatial domain, neglecting valuable frequency domain information. Furthermore, existing spatial-frequency fusion networks fail to fully exploit the advantages of both domains, resulting in limited fusion performance. To address this challenge, we propose a Spatial-Frequency Domain Cross-Learning Network (SFCFusion) for infrared and visible image fusion. Specifically, we have designed a frequency-domain learning branch that captures global feature information in the Fourier space, thereby more effectively preserving the global consistency of the source images. Additionally, we develop a spatial branch to extract local detail features and propose a Multi-scale Selective Enhancement Module (MSEM). Finally, when bridging the frequency and spatial branches, we observe that the feature information extracted from these two branches are complementary. To leverage this property, we design a Frequency-Spatial Cross-Guidance Module (FSCGM). This module employs a bidirectional guidance learning strategy to integrate critical information from both domains into each branch, thereby enhancing the quality of fused images. Extensive experiments on public datasets demonstrate that our method achieves significant advantages in terms of key fusion performance metrics and visual quality. It also exhibits robust performance in downstream detection tasks. Our code is available at https://github.com/HaodeShi/SFCFusion.