基于生成对抗网络的可见光与红外图像融合  

Visible and Infrared Image Fusion Based on Generative Adversarial Network

作　　者：刘锃亮 张宇[1] 吕恒毅[1] LIU Zengliang;ZHANG Yu;LYU Hengyi(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;College of Optoelectronics,University of Chinese Academy of Sciences,Beijing 100039,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学光电学院,北京100039

出　　处：《无线电工程》2022年第4期555-561,共7页Radio Engineering

基　　金：国家青年科学基金(62005269)。

摘　　要：图像融合是图像处理领域中非常重要的分支,可见光图像与红外图像的融合在机器感知、目标检测与追踪、监控、遥感和图像去雾等方面扮演着十分重要的角色。针对目前一些融合算法时效性差、复杂程度高、泛化程度低和融合后图片信息丢失量大等问题,在神经网络FusionGAN的基础上进行了改进。在其中引入了一种多尺度卷积PSConv和一种轻量化注意力模块ECA-Net,前者能够在更细粒度角度进行多尺度特征融合,后者能自适应地选择一维卷积核大小,从而实现性能上的提优。实验采用经典的红外与可见光数据集TNO和NIO数据集,经实验表明,改进后的算法在主观评价与客观评价下,与原算法和其他算法相比有着明显提高。Image fusion is a very important branch in the field of image processing.The fusion of visible image and infrared image plays a very important role in machine perception,target detection and tracking,monitoring,remote sensing,image defogging and so on.In order to solve the problems of some current fusion algorithms,such as poor timeliness,high complexity,low degree of generalization,and large amount of information loss after fusion,the neural network FusionGAN is improved.A multi-scale convolution PSConv and a lightweight attention module ECA-Net are introduced.The former can perform multi-scale feature fusion at a more fine-grained angle,and the latter can adaptatively select the size of one-dimensional convolution kernel to achieve performance optimization.Classical infrared and visible data sets TNO and NIO are used in experiment.Experimental results show that the improved algorithm is significantly improved as compared with the original algorithm and other algorithms under subjective and objective evaluation.

关 键 词：图像融合 红外图像 可见光图像 FusionGAN PSConv ECA-Net 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]