Multiple classifier systems: Theory, applications and tools

In many Pattern Recognition applications, the achievement of acceptable recognition rates is conditioned by the large pattern variability, whose distribution cannot be simply modeled. This affects the results at each stage of the recognition system so that, once this hasbeendesigned,itsperformancecannotbeimprovedoveracertainbound,despite the efforts in refiningeither the classification or the descriptionmethod. Starting fromthe earlynineties, someresearchgroupsconcentratedthe attention on a multiple classifier approach.The rationale of this approachlies in the assumption that, by suitably combining the results of a set of base classifiers, the obtained performance is better than that of any base classifier. In other words, it is claimed thattheconsensusofasetofclassifiersmaycompensatefortheweaknessofasingle classifier, while eachclassifier preservesits own strength. The implementation of a multiple classifier system implies the definition of a combiningruleorastrategyfordeterminingthemostlikelyclassasampleshouldbe attributedto,onthebasisoftheclasstowhichitisattributedbyeachbaseclassifier. Different combining rules and strategies, independent of the adopted classificationmodel,havebeenproposedsofar inthe openliterature.Inthis Chapterwe first presentasurveyofsuchapproaches,byconsideringbothdifferentarchitecturesand combiningrulesaswellasmethodsforconstructingensembles.Afterward,ataxonomy of the applications where a multiple classifier approach has been successfully appliedisorganized.Finally,someofthetoolscurrentlyavailableforimplementing a multipleclassifier system are presented.