MRPFusion: lightweight infrared and visible image fusion network based on multi-scale reparameterization

Infrared and visible image fusion enhances scene detail by leveraging the complementary information from different modalities. However, many existing methods face challenges in achieving a good trade-off between fusion quality and computational efficiency, which limits their applicability in real-time scenarios and on devices with limited resources. In addition, traditional methods tend to prioritize infrared information while neglecting critical texture details from visible light images. To address these challenges, a lightweight fusion network based on multi-scale reparameterization (MRPFusion) is proposed to improve fusion quality while reducing computational costs. During training, a reparameterized multi-scale feature extraction block is used to enhance texture detail extraction, and skip connections are employed to combine high-resolution details with low-resolution semantic information, thereby optimizing fusion quality. In the fusion phase, an adaptive attention collaborative fusion module, which combines channel and spatial attention, is introduced to regulate the fusion process, so that fine-grained fusion can be achieved from global to pixel levels, thereby preserving and enhancing information robustness. During inference, structural reparameterization simplifies convolutional operations to reduce graphics processing unit memory usage and accelerate inference. Experimental results on datasets including MSRS, TNO, and M3FD show that MRPFusion outperforms nine state-of-the-art fusion algorithms across multiple metrics while significantly reducing computational resource consumption. Its superior performance in downstream tasks such as object detection and semantic segmentation highlights its potential for high-level vision applications.