文章编号:1002-2082(2024)01-0107-11基于显著性的双鉴别器GAN图像融合算法谢一博,刘卫国,周顺,李梦晗(西安工业大学光电工程学院,陕西西安710021)摘要:针对红外图像与可见光图像在不同场景的特征表达不同的问题,提出一种基于显著性的双鉴别器生成对抗网络方法,将红外与可见光的特征信息相融合。区别于传统的生成对抗网络,该算法采用双鉴别器方式分别鉴别源图像与融合图像中的显著性区域,以两幅源图像的显著性区域作为鉴别器的输入,使融合图像保留更多的显著特征;并将梯度约束引入其损失函数中,使显著对比度和丰富纹理信息保留在融合图像中。实验结果表明:本文方法在熵值(entropy,EN)、平均梯度(meangradient,MG)、空间频率(spatialfrequency,SF)及边缘强度(edgeintensity,EI)4个评价指标中均优于其他对比算法。该研究实现了红外图像与可见光图像高效融合,有望在目标识别等领域中获得应用。关键词:图像处理;生成对抗网络;图像融合;显著性区域;目标识别中图分类号:TN219;TP391文献标志码:ADOI:10.5768/JAO202445.0102005Saliency-baseddualdiscriminatorGANimagefusionalgorithmXIEYibo,LIUWeiguo,ZHOUShun,LIMenghan(SchoolofPhotoelectricEngineering,Xi'anTechnologicalUniversity,Xi'an710021,China)Abstract:Toaddresstheproblemthatinfraredimagesandvisibleimageshavedifferentfeatureexpressionsindifferentscenes,ansaliency-baseddualdiscriminatorgenerativeadversarialnetworkmethodwasproposedtofusetheinfraredandvisiblefeatureinformation.Differentfromthetraditionalgenerativeadversarialnetwork,adualdiscriminatorapproachwasadoptedtodiscriminatethesaliencyregionsinthesourceimagesandthefusionimagesrespectivelyinthisalgorithm,andthesaliencyregionsofthetwosourceimageswereusedastheinputofthediscriminatorsothatthefusionimageretainedmoresalientfeatures.Thegradientconstraintwasintroducedintoitslossfunctionsothatthesalientcontrastandrichtextureinformationcouldretaininthefusionimage.Theexperimentalresultsshowthattheproposedmethodoutperformsothercomparisonalgorithmsinfourevaluationindexes:entropy(EN),meangradient(MG),spatialfrequency(SF)andedgeintensity(EI).Thisstudyachievesefficientfusionofinfraredimagesandvisibleimages,whichisexpectedtogainapp...