Infrared and visible image fusion based on structural re-parameterization

The purpose of infrared and visible image fusion is to enhance the detailed scene information in the source image by fusing the complementary information of different modalities. However, the existing deep learning methods have the problem of unbalanced fusion performance and computing resource consumption, and ignore the problem of noise in infrared images. Aiming at these two problems, this paper proposes an infrared and visible image fusion algorithm based on structural reparameterization. Firstly, the algorithm performs feature extraction on the two source images through a two-branch residual connection network with weight sharing, and the obtained features are cascaded to reconstruct the images. Then, the structural similarity loss and the content loss with bilateral filtering denoising are used to jointly guide the training of the network. Finally, after the training is completed, the structure reparameterization is performed to optimize the training network into a direct connection network. Qualitative and quantitative experiments are compared with seven leading deep learning fusion algorithms on multiple public data sets. The proposed fusion algorithm achieves the improvement of multiple evaluation indicators with lower resource consumption. The fusion results have richer scene information, stronger contrast and more in line with the visual effect of the human eye. © 2024 Northeast University. All rights reserved.