Image Annotation and Retrieval Using Dual Classifier Formulation
Keywords:
I mag e An notation, Asy m ptotic complexity, AutomaticImageAnnotation(AIA), TranslationModel(TM), continuous-spacerelevancemodel(CRM)Abstract
Automatic image annotation is a difficult and highly relevant machine learningtask.Recentadvanceshavesignificantlyimprovedthestate-of-the-art inretrievalaccuracy withalgorithmsbasedonnearestneighborclassification incarefullylearnedmetric spaces. Butthis comesat
apriceofincreasedcomputationalcomplexityduringtrainingand testing
.WeproposeFastTag,anovelalgorithmthatachievescomparableresultswithtwosimplelinearmappingsthat arecoregularizedinajointconvexlossfunction. The lossfunctioncanbeefficientlyoptimized inclosedformup-dates,whichallowsusto
incorporatealargenumberofimagedescriptorscheaply.Onseveralstandard real-worldbenchmark data sets, wedemonstrate
that FastTagmatches thecurrentstate-of-the-artintaggingqual-ity,yetreduces the trainingand testingtimesbyseveralorders
ofmagnitudeand haslowerasymptoticcomplexity
