Résumé: To obtain a better fusion effect for infrared and visible images, a generative adversarial network using a dual-branch generator with matched dense blocks is proposed. The dual-branch generator consists of two parallel sub-networks, namely the upper and lower branches, which are asymmetrical in structure. It could be applied to nonlinearly extract the textural and contrast information in multiple degrees of freedom. Based on the dual-branch structure, two dense blocks are optimally designed by selectively arranging reduced concatenation connections to effectively employ the shallow information. As a result, both are non-full connection and symmetrically added on the upper and lower branches, respectively. Additionally, a gradient loss function containing the mean square error function was applied in the generator loss function, which could help extract more textural detail information. With such a generator and under adversarial learning with the discriminator, it could allow the fused images to preserve more visible and infrared information while also produce satisfactory visual perception. Experiments were implemented based on the open datasets, which included contrast and optimization experiments. The results demonstrate that the proposed method has superiority in terms of more detail and salient contrast in faint features which is relative to other state-of-the-art methods, and the applied dual-branch structure with the matched dense blocks is an appropriate for better fusion effect. The proposed method could be applied in certain detection or monitoring fields for infrared and visible image fusion.

Langue: Anglais