LiMFusion: Infrared and visible image fusion via local information measurement

Most image fusion methods currently achieve satisfactory results under normal conditions. However, in complex scenes, the problem of information conflict between source images has not been effectively resolved. Sometimes, effective information from one source image may be masked by noise from another source image. In this paper, by organically combining the traditional decomposition strategy and the attention mechanism, we propose a novel infrared and visible image fusion network based on local information measurement, named LiMFusion. Specifically, the source image is decomposed by utilizing fourth-order partial differential equations to obtain high-frequency and low-frequency layers. Feature representation and information preservation capabilities are enhanced by well-designed spatial attention blocks as well as channel attention blocks combined with the UNet architecture. Moreover, a new localized image information measurement method based on histogram of oriented gradients and a spatial-aware loss function are proposed to make the fusion network more inclined to focus on the features of the region of interest. Extensive experimental results indicate that this method can more comprehensively reflect the brightness information and texture details of the source image, effectively addressing the challenges of fusion in complex environments. The source code will be publicly available at https://github . com /YQ -097 /LiMFusion.