简介:Anewmethodofsuper-resolutionimagereconstructionisproposed,whichusesathree-step-trainingerrorbackpropagationneuralnetwork(BPNN)torealizethesuper-resolutionreconstruction(SRR)ofsatelliteimage.ThemethodisbasedonBPNN.First,threegroupslearningsampleswithdifferentresolutionsareobtainedaccordingtoimageobservationmodel,andthenvectormappingsarerespectivelyusedtothosethreegrouplearningsamplestospeeduptheconvergenceofBPNN,atlast,threetimesconsecutivetrainingarecarriedontheBPNN.Trainingsamplesusedineachstepareofhigherresolutionthanthoseusedintheprevioussteps,sotheincreasingweightsstoreagreatamountofinformationforSRR,andnetworkperformanceandgeneralizationabilityareimprovedgreatly.Simulationandgeneralizationtestsarecarriedonthewell-trainedthree-step-trainingNNrespectively,andthereconstructionresultswithhigherresolutionimagesverifytheeffectivenessandvalidityofthismethod.
简介:Facerecognitionbasedonfewtrainingsamplesisachallengingtask.Indailyapplications,sufficienttrainingsamplesmaynotbeobtainedandmostofthegainedtrainingsamplesareinvariousilluminationsandposes.Non-sufficienttrainingsamplescouldnoteffectivelyexpressvariousfacialconditions,sotheimprovementofthefacerecognitionrateunderthenon-sufficienttrainingsamplesconditionbecomesalaboriousmission.Inourwork,thefacialposepre-recognition(FPPR)modelandthedualdictionarysparserepresentationclassification(DD-SRC)areproposedforfacerecognition.TheFPPRmodelisbasedonthefacialgeometriccharacteristicandmachinelearning,dividingatestingsampleintofull-faceandprofile.Differentposesinasingledictionaryareinfluencedbyeachother,whichleadstoalowfacerecognitionrate.TheDD-SRCcontainstwodictionaries,full-facedictionaryandprofiledictionary,andisabletoreducetheinterference.AfterFPPR,thesampleisprocessedbytheDD-SRCtofindthemostsimilaroneintrainingsamples.Theexperimentalresultsshowtheperformanceoftheproposedalgorithmonolivettiresearchlaboratory(ORL)andfacerecognitiontechnology(FERET)databases,andalsoreflectcomparisonswithSRC,linearregressionclassification(LRC),andtwo-phasetestsamplesparserepresentation(TPTSSR).