图像检索中相关反馈的半监督主动学习研究的中期报告AbstractThisreportpresentstheprogressofourresearchonsemi-supervisedactivelearningwithrelevantfeedbackforimageretrieval.Imageretrievalhasbecomeoneofthemostimportantapplicationsinthefieldofcomputervisionduetotheincreasingnumberofimagesavailableinvariousdomains.However,theperformanceofimageretrievalalgorithmsheavilyreliesonthequalityoftheannotatedtrainingdatasincemanualannotationisatime-consumingandlabor-intensivetask.Toaddressthisissue,activelearninghasbeenproposedtoautomaticallyselectthemostinformativesamplesforannotation.Meanwhile,relevantfeedbackfromuserscanbeincorporatedintoactivelearningtofurtherenhancetheperformanceofimageretrievalsystems.Inthisreport,wereviewrelatedworkonsemi-supervisedactivelearningandrelevantfeedback,anddiscussourproposedmethodsandpreliminaryexperimentalresults.IntroductionImageretrievalhasbeenwidelyusedinmanyfields,suchase-commerce,socialmedia,andvideosurveillance.However,thequalityofimageretrievalalgorithmsreliesheavilyontheavailabilityandqualityofannotatedtrainingdata.Manualannotationisatime-consumingandlabor-intensivetask,whichmakesitdifficulttoscaleuptolargedatasets.Inaddition,theannotationqualitymaybeinconsistentduetodifferentannotatorsorsubjectivecriteria.Toaddresstheseissues,activelearninghasbeenproposedtoautomaticallyselectthemostinformativesamplesforannotation.Activelearningisaniterativeprocessthatstartswithasmalllabeleddatasetandgraduallyexpandsittoreducetheannotationcostwhilemaintainingorevenimprovingtheperformanceofthemodel.Thekeyideaistoselectthesamplesthataremostinformativeforthemodel'straining,i.e.,thesamplesthatthemodelismostuncertainorconfidentbutwrongabout.Therearevariousselectioncriteriaforactivelearning,suchasuncertaintysampling,diversity,density,andrepresentativeness.However,activelearningonlyconsiderstheintrinsicpropertiesofthesamplesandignorestheadditionalinformationthatexternalsourcescouldprovide.Relevantfeedbackfromusers,suchasrelevancefeedback,preferencefeedback,orsimilarityfeedback,can