基于特征挖掘与跨层互补学习的花卉图像分类

打开文本图片集

中图分类号：TP391.4；TP181 文献标识码：A文章编号：2096-4706（2025）17-0052-06

Abstract：Flowerimageclasifcationisdevoted toidentifyingfowercategorieswithsubtlediferencesbetweenclasses andsignificantdiferenceswithinclasses.Thekeyishowtoextractstrongdiscriminantfeaturesfromhighlysimilarcategories. At present， mostof the mainstream methods use attention-based region localization methods，but these methods face two keylimitations：oneistofocus toomuchonafewsignificantregions，resultingintheneglectofpotentialkeydiscriminant regions;thesecondisthelackof modelingoffeaturecomplementarityresultinginsemanticredundancyoflocalfeatures.In viewof the aboveproblems，this paper proposes afower imageclasification method basedonfeature miningandcros-layer complementaryleaing.Firstly，afeature miningmoduleisdesignedtoforce thenetwork toexplorenewdiscriminantfeatures bystrengtheningthemostsignificantfeaturesandthensuppresing them.Secondly，cross-layerfeaturecomplementaryleaing moduleisconstructed，ndcross-layer interactionmechanismisusedtofusemulti-featurecomplementarysmantics toance discriminant features.Theexperimentalresults show that the method performs wellonthe OxfordFlowers102 dataset.

Keywords： multi-layer feature; feature enhancement; feature complementarity;flower image classification

0 引言

在计算机视觉领域，花卉图像分类是一项具有挑战性的识别任务，其核心在于区分外观相似、类别间差异细微的花卉子类[1]。（剩余10944字）