FaGAN: Frequency-aware Generative Adversarial Network for Infrared and Visible Image Fusion

Abstract

Recently, how to better maintain the frequency information from source images has become a key issue in image fusion. In this article, we first analyze this issue and then propose a frequency-aware generative adversarial network, namely FaGAN, to fuse infrared images and visible images. Specifically, our network is composed of a generator and two discriminators. The generator is used to generate the fused images. The two discriminators are used to distinguish the high-frequency information differences and the low-frequency information differences between the source images and the fused images, respectively. In addition, we propose a specific frequency loss to better constraint the generator, which will better retain the high-frequency information and the low-frequency information in the fused images. Extensive experiments on VIFB (visible and infrared image fusion benchmark) well demonstrate that our network achieves competitive performance in terms of both objective evaluation metrics and visual effect.