欢迎来到天天文库
浏览记录
ID:54372443
大小:792.02 KB
页数:7页
时间:2020-04-30
《光滑逼近超完备稀疏表示的图像超分辨率重构.pdf》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、第39卷第2期光电工程Vl01.39.NO.22012年2月Opto—ElectronicEngineeringFeb.2012文章编号:1003—501X(2012)02—0123—07光滑逼近超完备稀疏表示的图像超分辨率重构路锦正,2,张启衡,徐智勇,彭真明2(1.中国科学院光电技术研究所,成都610209;2.电子科技大学光电信息学院,成都610054;3.中国科学院研究生院,北京100049)。摘要:为改善单帧降质图像的分辨率水平,提出了一种新的基于稀疏表示的学习法超分辨率图像重构方法。针对信号在既定的欠定超完备字典下的非稀疏性问题,采
2、用光滑的递减函数逼近范数以避免对稀疏度先验的依赖,从而实现待重构图像块的有效稀疏表示,同时通过梯度下降的迭代优化获得稳定的收敛解。与双立方插值相比,图像的三倍超分辨实验显示,图像峰值信噪~b(PSNR)提高2dB,框架相似t~(SSIM)fP.善0.04,重构图像剔除了更多的模糊退化及边缘伪迹。该方法适于单帧降质图像的超分辨率增强。关键词:稀疏表示;超完备字典;光滑。范数;超分辨率重构中图分类号:TP391.4文献标志码:Adoi:10.3969~.issn.1003.501X.2012.02.023ImageSuper-resolutionR
3、econstructionBasedonSmoothlyApproximateOver-completeSparseRepresentatiOnLUJin.zheng,,ZHA_NGQi.heng,XUZhi.yong,PENGZhen.ming2(1.InstituteofOpticsandElectronics,ChineseAcademyofSciences,Chengdu610209,China;2,SchoolofOptoeleetronieInformation,UniversityofElectronicScienceandTec
4、hnology,Chengdu610054,China;3.GraduateUniversityofChineseAcademyofSciences,Beijing100049,China)Abstract:Toimproveresolutioncapacityofthedegradedimage,alearning—basedsuper-resolutionreconstructionmethodviasparserepresentationoverover-completedictionaryisintroduced.Duetonon—sp
5、arsestrepresentationofsignalwithrespecttogivenill--conditioneddictionary,thesuggestedsmoothedL0normsparse—-representationtechniqueoverblindsparsitywithcontinuousdescendingfunctioncanexhaustivelyexploitgivenspecificdictionary,achievingefectivesparsedecompositionoflowresolutio
6、nimagepatch.Afterwards,thestableandconvergentsolversareobtainedfromoptimizationofgradientsteepestdescent.Experimentalresultsdemonstratethat,comparedtoBicubicinterpolation,thePowerSignaltoNoiseRatio(PSNR)gainofimagethrice-zoomediscloseto2dB,andtheimprovementofStructuralSimila
7、rity(SSIM、isalmost0.04.Moreover,thesuper-resolvedimageseliminatedexcessiveblurringdegradationandannoyingedgeartifacts.Theproposedmethodcanbeefectivelyappliedtoresolutionenhancementofdegradedsingle-image.Keywords:sparserepresentation;over-completedictionary;smoothedL0norm;sup
8、er-resolutionreconstruction0引言基于学习的图像超分辨率重构(Super-resolutionReconstruction,
此文档下载收益归作者所有