鲁棒迁移判别分析-综合字典对学习的带噪声图像分类

打开文本图片集

关键词：鲁棒；迁移学习；字典对学习；噪声图像分类

中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）04-036-1258-08

doi：10. 19734/j.issn. 1001-3695.2025.06.0261

Robust transfer discriminative analysis-synthesis dictionary pair learning for noisy image classification

Zhou Guohua 1，2 ，Han Shaoyong³4，Xu Yiqing1，Gu Xiaoqing§，Yin Xinchun2+ （1.DeptofIfoationEnring，ngouVcatioalItuteofdstchlog，gougs36，C；lee ofInformationEngering，YngouUiersityYngouJgu217，ina；3.SholofahematicsndComputerSieeo glingUniversityTnglingAnui24461，hina；4stdoctoralSientficesearchWkstationBankofZhengou，Zhengzo401， China；5.Scholof ComputerScienceandArtificialInteligence，Changzhou Uniersity，Changzhou Jiangsu213164，China）

Abstract：Real-world images are mostlyaccompanied bycomplex noise.Atthe same time，the image sources arecomplex and theirdatadiferencesarediferent.Toobtainaccuraterepresentationofnoisyimagesandimproverecognitionability，thisstudy proposedarobust transferdiscriminative analysis-synthesisdictinarypairleaingalgorithm（RTDAS-DPL）.Firstlyitapplied Gausiandistributionand Laplacian distribution todescribe the mostcommon Gaussiannoise andsalt-pepper noisein images， which brokeawayfrom the traditionalsingleGausian noise assumptionandenhancedthealgorithm’srobustnessincomplex noisyenvironments.econdlyitonstructedadomain-ivariantsubspaceviasharedanalsis-sythesisdictioarypisacross domains，anddesignedamaximummeandiscrepancystrategybasedonsparsecoeficients tominimizedistributionaldiferences betweensourceandtargetdomains.Then，itintroducedthediscriminativeconstraints tosimultaneouslyoptimizeintra-classco hesionand inter-classseparation，whichcould reduce thecorrelationamong atoms of diferentclasses while increased thecompactness withinthesameclas.Thus itobtainedhighlydiscriminativedictionarypairsandclasifiers.Experimentalesultson internet social media imageand cropleaf image datasetsdemonstrate the efectivenessof theproposed algorithm.

Key Words：robustness；transfer learning；dictionary pair learning；noisy image classification

0 引言

随着人工智能和5G网络等技术的高速发展，互联网每天都产生海量复杂的图像数据。（剩余16727字）