卷积神经网络结合显著性目标掩图的红外与可见光图像融合  

Infrared and Visible Image Fusion Based on Convolutional Neural Network Combined with Saliency Target Mask

作　　者：万刘永 程健庆[1] 刘义海 WAN Liu-yong;CHENG Jian-qing;LIU Yi-hai(Jiangsu Automation Research Institute,Lianyungang 222006,China)

机构地区：[1]江苏自动化研究所,江苏连云港222006

出　　处：《舰船电子对抗》2022年第1期63-67,共5页Shipboard Electronic Countermeasure

摘　　要：传统的基于多尺度变换理论的红外与可见光融合,提取特征单一,融合规则需要手动设计,难以应对多场景的需求,而深度学习的方法具有良好的特征提取能力,能够对多种特征进行学习。使用基于卷积神经网络的方法对图像融合进行研究,将网络分为特征提取网络和特征融合网络2个部分。首先使用图像处理软件获得红外图像的显著性目标掩图,然后以目标掩图为基础定义卷积神经网络的损失函数,最后使用公共数据集对神经网络进行训练,并与传统的多尺度变换方法进行对比。实验结果表明,从主观评价角度来看,融合结果更符合人眼习惯,更有利于人眼识别,同时在客观评价指标上也均有所提高。Traditional infrared and visible light fusion is based on the theory of multi-scale transfor-mation,and for it,the extraction feature is single,and the fusion rule needs to be manually de-signed.It is difficult to deal with a variety of scenarios,while the deep learning method has good feature extraction ability which is able to learn a variety of features.This paper uses the method based on convolution neural network to study image fusion.The network is divided into two parts,feature extraction network and feature fusion network.Firstly,the salience target mask of infrared image is obtained by using image processing software;and then the loss function of convolutional neural network is defined on the basis of the target mask;finally,the neural network is trained by using public data set,and compared with the traditional multi-scale transformation method.The experimental results show that the fusion results are more consistent with human eye habits and more conducive to human eye recognition from the perspective of subjective evaluation,and the ob-jective evaluation indexes also has been improved.

关 键 词：图像融合 卷积神经网络 显著性目标掩图 梯度损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]