基于對(duì)比學(xué)習(xí)的三維模型分類

打開(kāi)文本圖片集

關(guān)鍵詞：三維模型分類；對(duì)比學(xué)習(xí)；卷積神經(jīng)網(wǎng)絡(luò)；注意力機(jī)制；遷移學(xué)習(xí)DOI：10.15938/j. jhust.2025.02.004中圖分類號(hào)：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號(hào)：1007-2683（2025）02-0032-10

Abstract：At present，3Dmodelclasification hasbeenaresearch hotspot.Masive3Dmodels notonlyhavediversityineach class，butalsoavesimilaritisbetweenclases，whichseriouslyafect theclasificationaccuracyof3Dmodels.Weproposea3D modelclassficatiomethodbasedootrastielaingIntstod，thengisdidedintompledisiatioagead aclassifictionstage.Inthestageofsamplediscrimination，3Dmodelsofthesamecategoryaremutuallpositivesamples，and3D modelsofothercategoriesaremutuallyegativesampls.Thecontrastivelossisusedtoconstraintesamplefatures，ndtepositive andnegativesamplesareappedtothesingle-centerunithypersphereinthesamespace toobtainagoodsemanticrepresentationspace of3Dmodellasification.Inaditon，inodertocapturethecorelatinbetweentheviewsandthekeyareasintheviews，amultiheadslf-attentionmoduleandspatialatentionmoduleareintroducedinteaper.Moreover，thechaelatentionisddednthe multi-headself-atentionmoduletoobtainthechanneldimensioninformation.Intheclasificationstage，thenetworkmodelis trasferredtotheclassficationtaskbyfine-uningtenetworkparametrstocompletethe3Dmodelclasification.Theexperiental results show that the classification accuracy of the 3D model respectively reaches 99.4% and 97.5% on the ModelNet1O and ModelNet40 datasets.

Keywords：3Dmodel classification；contrastive learning；convolutionneural network；atention mechanism；transferleaing

0 引言

人類的視覺(jué)感知具有三維立體性，并且三維模型相較于文本、圖像等擁有更加豐富的信息，因而三維模型在機(jī)器人、工業(yè)設(shè)計(jì)、無(wú)人駕駛等新興領(lǐng)域有著舉足輕重的作用。（剩余16242字）