融合显著信息的层次特征学习图像分类.pdf

《融合显著信息的层次特征学习图像分类.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、计算机研究与发展DOI：10．7544／issnl000—1239．2014．20140138JournalofComputerResearchandDevelopment51(9)：1919—1928，2014融合显著信息的层次特征学习图像分类祝军赵杰煜董振宇(宁波大学信息科学与工程学院浙江宁波315211)(jun．zhu．CS@gma1．corn)ImageClassificationUsingHierarchicalFeatureLearningMethodCombinedwithImageSaliencyZhuJun，ZhaoJieyu，andDo

2、ngZhenyu(CollegeofInformationScienceandEngineering，NingboUniversity，Ningbo，Zhejiang315211)AbstractEfficientfeaturerepresentationsforimagesareessentialinmanycomputervisiontasks．Inthispaper，ahierarchicalfeaturerepresentationcombinedwithimagesaliencyisproposedbasedonthetheoryofvisuals

3、aliencyanddeeplearning，whichbuildsafeaturehierarchylayer-by—layer．Eachfeaturelearninglayeriscomposedofthreeparts：sparsecoding，saliencymaxpoolingandcontrastnormalization．Tospeedupthesparsecodingprocess，weproposebatchorthogonalmatchingpursuitwhichdiffersfromthetraditiona1method．Thesa

4、lientinformationisintroducedintotheimagesparserepresentation，whichcompressesthefeaturerepresentationandstrengthensthesemanticinformationofthefeaturerepresentation．Simultaneously，contrastnormalizationeffectivelyreducestheimpactoflocalvariationsinilluminationandforeground—backgroundc

5、ontrast，andenhancestherobustnessofthefeaturerepresentation．Insteadofusinghand—crafteddescriptors，ourmodellearnsaneffectiveimagerepresentationdirectlyfromimagesinanunsuperviseddata～drivenmanner．ThefinalimageclassificationisimplementedwithalinearSVMclassifierusingthelearnedimagerepre

6、sentation．Wecompareourmethodwithmanystate—of——the—artalgorithmsincludingconvolutionaldeepbeliefnetworks，SIFTbasedsinglelayerormulti—layersparsecodingmethods，andsomekernelbasedfeaturelearningapproaches．TheexperimentalresultsontwocommonlyusedbenchmarkdatasetsCaltech101andCaltech256sh

7、owthatourmethodconsistentlyandsignificantlyimprovestheperformance．Keywordsfeaturelearning；hierarchicalsparsecoding；imagesaliency；imageclassification；saliencymaxpooling摘要高效的图像特征表示是计算机视觉的基础．基于图像的视觉显著性机制及深度学习模型的思想，提出一种融合图像显著性的层次稀疏特征表示用于图像分类．这种层次特征学习每一层都由3个部分组成：稀疏编码、显著性最大值；F聚(saliencym

8、axpooling)和对比度归一化．通过在图像层次稀