LESFuse: A lightweight dual-domain collaborative framework for high-fidelity visible-infrared image fusion

Visible-infrared image fusion is crucial for robust perception in challenging environments, yet the inherent modality gap often leads to structural distortion and detail loss. To address this, we propose LESFuse, a novel lightweight fusion paradigm that establishes a dual-domain collaborative mechanism. Our approach introduces an intensity-structure interaction model to enforce spatial consistency and a learning-based frequency decomposition strategy to disentangle and enhance multi-scale features. Extensive experiments on three public datasets (MSRS, RoadScene, and TNO) against seven state-of-the-art methods demonstrate that LESFuse achieves superior performance. Quantitatively, our method attains the highest scores across multiple metrics, achieving a Mutual Information (MI) score of 3.71 on the MSRS dataset, significantly outperforming the second-best method. In downstream object detection tasks, LESFuse yields the highest mean Average Precision (mAP@0.5) of 83.16%. Furthermore, the framework maintains exceptionally low computational costs, requiring only 0.03M parameters and 1.85 GFLOPs, with an average inference time of 0.11 s on a CPU. These results confirm LESFuse's effectiveness and real-time capability for deployment on resource-constrained platforms.