基于GAN的红外与可见光图像融合算法

Infrared and Visible Image Fusion Based on GAN

针对生成对抗网络中目标消失的问题,我们提出一种用于红外和可见光图像融合的新型生成对抗网络。我们提出的GAN图像融合算法包含三个部分。第一部分利用视觉显著性图提取红外图像的显著区域,并对这些区域的信息进行保存。第二部分将可见光图像、红外图像和显著信息在发生器中进行充分融合,得到预融合图像。在第三部分中,鉴别器试图区分预融合图像和可见光图像,以便基于对抗机制从可见光图像中学习细节。实验数据证明,我们的方法在定性和定量上都优于其他主流方法。

To overcome the object vanishing problemin Generative Adversarial Network(GAN),in this paper,we present a novel GAN for infrared and visible image fusion.Our proposed method contains three parts.In the first part,the salient regions of infrared image are extracted by visual saliency map and the information of these regions are preserved in purpose.In the second part,the visible image,infrared image and salient information are merged thoroughly in the generator to gain a pre-fused image.In the third part,the discriminator attempts to differentiate the pre-fused image and visible image,in order to learn details from visible image based on the adversarial mechanism.Experimental data testify that our method outperforms other leading methods quantitatively and qualitatively.