提出了一种应用于红外图像仿真的级联多尺度信息融合生成对抗网络,能由可见光图像估计对应的红外图像。针对可见光与红外图像特征之间的关联与区别,该网络采用级联的对抗网络结构:第一级对抗网络以语义分割图像为辅助任务,使用大感受野的卷积网络结构,重建红外图像的结构信息;第二级对抗网络以可见光的灰度反转图像为辅助任务,采用小感受野的网络结构,补充红外仿真图像的细节纹理信息,并使用多尺度融合模块整合多感受野信息以提升算法精度。在先进算法的通用数据集上进行实验,结果表明,级联多尺度信息融合对抗网络能够实现可见光到红外图像的转换,可得到结构与纹理都较正确的红外仿真图像,在多种客观指标与主观感受上均优于其他类似算法。

In this paper, we propose a cascade multi-scale information fusion generative adversarial network (CMIF-GAN) for infrared image simulation, which can estimate the infrared map from a visible image. Inspired by the connections and differences between visible and infrared features, CMIF-GAN adopts a cascaded structure composed of two levels of adversarial networks. With a large overall receptive field, the first-level adversarial network focuses on reconstructing structural information of the infrared image, and adds a semantic segmentation image task as auxiliary information. To enrich detailed texture information of the infrared image, the second-level adversarial network uses the grayscale inverted visible (GIV) images as auxiliary information and adopts a small overall receptive field network. Otherwise, the second-level adversarial network can integrate the multiple receptive information by a multi-scale fusion module (MFM) to improve algorithm accuracy. Experiments on public dataset demonstrate that CMIF-GAN can efficiently translate visible images to corresponding infrared images, and outperform previous methods in objective metrics and subjective vision.