LGFusion: Frequency-Aware Dual-Branch Integration Network for Infrared and Visible Image Fusion

Abstract

Multimodal image fusion aims to generate a fused image that integrates complementary advantages from different modalities, for example, target saliency from infrared images and texture details from visible images. However, existing methods struggle to jointly model global and local features, separate shared and modality-specific information, and provide semantic interpretability. To address these challenges, we propose LGFusion, a dual-branch parallel fusion framework that enables structure-aware and detail-preserving feature extraction and fusion. We introduce a frequency-based structural assumption: low-frequency components across modalities are correlated and represent shared background and layout, while high-frequency components are modality-specific, capturing details like thermal patterns or textures. Based on this, we enhance the consistency of low-frequency features and increase the discriminability of high-frequency features to improve the interpretability and controllability of the fusion process. Additionally, we design an efficient feature extraction module combining Efficient Attention and the lightweight MambaBlock, which models long-range dependencies and global context at low computational cost. To further enhance local detail modeling, we integrate a high-pass filter with a Transformer encoder, effectively capturing high-frequency features such as edges and textures. Extensive experiments show that our method achieves state-of-the-art performance in infrared visible image fusion and demonstrates strong generalization in medical image fusion tasks.