AUTOMATIC SYSTEM FOR PHYTOPLANKTONIC ALGAE IDENTIF

143AUTOMATICSYSTEMFORPHYTOPLANKTONICALGAEIDENTIFICATIONJos6LuisPech-Pacheco',2,GabrielCristobal',Josu6Alvarez-Borrego2andLeonCohen3'InstitutedeOptica(CSIC).Serrano121.28006Madrid.Spain.*DivisidndeFisicaAplicada,DepartamentodeOptica(CICESE).Krn107CarreteraTijuana-Ensenada,EnsenadabajaCalifornia,Mexico.'HunterCollegeandGraduateCenterofCUNY.695ParkAvenue,NY10021ABSTRACTWeproposeanautomaticsystemfordiatomlocalizationandidentificationwithamodularstructure.Themaincontributionofthisworkistoprovideacompleteautomaticsystemfortheanalysisofphytoplanktonicsamplesinbrightfieldmicroscopy.Theoverallprocedureconsistsintwoparts:first,framegatheringatlowmagnificationandsecond,furtheranalysisathighermag-nification.Atlowmagnificationthegoalistoobtainapanoramicoverviewofthefullsamplebytilingeachframe.Subsequentprocessingstepswillprovidethelocalizationandsizeofeachparticleineachframeforfurtheranalysis.Thelocalizationmethodbasedonimagefusiontechniquesprovidesmorerobustandaccurateparticledetectionthanothermethodsreportedintheliterature.Fromthesizeinformationweobtainausefulcueabouttheobjectivetouse.Athighermagnificationwedevelo-pednewautofocusingtechniquesprovidingafastandaccuratefocusedimage.Becauseparticlespresentavolumetricstruc-ture,weproposetheuseofmultifocusfusiontechniquesformerginginasingleplanethefocusedpartsfromneighbouringthebestfocusedimage.Thenweappliedaparticleselectionanalysistoreducethenumberofimagestoanalyze,i.e.todiscrimi-natediatomsfromdebris.Thisisthemostchallengingstep,duetolargevariabilityofshapes,diatomfragmentation,particleoverpopulationanddiatomhiding.Thelatterisnotdescribedinthepresentpaperandwillbethesubjectforaforthcomingpublication.Finally,fordiatomidentificationweusethescaletransformtechniqueandacepstrum-basedcross-correlationtechnique.Keywords:automaticslidescanning,identification,registration,cepstrumanalysis.RESUMENSeproponeunsistemaautomaticoconunaestructuramodularparalalocalizacidneidentificacidndediatomeas.LAcon-tribucidnprincipaldelpresentetrabajoesproporcionarunsistemaautomdticocompletoparaelandlisisdemuestrusfto-plactdnicasenmicroscopiadetransmisidn.Elprocedimientogeneralconsisteendospartes:enprimerlugai;laadquisicidndeimdgenesenbajnsaumentosyunasegundafusedeanalisisaaumentosmayores.Enbajosaumentoselobjetivoesobte-nerunuvisidnpanorarnicadelamuestracompletapormediodeunmosaicode10sdiferentescampos,demaneruqueeneta-passucesivasdeprocesadopuedadeterminarselalocalizacidnytamafiodecadaparticulaparasuposterioranalisis.Elme'tododelocalizacidnestabasadoenlautilizacidndete'cnicasde,fusidn,proporcionandounamasrobustayprecisadetec-cidndelasparticulasqueotrasme'todosdescritosenlaliteratura.Lainformacidndeltamafio,extraidaenestaetapa,pro-porcionaunaclovejihndamentdUlahoradeelegireltipodeobjetivoautilizarenmayoresaumentos.Paraestosultimosaumentos,proponemosluutilizacidndenuevaste'cnicasdeautoenfoquequeproporcionandeunamanerarapidalamejorimagenenfocada.Dadoquelasparticulaspresentanunaestructuravolume'trica,seproponeelusodete'cnicasdefusidndeimcigenesmultifoco,parapresentarenunsimpleplanolaspartesenfocadasde10splanoscercanosalaimageridemejorfoco.Acontinuacidnseefectuaunprocesodeseleccidndeparticulasparareducirelnumerodeimagenesaanalizui;esdecir;conobjetoderealizarunadiscriminacidndeltipodiatomea-materiaresidual.Estaultimaetapapresentuunagrandifcultad,debidoalagranvariabilidaddeformas,alafragmentacidndediatomeas,Ulasuperpoblacidndeparticulasyalasuper-posicidndece'lulas.Estaultimapartequedafueradelalcancedelpresentearticuloyseraobjetodeunaprdximapublicacidn.Finalmente,paruelprocesodelaidentificacidndediatomeassehanusadoticnicasbasadasenlatransforrnadadeescalayte'cnicasdecorrelacidncruzadaatrave'sdelcepstrum.Palabrasclave:Analisisautomaticodemuestras,identificacidn,registro,analisiscepstral.LimneticaZO(1):143.158(2001)0Asociaci6nEspaiioladeLimnologia,Madrid.Spain.ISSN:0213-8409144Pech-Pachecoetal.INTRODUCTIONAlthoughsomestudieshavelookedattheanaly-sisofimagesatmediummagnification,theydonotconsidertheproblemastepaheadi.e.startingtheautomatizationprocessfromasingle(orafew)low-mediummagnificationimages.Thisisaverychallengingproblem,becausethesamplemaybecontaminatedwithdebrisoritmayevencontainbrokenspecimens.Weidentifytwobroadclassesofdiatomidentificationsystems:semiautomaticandautomatic.Semiautomaticmethodsrequirehumaninteractiontoselectini-tialspecimensfromalowresolutionimage.Obviously,thisgroupofmethodsprovidesgoodresultsbutitistime-consumingandtedious.MethodsfordiatomdetectionandidentificationhavebeenstudiedinCairnseta1.,1972;Cairnsetal.,1979;Culverhouseetal.,1996andPech-Pacheco&Alvarez-Borrego,1998.Cairnsetal.(1972)haveproposedsomediatomidentificationmethodsbasedoncoherentopticsandhologra-phy.However,theydonottackletheproblemoftheautomatizationoflowresolutionimages.Culverhouseetal.(1996)derivedsomemethodsforphytoplanktonidentificationbasedonneuralnetworksbut,again,theydonotprovideafullyautomaticmethod.Pech-Pacheco&Alvarez-Borrego,(1998)haveproposedahybridoptical-digitalmethodforidentificationoffivedifferentspeci