LIGFusion: a light-integrated and guided fusion method for multimodal image fusion

Multimodal image fusion plays a crucial role in complex scenarios such as nighttime surveillance and autonomous driving, with infrared-visible image fusion receiving significant attention. We propose light-integrated and guided fusion (LIGFusion), a fusion guidance method designed to address the challenges of uneven illumination, limited global information extraction, and high-frequency information loss in existing multimodal image fusion techniques. LIGFusion introduces a light-discriminative perception module and a light-adaptive loss function, enabling dynamic adjustment of feature extraction and fusion strategies based on illumination conditions. To enhance global feature extraction, a transformer-based module is incorporated to mitigate the limitations of convolutional neural networks. In addition, difference of Gaussian feature extraction and fusion are employed to preserve high-frequency details during the forward propagation of the fusion network. Experimental results on the multi-spectral road scenarios for practical infrared and visible image fusion (MSRS), RoadScene, and the TNO multiband image data collection (TNO) datasets demonstrate that LIGFusion achieves superior performance in thermal target saliency, detail preservation, and noise suppression. The proposed method outperforms mainstream approaches such as TIMFusion and TUFusion across multiple evaluation metrics. Ablation studies validate the effectiveness and rationality of the proposed network modules. Furthermore, LIGFusion proves effective in object detection and semantic segmentation tasks, providing high-quality fused images to support applications in complex real-world scenarios. The code of LIGFusion is available at https://github.com/zuiyuejvshi/LIGFusion