FIPFuse: Visible and Infrared Image Fusion Based on Frequency Information Perception

The fusion of visible and infrared images aims to generate fused images that has the advantages of both modalities. Traditional methods typically rely on manually designed strategies for feature extraction and fusion, resulting in increased computational complexity. The existing deep learning methods ignore the utilization of the frequency information of the source images. These methods are typically unsupervised and constrained by loss functions during the training. However, the limitations of previous loss functions have led to training that focuses excessively on the local information region of the source images, ignoring important global information. In order to solve the above problems, this paper proposes a visible and infrared image fusion method based on frequency information perception designed as FIPFuse. The frequency information perception module in FIPFuse is utilized in both the image reconstruction and image fusion stages, effectively extracting high-frequency and low-frequency information from the source images. A random structure similarity loss was proposed to train the model, with the objective of guiding the network to sufficiently focus on the source image information, thereby effectively fusing the source images. Extensive experiments on public datasets have demonstrated that FIPFuse outperforms other representative methods in both qualitative and quantitative assessments.