图像语义分割算法研究

《图像语义分割算法研究》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、万方数据AbstractWiththebloomofIntemetandthedevelopmentofmultimediatechnology,thereisaprolongedexplosiononthenumberofmultimediadata，represemedbywebimages．Massivedataalsobroughtchallengescomingfrommanyaspects，suchasthestorage，retrievalandmanagemem．Majorresearchenginesarebasedonth

2、etextdescriptionsofmultimediadata，whichisnotalwaysrelevanttothedataandthusleadstopoorperformances．Moreover,facinguptola唱e-scalewebimages，manymultimediadatadonothaveanytextdescriptionatall。Sohowtoautomaticallyandeffectuallyorganizeandmanagemassivedata,aswellashowtoeffectivel

3、yandefficientlyunderstandthesemanticmeaningofmassivedatahavebecometheburningquestions．Inthisthesis，wefocusontwocrucialtasksinlarge—scaleimageretrievalandHa-derstanding，i．e．，automaticimageannotationandimagesemanticsegmentation．Wetakeinsightintothecriticalproblemsofthesetasks

4、．andproposenovelapproaches．Automaticimageannotationtaskamistotraincomputerswithhuman·-labeledim--agestoaddsemanticlabelsonanunlabeledimage，asalistofobjectsthatappearintheimage．Comparingwithcontent-basedimageretrievalwhichestimatesthesimilar-ityamongimageswiththefeatureextra

5、cted，retrievalonsemanticlabelsismuchmorerapid．Theproliferationofwebimageslargelyenrichestheavailabilityofimageswithhuman-addedlabels．Soitisofgreatimportancetodevelopautomaticimageannotationtechniquesindealingwithlarge·scaleimageretrievalandunderstanding．Basedonimage-levelfe

6、atureandsemantic，themainlimitationofimageannota-tionmodelsisthattheycouldneitherobtainnorprovidethelocationofeachsemanticlabelintheimage，andthusleadstoinaccurateprediction．Aimingtoprovidethese-manticcategoryofeveryregionandpixel，semanticsegmentationrecentlyhasattractedgreat

7、researchinterests．Inordertobetterlearnthesemanticofimagesandsuperpixels，weproposeafullysupervisedsemanticsegmentationapproachusingmultiplegraphswithblock-diagonalconstraints．Describingimagesfromdifferentviews，eachfeatureisofitsuniqueprosandcons，anditisdifficulttoidenti母thei

8、mportanceoffeaturestodifferentcategories．Theproposedfully-supervisedapproachconstr