基于馬氏距離局部邊界fisher研究降維算法

《基于馬氏距離局部邊界fisher研究降維算法》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學術(shù)論文-天天文庫。

1、基于馬氏距離局部邊界Fisher研究降維算法　　文章編號:10019081(2013)07193005doi:10.11772/j.issn.10019081.2013.07.1930摘要：針對人臉識別應(yīng)用中的高維數(shù)據(jù)圖像以及歐氏距離不能準確體現(xiàn)樣本間的相似度的問題，提出了一種基于馬氏距離的局部邊界Fisher分析(MLMFA)降維算法。該算法從現(xiàn)有的樣本中學習得到一個馬氏度量，然后在近鄰選擇以及新樣本降維過程中用馬氏距離作為相似性度量。同時，通過馬氏度量構(gòu)造出類內(nèi)“相似”圖和類間“代價”圖來描述數(shù)據(jù)集的類內(nèi)緊湊性和類間分離性。MLMFA很好地保持了數(shù)據(jù)

2、集的局部結(jié)構(gòu)。用YALE和FERET人臉庫進行實驗，MLMFA的最大識別率比傳統(tǒng)基于歐氏距離算法的最大識別率平均分別提高了1.03%和6%。實驗結(jié)果表明，算法MLMFA具有很好的分類和識別性能。關(guān)鍵詞：馬氏距離；局部邊界Fisher分析；降維；人臉識別中圖分類號：TP391.413文獻標志碼:A英文標題4MahalanobisdistancebasedlocalmarginalfisheranalysisdimensionalityreductionalgorithmDimensionalityreductionalgorithmoflocalmarginal

3、FisheranalysisbasedonMahalanobisdistance英文作者名LIFeng1*,WANGZhengqun1,XUChunlin2,ZHOUZhongxia1,XUEWei1英文地址(1.CollegeofInformationEngineering,YangzhouUniversity,YangzhouJiangsu225127,China;2.DepartmentofLaserApplicationTechnology,NorthLaserTechnologyGroupCompanyLimited,Yangzho

4、uJiangsu225009,China英文摘要)Abstract：ConsideringhighdimensionaldataimageinfacerecognitionapplicationandEuclideandistancecannotaccuratelyreflectthesimilaritybetweensamples,aMahalanobisdistancebasedLocal4MarginalFisherAnalysis(MLMFA)dimensionalityreductionalgorithmwasproposed.AMahalanobisdist

5、ancecouldbeascertainedfromtheexistingsamples.Then,theMahalanobisdistancewasusedtochooseneighborsandtoreducethedimensionalityofnewsamples.Meanwhile,todescribetheintraclasscompactnessandtheinterclassseparability,intraclass“similarity”graphandinterclass“penalty”graphwereconstructedbyusin

6、gMahalanobisdistance,andlocalstructureofdatasetwaspreservedwell.WiththeproposedalgorithmbeingconductedonYALEandFERET,MLMFAoutperformsthealgorithmsbasedontraditionalEuclideandistancewithmaximumaveragerecognitionrateby1.03%and6%respectively.Theresultsdemonstratethattheproposedalgorithmhasve

7、rygoodclassificationandrecognitionperformance.ConsideringhighdimensionaldataimageinfacerecognitionapplicationandEuclideandistancecannotaccuratelyreflectthesimilaritybetweensamples,aMahalanobisdistancebasedLocalMarginalFisherAnalysis(MLMFA)dimensionality4reductionalg