FaceDetection Algorithm andFeature Performance
Jilmil Saraf · 2007
Theperformance ofthreewellknownfacede- tection algorithms andfouralternative typesoffeatures are characterized using facedatafromtheFaceRecognition Grand Challenge. Thethreealgorithms area Semi-Naive Bayesian Classifier, a neural networkcalled a SNoW,anda Cascade Classifier usingHaarwavelets. Forthefirst twoalgorithms, ROC analysis isusedtoasses therelative valueofwavelet features compared tosimpler pixel features. Nouniversally best feature isobserved, andforimagery acquired underuncon- trolled lighting, pixels perform slightly better thanwavelets. TheCascade Classifier isfoundtobeimpossible totrain inthe samefashion astheotheralgorithms, butitisalsofoundto perform verywellusing atraining configuration supplied along withthealgorithm aspartoftheOpenCVlibrary. I.INTRODUCTION Detecting andthenlocalizing faces isuniversally thefirst stepinfacerecognition. Theresearch intohowbestto develop reliable facedetection algorithms isrelatively ma- ture. Several academic systems (11),(10)havedemonstrated relatively highlevels ofperformance onlarge datasets. Further, facedetection hasentered therealm ofcommercial products (2). Despite thissuccess, important openquestions remain regarding therelative merits ofdifferent decision pro- cedures, i.e. algorithms, andfeatures. Herewe compare threealgorithms alongwithfourtypesoffeatures on a series offacedetection problems derived fromim- agerycollected fortheFaceRecognition GrandChallenge (FRGC2.0) project (ttp://face. nist.gov/frgc/}) .Tethree algorithms considered areaSemi-Naive Bayesian Classifier, a typeofneural network called SNoW,anda cascade ofclassifiers. TheSemi-Naive Bayesian Classifier isbasedupontheworkdonebySchneiderman atCarnegie Mellon University (14). TheSNoW classifier isbasedupon analgorithm highlighted inasurvey offacedetection algo- rithms presented byYangetal.(8). Thecascade classifier is arguably thebestknownofthethree andisbased uponthe workofViola andJones(11). Thefourfeatures considered are5-3linear phase wavelets, Haarwavelets, Sobeledgesandpixels. The5-3linear phasewavelets areinteresting because theywereusedby Schneiderman inconjunction withtheSemi-Naive Bayesian Thismaterial isbasedinpartuponworksupported bytheNational