LIALFP: Multi-band images synchronous fusion model based on latent information association and local feature preserving

The fusion of multi-band images, i.e., far-infrared image (FIRI), near-infrared image (NIRI) and visible image (VISI), is mainly confronted with three-sided challenges: One is whether image fusion process could be performed synchronously. Most of existing algorithms are aimed at two fusion targets, which makes them have to adopt sequential way to merge multiple images. Unfortunately, this may make their results vulnerable to ambiguity or artifact. The second is the ground truth of fused results could not be obtained at all in some image fusion fields (e.g., multi-band images). That leads immediately to the failure of supervised methods to give full play to their advantages. Third, the latent projection between the result and source images is often not directly considered. Notably, the relation involves not only the fusion result with all inputs, but also with each original. In order to solve aforementioned problems, this paper establishes a unsupervised representation learning model for synchronous multi-band images fusion. First, significant pixel fusion features are extracted to ensure that the primary information can be integrated. Secondly, the potential relationship between the result and the whole originals is assumed to be the linear mapping, reducing the unpredictability of these fusion results. In addition, these transformation matrices have been given the function of feature selection, which could choose discriminant features and project them into the fusion space. Then, the locally significant features of each source are captured by designed graph Laplacian matrix. Finally, experiments show the rationality and superiority of our algorithm through comparison with a variety of recent advanced algorithms from subjective judgment and objective indicators.