DAGAN: A Generative Adversarial Network Based on Attention and Dual Discriminator for Infrared and Visible Fusion

In order to solve the problem of infrared and visible image fusion in which there exists incomplete display of source image edge information in the fused image as well as loss of some details in the visible image, this paper proposes a new approach to realize the framework of image fusion and detail complementarity called dual attention generative adversarial network (DAGAN). DAGAN can accurately capture the intensity distribution of different targets, while focusing on extracting the target edge information and background texture detail information. Specifically, the generator uses a dual attention network to simultaneously focus on the distribution of local key information in the visible light domain and the infrared domain, where the multi-scale attention of the dual discriminators can extract attention maps from different source images. Experimental results indicate that the method adopted in this paper shows more significant advantages compared with other fusion methods, among which the structural similarity, mutual information, information entropy, standard deviation, difference correlation synthesis, and correlation coefficient have been greatly improved.